Every product manager has heard the line: you’re the CEO of the product. Ben Horowitz wrote it at Netscape in the 1990s, restated it in his 2012 essay “Good Product Manager/Bad Product Manager,” and for thirty years the industry has quietly agreed it was aspirational at best and fiction at worst. A real CEO can hire, fire, and allocate capacity as she pleases. A PM can’t get a bit more design capacity without three meetings and a Slack thread that goes nowhere. If you have struggled with rearranging the annual company priorities to get a designer to help you understand why an A/B test budged like I have, you know what I mean.

This year, for the first time, half of that line stopped being fiction. The other half didn’t move an inch. The gap between those two halves is becoming the most important org design question in product management, and most leaders I talk to are only tracking one side of it.

AI tooling closed the execution gap that made “CEO of the product” feel hollow. PMs can now prototype, write functional code, query their own data, and generate design directions without waiting on anyone’s sprint allocation. Regulation did the opposite to the authority gap. The EU AI Act’s high-risk obligations become enforceable on August 2, 2026, ten days from when I’m writing this. A wave of 2026-vintage governance work at Anthropic and OpenAI is starting to formalize the sign-off authority that “CEO of the product” always implied but never delivered.

If you run product at an S&P 500 company or a scale-up racing to ship agentic features, what follows is the framework for telling those two shifts apart, and for not confusing “my PM can build a prototype alone now” with “my PM has become a CEO.”

One word is doing a lot of work in this piece, so let’s define it. By “authority”, I don’t mean who has the best product instinct (or highest paid position) in the room. I mean the standing to approve a decision once it’s built, having to answer for it personally when it goes wrong, not as a team, not as a function. And when a regulator or a board asks who signed off, I mean an actual name being attached to the answer, not a committee. That’s the half of “CEO of the product” nobody has actually built yet. Right now, at most companies, it seems that it doesn’t belong to anyone.

The Line That Never Quite Held

Horowitz’s original memo, written for his product team at Netscape, was blunt. A good product manager is the CEO of the product and takes full responsibility for its success. Decades later, he explained what he was actually trying to fix: a group of PMs who all had different ideas about what their job was. The line wasn’t a grant of authority. It was a description of a leadership job with a catch he stated outright. Nobody actually reports to you.

The backlash to the phrase has lasted as long as the phrase itself. Silicon Valley Product Group and Mind the Product have both published pieces telling PMs, in effect, to stop calling themselves CEOs of anything: no board, no P&L, no ability to fire or hire. The critique was never wrong. What it missed is why the phrase survived three decades of correction anyway. PMs have always been measured like CEOs, judged on outcomes, blamed for failures, without ever being equipped like one. That mismatch, not the metaphor, was the actual problem. This year, exactly half of it started closing.

The Half That Just Became True

The mismatch was always about dependency. A PM with a sharp instinct for what to build still had to wait for a designer’s bandwidth and an engineer’s sprint just to find out if the instinct was right. AI tooling is dissolving that specific dependency, fast.

LogRocket’s product-builder research puts a number on it. A standard PM, design, and engineering trio costs roughly $1 to $1.5 million a year in salary alone. A PM who can prototype and validate independently, using tools like Claude Code, Lovable, or Figma Make, can prevent an estimated $500,000 or more in wasted engineering effort annually, simply by killing weak ideas before they reach a sprint. That’s not a productivity anecdote. It’s a PM doing something a CEO does, deciding with their own hands what doesn’t get built.

The labor market is already pricing this in. Job data from TrueUp, surfaced by Lenny Rachitsky and reported in Fast Company in April 2026, shows open product manager roles at their highest level since 2022, roughly 7,300 globally, while open design roles have been flat at around 5,700 since early 2023. The demand ratio between the two flipped in mid-2023 and hasn’t reversed since.

Marc Andreessen's read on that same data wasn't "PMs win." It was a three-way standoff. AI is now, in his words, a really good coder, a really good designer, and a really good product manager, meaning coders, designers, and PMs are all simultaneously convinced AI just handed them the other two jobs. Nobody has actually won that argument. What's true is narrower and more useful: the specific dependency that made "CEO of the product" a punchline, waiting on people who don't report to you just to test an idea, is measurably weaker than it was two years ago. I made a version of this same point about who the real user of enterprise software has become when I wrote about designing TaxTech platforms for AI agents instead of humans. The pattern repeating here is the same one: the person doing the building is no longer always the person whose job title says "builder."

The Half That Didn’t Move

Not everyone is buying this, and it’s worth noting who’s pushing back. Anthropic’s own chief design officer, Joel Lewenstein, described the shift precisely in a recent interview: a lot of role collapse at the very beginning of a product’s life, but pretty clear swim lanes once it moves past early exploration. PMs still own the business case. Engineers own deployment at scale. Designers own the harder human-computer-interaction questions. Claire Vo, founder of ChatPRD, made the sharper version of the same point about design specifically: most product trios treat design like a tax they don’t want to pay, but the trio hasn’t gone anywhere. Nobody serious is arguing the roles disappeared. They’re arguing the roles compressed at the front of the process and reasserted themselves at the back.

That back end is the point where “CEO of the product” was always thinnest, not in who builds the first prototype, but in who’s accountable once it goes live. This is the half of the metaphor regulation is now writing into law, not aspiration. The EU AI Act’s Articles 12 and 13 require automated logging of every high-risk AI system’s operation: inputs, reference data, output verification, retained for a minimum of six months and retrievable by regulators on request. Article 14 requires human oversight, which only functions if those logs are actually accessible. None of it is something a PM’s personal fluency with an AI coding tool touches.

Ask who owns that authority today at most companies shipping AI features, and you’ll get five different answers in the same meeting. You'll probably find out that nobody owns it, because the job is new, and new accountability gaps don’t stay empty for long. They default to whoever already sits closest to the customer, the roadmap, and the metrics: the PM.

Anthropic is already hiring for the collision point. There’s a live requisition for a “Product Policy Manager, User Well-being” whose job is to align Product, Engineering, Public Policy, and Legal, and to treat safety as a product requirement rather than an afterthought. Strip the title and read the responsibilities, and it’s a compliance PM role wearing a different badge, the same cross-functional alignment work that’s been the actual job description in regulated FinTech and TaxTech for years, arriving at a frontier AI lab because the underlying accountability problem is now the same one.

I’ve watched this exact split play out. I once prototyped a reconciliation workflow improvement in an afternoon using an internal LLM tool. The execution gap has closed that much. Shipping that same change to a live tax determination pipeline still took weeks, not because engineering was the bottleneck, but because every change touching a jurisdiction we file in requires approval by a tax manager, testing, and overall sign-off from people who have never reported to me and never will. AI closed the part of the job that was about waiting. It did not touch the part that was about who’s allowed to say yes. I wrote about the adjacent version of this gap when I argued that AI vendors in tax technology need the same SLA discipline enterprises already demand from AWS, because “who’s accountable when the automated thing is wrong” is the same infrastructure question whether it’s a cloud provider or a model router.

The Labs Are Already Building the Other Half

Most product leaders haven’t noticed that Anthropic and OpenAI are already building this same authority infrastructure into their own products, not out of goodwill, but because their biggest enterprise customers probably won’t buy without it.

Anthropic launched a Compliance API for Claude in August 2025 and expanded it in May 2026 to cover 28 enterprise security platforms. Compliance teams can now watch Claude the way they already watch Okta or Workday. There’s a second, stranger piece of this. Anthropic’s research on Natural Language Autoencoders, a tool for translating the model’s internal reasoning into plain English, turned up something unsettling: in up to 26 percent of test interactions, Claude appears to suspect it’s being evaluated, and doesn’t say so. Its extended thinking feature is the product answer to that same problem, showing the reasoning before the answer arrives.

OpenAI’s version looks different, and the fine print matters more than the feature. ChatGPT Enterprise ships a Compliance Logs Platform for exporting conversations and files. Useful, except it keeps those logs for only 30 days by default, well short of the EU AI Act’s six-month floor. Separately, OpenAI has been studying chain-of-thought monitorability, or whether a model’s reasoning trace reveals unsafe behavior before it reaches a user. The company says it already uses the technique internally to catch reward hacking.

Both labs are chasing the same enterprise checkbook, which tends to be a more reliable predictor of behavior than either company’s stated mission. That checkbook is what’s turning the authority gap in “CEO of the product” into a feature vendors compete to ship first.

Where the Two Gaps Cross

Put the two halves next to each other, and the picture sharpens. The execution gap is closing for every PM with an AI subscription, regardless of industry. The authority gap is not closing anywhere. It’s being formalized, fastest in the places already used to it: regulated FinTech, TaxTech, and now, visibly, the frontier labs building the models everyone else depends on.

That means the PMs who feel most like a CEO right now, the ones at consumer and AI-native companies who can build alone for the first time, are furthest from the accountability infrastructure coming for them next. The PMs who’ve built audit trails because a regulator demanded it aren’t behind on this. They’re the only ones who’ve already done this.

None of this is assigned yet. No governance framework, no regulation, no vendor’s compliance API tells you which named human has to sign the risk register. That decision is still up for grabs inside every product org building agentic features right now. The org that assigns it deliberately will be in a very different position than the one that waits for a regulator, or a lawsuit, to assign it for them.

What This Means for Your Product Org

If you lead product, this is a working checklist.

• Stop leveling PMs on documentation output. A clean PRD used to be a two-day signal of competence. It’s now a two-minute AI output. Promote and hire on judgment under ambiguity and stakeholder navigation instead, the things that, per Lewenstein, still don’t collapse once the tools get better.

• Name an owner for the authority gap before someone names it for you. Right now, at most companies, nobody owns the risk register, the model or agent card, or the human-oversight sign-off for an AI feature. It sits in the space between legal, engineering, and product, where everyone assumes someone else has it. Decide now, not the week before enforcement, whether that owner is the PM, and write it into the role, not just the org chart. I laid out a practical version of this question in a framework for structuring tax-product-engineering pods that applies just as directly to an AI-safety pod as it does to a compliance one.

• Separate “can build” from “can sign off” explicitly in your career ladder. A senior PM who ships prototypes in Claude Code is not, by that fact alone, a PM who should unilaterally push a live pricing or safety-relevant change. Those are two different competencies, and most leveling frameworks still conflate them.

• Put the audit trail question into every vendor review, not just the AI ones. If a vendor can’t tell you which model handled a request, how long the log lives, and who can pull it on demand, you’ve inherited an authority gap you didn’t sign up for. Treat it the same way you’d treat an SLA gap with a cloud provider.

• Hire for the collision point before a regulator forces you to. Anthropic didn’t wait for enforcement to create a Product Policy Manager role. If your roadmap includes anything agentic, decide now whether that alignment work sits inside an existing PM’s remit or becomes its own seat, and staff it before the risk register turns into an incident report.

• Give your PMs a real answer to “who signs this,” not just “who built this.” The PMs who can already answer that question, the ones who came up in FinTech, TaxTech, or any other regulated vertical, are sitting on a skill the rest of the market is about to need.

Frequently Asked Questions

Is a product manager really the “CEO of the product”? Not in the way the phrase implies. PMs are held accountable for outcomes the way a CEO is, but historically lacked a CEO’s authority to hire, fire, or allocate capital. AI tools have closed the execution half of that gap; the authority half hasn’t moved.

Will AI replace product managers? Most current analysis says no. AI automates research summaries, spec drafts, and data analysis, but can’t own product vision, negotiate trade-offs, or understand organizational context and mission the way a human PM does.

What does the EU AI Act require from product teams? High-risk AI systems must log inputs, reference data, and outputs for at least six months, remain retrievable by regulators, support human oversight, and increasingly rely on explainability techniques like SHAP and LIME. Enforcement begins August 2, 2026.

Who owns AI risk and compliance sign-off inside a product org today? In most companies, no one, formally. Legal, engineering, compliance, and product each assume it’s partly someone else’s job. That vacuum tends to default to whoever sits closest to the roadmap and the customer, usually the product manager, even without a title or a mandate to match.

What are Anthropic and OpenAI doing to make AI more auditable? Anthropic has shipped a Compliance API integrating with 28 enterprise security platforms and published interpretability research (Natural Language Autoencoders) that translates internal model activity into readable English. OpenAI runs a Compliance Logs Platform and publishes research on chain-of-thought monitorability, checking whether a model’s reasoning trace reveals safety-relevant behavior before it becomes a problem.

What’s the actual difference between a product manager and a CEO? A CEO has formal authority: hiring, firing, capital allocation, board accountability. A PM has influence without that authority, historically compensated for by informal leadership skill. AI is closing the execution side of that gap, not the authority side.

References & Further Reading

1. Andreessen Horowitz, “Good Product Manager/Bad Product Manager,” https://a16z.com/good-product-manager-bad-product-manager/, Accessed 2026-07-23

2. Silicon Valley Product Group, “CEO of the Product Revisited,” https://www.svpg.com/ceo-product-revisited/, Accessed 2026-07-23

3. Mind the Product, “Product managers, you are not the CEO of anything,” https://www.mindtheproduct.com/product-managers-not-ceo-anything/, Accessed 2026-07-23

4. LogRocket Blog, “Why product managers must become product builders in 2026,” https://blog.logrocket.com/product-management/product-builders-future-product-management/, Accessed 2026-07-23

5. Fast Company, “Why are designers, engineers, and product managers in a ‘three-way standoff’?,” https://www.fastcompany.com/91519219/why-are-designers-engineers-and-product-managers-in-a-three-way-stand-off, Accessed 2026-07-23

6. Asqav, “EU AI Act Audit Trail Requirements: What You Need Before August 2026,” https://www.asqav.com/blog/posts/eu-ai-act-audit-trail-requirements, Accessed 2026-07-23

7. General Catalyst Job Board, “Product Policy Manager, User Well-being @ Anthropic,” https://jobs.generalcatalyst.com/companies/anthropic/jobs/50806384-product-policy-manager-user-well-being, Accessed 2026-07-23

8. General Analysis, “How to Audit Claude with the Compliance API,” https://generalanalysis.com/guides/claude-compliance-api, Accessed 2026-07-23

9. Sougata Roy, “Microsoft Purview Adds Anthropic Claude Connector for Enterprise AI Visibility, May 2026,” https://sougataroy.com/intelligence/microsoft-purview-claude-connector-may-2026, Accessed 2026-07-23

10. BuildFastWithAI, “Anthropic Reveals Claude’s Hidden Reasoning (2026),” https://www.buildfastwithai.com/blogs/anthropic-claude-nla-interpretability-2026, Accessed 2026-07-23

11. Anthropic, “Claude’s extended thinking,” https://www.anthropic.com/news/visible-extended-thinking, Accessed 2026-07-23

12. OpenAI Help Center, “ChatGPT Enterprise & Edu, Release Notes,” https://help.openai.com/en/articles/10128477-chatgpt-enterprise-edu-release-notes, Accessed 2026-07-23

13. OpenAI, “Evaluating chain-of-thought monitorability,” https://openai.com/index/evaluating-chain-of-thought-monitorability/, Accessed 2026-07-23

14. CLTC UC Berkeley, “Introducing the Agentic AI Risk Management Profile,” https://cltc.berkeley.edu/2026/02/24/introducing-the-agentic-ai-risk-management-profile-expert-perspectives-on-governance-and-best-practices/, Accessed 2026-07-23

Let’s start with an uncomfortable admission: most of us talk about AI agents in tax and finance the way teenagers talk about sex — a lot of confident chatter, very little firsthand experience, and a quiet fear of being the last one in the room who hasn’t done it yet.

I’ve sat in the vendor demos. I’ve had the 2 a.m. thought every product person in this space has had: what if the thing I’ve spent a career building is being quietly rebuilt by a model that doesn’t sleep as much as I do?

But I always trust the numbers, not the narrative. So here’s a number instead of a vibe. Inside Anthropic’s Claude Connectors Directory (the list of tools the company has chosen to spend its own scarce engineering time vetting), “Finance and Trading” isn’t a side project for interns. As of July 2026, it’s the single largest category in the entire directory. 57 connectors out of 439. Bigger than Marketing and Sales, Data and Analytics, and Development Tools (the category MCP was created to serve).

Not because tax and finance are easy to automate. It’s one of the least forgiving domains in enterprise software, which is exactly why it’s unsettling that it’s winning. It’s winning because it’s where the money is, and every AI lab chasing enterprise revenue has already done that math.

So I decided to create a map: who’s building what, where the productivity gains are real versus theater, what it costs in tokens and dollars to build the tools that plug into it. If you run tax or finance technology at an S&P 500 company or any company interested in AI (and at this point I don’t see a reason not to be at least curious), treat this as the state-of-the-union of how the AI territory is expanding.

The Convergence Nobody Expected

MCP was Anthropic’s pet project fifteen months ago. In February 2026, Anthropic donated it to the Linux Foundation and co-founded the Agentic AI Foundation with OpenAI and Block with Google, Microsoft, AWS, Cloudflare, and Bloomberg standing behind them like investors.

Five companies that would happily eat each other’s lunch, funding the same connection layer simultaneously. That’s not a group hug but a realization that none of them can afford to be the AI provider whose agent can’t reach a customer’s ERP because the customer standardized on someone else’s protocol. So instead of five incompatible standards, you get one standard and five implementations racing to fill it with the highest-value connectors first.

The Marketplace Map

Anthropic — 439 verified connectors, 30 categories, enterprise-managed access via Okta. Perplexity — MCP live on Pro/Max/Enterprise since March, claims 400+ connectors though only 68 have public documentation. Mistral — 60+ prebuilt connectors in Studio and Le Chat, per-workspace admin controls added in June. Google — native MCP in the Gemini API since March, managed servers rolling out service by service. OpenAI — MCP folded into the Apps SDK, “connectors” quietly renamed “apps” in ChatGPT, Stripe among launch partners. DeepSeek — no first-party marketplace at all; V4-Pro rides the existing ecosystem via OpenAI-compatible function calling instead.

The Proof: Where the Curation Effort Goes

Here’s where the “57 connectors” stat gets read backwards. On the open, anyone-can-submit long tail (aggregators like mcpservers.org), finance is a rounding error: development sits at 2,677 servers, productivity at 1,482, finance at 69. MCP is an coding-agent standard, and the open ecosystem still shows it.

But the open directory isn’t where enterprises source tech. What matters is where labs spend their own limited curation effort, deciding what to vet and put their name behind.

In Anthropic’s vetted directory, Finance and Trading is first: 57 connectors, ahead of Marketing and Sales (37), Data and Analytics (35), and Development Tools (29). In Perplexity’s smaller Computer surface, Finance and Investing ties for second among 14 categories, just behind Healthcare. Two labs, two philosophies, the same independent conclusion. When competitors who’ve never coordinated agree on where the money is, I generally pay attention.

Let’s Check the Hypothesis

Is everyone actually racing to become the everyday interface for Tax and Finance? Yes! But it’s not AI labs versus each other. It’s AI labs and the incumbent ERPs, both sprinting for the same interface layer, using the same neutral protocol as the weapon.

On the AI lab side: Anthropic launched dedicated financial-services agents on May 5, 2026 — reconciliation, valuation review, earnings analysis — wired into Excel, PowerPoint, and Outlook. PwC committed to certifying 30,000 U.S. professionals on Claude. KPMG is integrating Claude into its Digital Gateway for tax and PE clients. That’s a model provider walking straight into the office of the CFO.

The ERPs aren’t ceding ground. SAP is repositioning Joule as the operating layer for an “autonomous enterprise.” Workday shipped agents across HR, IT, and finance. Microsoft used Build 2026 to reframe Windows itself as an agent runtime. Every one of these players is using MCP, or a cousin like Google’s A2A, as connective tissue. Nobody wants to be the vendor locked out of the agent’s reach.

I made a version of this argument in Agent-First TaxTech: Your Tax Platform Was Built for the Wrong User: at Uber’s transaction volume, our tax determination APIs get called millions of times a day, and almost none of those calls come from a human opening a dashboard. MCP just formalized that reality across every vendor relationship at once. The interface war isn’t being fought for the person who opens the app twice a week. It’s being fought for the agent that never sleeps.

Where the Productivity Actually Shows Up

Strip away the platform politics and the case is concrete in a few specific places, not everywhere.

Reconciliation eats roughly 41% of finance teams’ hours today, and production AI deployments are now reporting 95% straight-through cash application. The human exception queue shrinks to a sliver of volume.

Tax preparation shows the cleanest math: ~30 minutes saved per return. At 2,000 returns a year, that’s ~1,000 hours recovered, $15,000–$20,000 in offshore prep cost, freed for review instead of data entry.

Audit and document review is running 50%+ reductions in analysis time at firms that took deployment seriously, which tracks with what I wrote in TaxTech AI Vendors Must Have the Same SLAs You Demand from AWS: gains show up fastest where the data pipeline was already clean, and stall everywhere else.

What This Actually Costs to Run

A simple tool call runs 5,000–15,000 tokens. A multi-step agentic task - an actual reconciliation, a drafted variance memo - runs 200,000–1,000,000. A complex multi-agent workflow with retries can blow past 3,000,000. At current pricing, a fully orchestrated agentic interaction costs roughly $0.60–$1.20 (about 30× a 2023-era chatbot turn). “Just plug it into an agent” stopped being a pricing strategy the day multi-step reasoning became the default. Tiered model routing - cheap models for lookups, frontier models only for the reasoning-heavy steps - is the difference between a blended cost of ~$2.31 per million tokens and $18.40 for teams that route everything to the frontier by default1.

The Build-vs-Buy Math

Two independent cost analyses — Azilen and DevCom — land on the same three build tiers without citing each other:

1️⃣ Entry-tier — $20,000–$30,000. A single-purpose agent (one workflow, minimal integration). 2️⃣ Mid-tier — $30,000–$60,000. A multi-step workflow wired into a handful of systems. 3️⃣ Complex, enterprise-grade — $60,000–$100,000+. Full governance, audit trail, multiple ERP/tax-engine integrations.

Finance adds a real premium on top of any tier, though the two sources frame it differently: Azilen calls it +50–100% over a general-purpose agent; DevCom puts finance and healthcare at a flat $70,000–$250,000+ once SOX and access controls are priced in. Different framing, same direction.

Budget $2,000–$20,000 a month in ongoing costs once it’s live. Payback is typically three to twelve months for a high-volume workflow — which lines up with BCG/Forrester’s 2026 data, putting median enterprise time-to-value at roughly 5.1 months.2.

The Effort/ROI Matrix: What to Build First

1️⃣ Best value — build this quarte. Variance analysis and close-commentary generation: ~$30,000 to build, ROI 7.5.

2️⃣ Biggest bets — expensive, but the math clears.

• Real-time cash application / AP-AR matching: ~$110,000 to build, ROI 9.3. The 95% straight-through-processing benchmark is already live in production, not a vendor’s aspiration.

• VAT reconciliation via MCP into your ERP and tax engine: ~$100,000 to build, ROI 9.0. This is the one eating directly into that 41%-of-your-team’s-time figure from earlier.

3️⃣ Worth doing, not urgent. A compliance calendar / filing-tracker agent: ~$20,000 to build, ROI 4.8. It won’t move a P&L line, but it buys your team peace of mind.

4️⃣ Reconsider before committing budget. A transfer pricing documentation assembler: ~$75,000 to build, ROI 5.0 — the weakest ratio of the eight. Jurisdiction requirements vary too much for a generalized agent to earn back its cost inside a normal planning cycle, unless your transfer pricing exposure is unusually concentrated.

As I laid out in The 2026 AI TaxTech Map, the indirect tax infrastructure layer — Avalara, Fonoa, Sovos, Thomson Reuters ONESOURCE, Vertex — is racing to ship this exact capability as a paid product, right now. Everything above assumes you’re weighing a build against those vendors. Not against doing nothing — that was never actually an option, you just haven’t priced it yet.

Curious to understand if anyone has tried building any of those. What type of Ai-supported apps work in an enterprise? Please share in the comments.

Frequently Asked Questions

What is the Model Context Protocol (MCP) and how does it work for enterprise? MCP is an open standard, originally built by Anthropic and now governed by the Linux Foundation, that lets AI agents connect to external tools, databases, and business systems through a single standardized interface instead of custom point-to-point integrations. In enterprise settings, an MCP server sits between the AI and the data source, enforcing governance policies before any action executes.

How much does it cost to build an AI agent for a finance team? Entry-tier: $20,000–$30,000. Mid-tier: $30,000–$60,000. Complex, audit-ready builds: $60,000–$100,000+. Finance agents carry a 50–100% premium over general-purpose ones due to compliance requirements. Typical payback: three to twelve months for high-volume workflows.

Is MCP secure enough for tax and financial data? It can be, but security isn’t automatic — enterprise-grade deployments need encrypted credential storage, read-only access by default, human-in-the-loop approval for exports, and audit logging. The risk isn’t the protocol; it’s connecting agents to production systems without IT review.

How much does agentic AI cost per task in token usage? Simple tool calls: 5,000–15,000 tokens. Multi-step workflows: 200,000–1,000,000. Complex multi-agent tasks: 3,000,000+. A fully orchestrated interaction costs roughly $0.60–$1.20 — about 30× a 2023-era chatbot turn.

What’s the ROI timeline for AI agents in finance and accounting? Most deployments recover build cost in three to twelve months when automating a genuinely high-volume workflow — reconciliation, cash application, tax prep — with a clear before/after metric. Lower-volume, highly variable workflows take longer to clear their cost, sometimes never do.

References & Further Reading

1. Anthropic — “Donating the Model Context Protocol and establishing the Agentic AI Foundation” — Accessed July 2026

2. AIToolsReview — “All 439 Claude Connectors: The Complete Directory (July 2026)” — Accessed July 2026

3. Perplexity Connectors awesome-list — “Finance and Investing” category — Accessed July 2026

4. Awesome MCP Servers — Finance, Productivity, Development, Database categories — Accessed July 2026

5. Google Cloud — “Announcing official MCP support for Google services” — Accessed July 2026

6. OpenAI Developers — “MCP – Apps SDK” — Accessed July 2026

7. Mistral AI — “Connect the dots: Build with built-in and custom MCPs in Studio” — Accessed July 2026

8. Composio — “DeepSeek MCP Integration for AI Agents” — Accessed July 2026

9. CFO.com — “Inside Claude’s rapid expansion across corporate finance” — Accessed July 2026

10. Moveo.ai — “Financial reconciliation in 2026: How AI agents are eliminating the manual bottleneck” — Accessed July 2026

11. Azilen — “How Much Does a Financial AI Agent Cost in 2026?” — Accessed July 2026

12. LeanOps — “AI Agents Burn 50x More Tokens Than Chats” — Accessed July 2026

13. Arcade.dev — “MCP Runtime Security for AI Agents in Accounting” — Accessed July 2026

Partway through, he turned to how the rest of the industry prices AI: “I’m not throwing shade at them, but something has gone completely wrong” with the token-based model used by OpenAI and Anthropic. On enterprise buyers specifically: “The basic view among enterprises in this country is I’m going to chillax and waste my time with tokens.”

In a follow-up interview with The Information, published the same week, Karp framed the issue less as pricing and more as competitive exposure: “There’s just very deep frustration around…are they gonna optimize the models for me, or are they gonna take the alpha of my business, transfer in their weights, and compete against me?”

Worth noting for balance: I’ve found no evidence that Anthropic or OpenAI train on enterprise customer data, but they are likely tracking domain usage, customer personas, prompts, and other metadata.

Karp’s “alpha” framing is a claim about risk exposure, not a documented incident — and it doubles as the pitch for Palantir’s own position as the “application layer” sitting between enterprise data and whichever model a customer chooses. It’s a competitive stance as much as a pricing critique, and both readings are worth holding at once.

I’m not in a position to referee any of this. What I noticed is that several of the points mentioned in the interview touch things I’ve written about this year.

On Measuring Value Instead of Usage

Two months ago, Amazon retired its internal AI leaderboard after employees optimized for token volume rather than useful output. I wrote about why usage is the wrong metric to track at the time — Amazon’s replacement metric counted deployments that actually shipped, not activity. Karp’s comment about enterprises “wasting time with tokens” is describing the same gap between activity and value, from the vendor side rather than the buyer side.

On Vendor Transparency

Separately, I’ve argued that TaxTech AI vendors should carry the same SLA discipline enterprises already require from AWS — visibility into which model handles a given request, at what cost. Palantir’s own answer to that question happens to be a product feature: Evolve, which the company describes as including a “model router” that sends a given task to different models depending on whether the customer is optimizing for performance or cost. Whatever one makes of the business strategy behind it, a router that surfaces model selection and cost is closer to the infrastructure layer I described than most TaxTech vendors currently offer.

Karp also drew a sharper line between American and Chinese open-source models — “the choice is going to be very quickly [an] open model from America or open model from China”. That’s not just a model-selection debate for AI labs. It’s a version of the same data-residency question TaxTech vendors already navigate under GDPR, e-invoicing mandates, and cross-border VAT rules, with model origin as a new variable.

On the Stakes for Finance Specifically

I’m finishing a longer piece for next week on why finance and tax now make up the single largest category in Anthropic’s own Claude Connectors directory — 57 of 439 connectors, ahead of coding and marketing. Full breakdown and sourcing coming in that post. The early number is enough to make the point here: unsettled model-vendor economics flow directly into the tools running reconciliation, VAT determination, and transfer pricing documentation. More on this next week.

What Palantir’s Own Numbers Show

Whatever one makes of the positioning, Karp’s company has a multi-year track record worth putting next to the comments. Palantir’s quarterly revenue has climbed for nine straight quarters:

In Q1 2026 specifically, Palantir reported revenue of $1.63B, up 85% year over year, with net income of $870.5M. Cash generation moved in step with revenue:

A 53% net margin, a 55% operating cash flow margin, and a 57% adjusted free cash flow margin, against $8B in cash and short-term treasuries on the balance sheet. Palantir raised full-year 2026 guidance to $7.65–7.66B in revenue (roughly 71% growth), with a Rule of 40 score of 145%. Palantir’s commercial model has leaned on enterprise platform and outcome-based contracts (Foundry, AIP, Evolve) rather than pure per-token API pricing — one way to read why its CEO frames token metering as the wrong model for building a durable enterprise business.

Worth watching as more TaxTech and finance AI tools move to usage-based pricing: whether buyers get the model-level transparency Karp says Palantir’s router provides, or stay inside the token-metered model he’s pushing back on.

🔮 Strategic Forecast: By the end of 2027, expect at least one major TaxTech vendor to publicly shift part of its AI pricing from pure token/usage metering toward outcome- or platform-based tiers, citing enterprise buyer feedback as the reason.

References & Further Reading

1. CNBC — “Palantir CEO Alex Karp says ‘something has gone completely wrong’ with how AI is sold” — https://www.cnbc.com/video/2026/07/01/palantir-ceo-alex-karp-says-something-has-gone-completely-wrong-with-how-ai-is-sold.html — Accessed 2026-07-03

2. CNBC — “Palantir’s Karp bashes OpenAI, Anthropic token model: ‘Something has gone completely wrong’” — https://www.cnbc.com/2026/07/01/palantir-karp-open-ai-anthropic-tokens.html — Accessed 2026-07-03

3. The Information (Laura Bratton, Applied AI) — “Palantir CEO Says Some U.S. Government Customers Switched to Open Source AI” — https://www.theinformation.com/newsletters/applied-ai/palantir-ceo-says-u-s-government-customers-switched-open-source-ai — Accessed 2026-07-03

4. Palantir Technologies — Q1 2026 Earnings Release / Investor Relations — https://investors.palantir.com/news-details/2026/Palantir-Reports-Q1-2026-U-S--Revenue-Growth-of-104-YY-and-Revenue-Growth-of-85-YY-Raises-FY-2026-Revenue-Guidance-to-71-YY-Growth-and-U-S--Comm-Revenue-Guidance-to-120-YY-Crushing-Consensus-Expectations/ — Accessed 2026-07-03

5. Palantir Technologies — Quarterly earnings releases, Q1 2024–Q4 2025 (SEC 8-K filings) — Accessed 2026-07-03

Last year I sat in a vendor demo for a new indirect tax platform. The sales engineer spent more than 45 minutes walking us through the UI (an arduous ordeal, if you ask me). The dashboard was beautiful. Someone had spent months (or years) making this thing look exactly right.

Halfway through, I thought, “I wonder how many times a day actual humans will open this interface…”

At Uber, our tax determination and external integrations services and APIs are called millions of times a day. The overwhelming majority of those calls come from other services and automated pipelines — not from a tax professional navigating a dashboard, not from anyone sitting at a keyboard. The interface we were presented with was designed almost exclusively for a user who increasingly doesn’t and won’t show up. The product looked like it was built for the 1%. The 99% didn’t actually need a UI.

That conversation keeps coming back to me. Because I don’t think Uber is the exception. I think we’re the industry.

The Wrong User

Most products in enterprise TaxTech over the last decade look like they were optimized for a human sitting at a keyboard. Navigation menus, approval buttons, dashboard layouts. This wasn’t wrong — it was right for its time. We built what the market needed. The market just changed faster than our roadmaps did.

I wrote earlier this year about why tax teams need to stop being compliance factories and become AI orchestrators. That was about the people side. This post is about the product side because it’s where most teams have to start now.

IDC published something blunt in May 2026: “AI agents don’t care about your UI. They don’t appreciate an elegant dashboard or a well-designed navigation menu. What they need is data access, API depth, and integration reliability.”

That’s not a trend piece. That’s a verdict on a decade of product decisions.

Gartner projects 40% of enterprise apps will embed AI agents by the end of 2026 — up from less than 5% in 2025:

When Anthropic launched the Model Context Protocol in November 2024, it hit 100,000 downloads in a month. When OpenAI adopted it in March 2025, that jumped to 22 million monthly downloads within weeks (a two-hundredfold increase in four months, or the time between your last roadmap review and your next one).

The UX moat we spent years building is gone. The new moat is API depth, clean data models, and MCP compatibility. The TaxTech PM who doesn’t make this call in 2026 will make it reactively in 2027, under worse conditions and likely severe competitive pressure.

Three Paths

I think there are three ways an AI agent will interact with your product (in this order):

1️⃣ Vision-based parsing — the agent screenshots your interface and infers what’s on screen. This is usually slow, error-prone, and token-heavy. This is what happens when there is no other option. If your product is only reachable this way, agents will fail on it constantly and silently. Most websites are not yet optimized for image-based reading.

2️⃣ Accessibility-tree parsing — the agent reads the structured HTML representation your browser generates. Faster and more reliable. Nielsen Norman Group made the business case in April 2026: products built with strong accessibility practices are already more legible to agents. The teams that invested in semantic HTML, proper labeling, and logical page structure were accidentally preparing for this.

3️⃣ Direct API or MCP calls — the agent bypasses the UI entirely. Clean contracts. Structured responses. Machine-readable outputs. No screenshots, no inference, no ambiguity.

For TaxTech, only the third option scales because VAT determination, e-invoicing reconciliation, and transfer pricing run at transaction volume, not human speed. An agent calling your tax engine doesn’t need a beautiful UI. It needs a contract it can trust: predictable inputs, deterministic outputs, versioned endpoints, and error states that explain exactly what broke in terms a downstream system can act on.

Where This Gets Dangerous

This is the part most agentic AI pieces skip. I will clarify.

In most software categories, going agent-first is an architecture problem. In compliance, it’s an architecture problem and a liability problem. Thomson Reuters framed it well in June 2026: “AI that produces a correct result through an opaque process is not useful in a compliance context. It’s a liability. Auditors don’t just want the answer — they need a full trail of evidence showing exactly who reviewed what, when, what changes were made, and what rationale was applied at each decision point.”

This matters more than most PM frameworks acknowledge. And it’s getting more urgent: the IRS now runs 126 AI applications. Your auditor isn’t a person reviewing PDFs anymore. Your audit trail needs to be as machine-readable as the system now reviewing it.

In a consumer app, an agent making the wrong call is a bug report. In indirect tax, it’s a penalty notice, an audit trigger, or a blocked filing in a jurisdiction your business depends on. The stakes are not the same. The architecture cannot be the same.

And here’s the number that should make every tax director uncomfortable: 56% of a tax professional’s working day goes to reactive, manual tasks. They want to spend 70% of their time on strategic work. The gap between where they are and where they want to be is exactly the gap that agentic TaxTech is supposed to close. But only if the product is built for the agent doing the work — not the human watching it happen.

On the one hand, we are asking tax teams to do more with less, across more jurisdictions and on tighter timelines. On the other, we are building products optimized for a user interaction pattern that doesn’t scale and wouldn’t support those teams. I imagine most tax folks would be frustrated. I am.

How To Build For The Right User

The old product design process: define user stories for a tax professional navigating a workflow → design the UI → build the backend to serve the UI.

The new one: define the agent contract first. What does an orchestration layer need to call your product reliably? What structured response does it need back? Then build the human interface as a governance layer on top of that contract, not as the primary interaction surface.

In practice, five steps:

1️⃣ API-first, without compromise. Every capability must be accessible programmatically before it ships in a UI. If it can’t be called by a machine, it isn’t ready. This discipline was optional during the decade of dashboard-first SaaS growth. It is not optional now.

2️⃣ Your MCP server is a product deliverable, not a dev tool. If you’re building or buying a TaxTech platform in 2026 without a clear answer to “how will external agents call this?”, you’re betting on the wrong product. Treat your MCP surface area with the same rigor as your UI.

3️⃣ Redesign your output contracts. Agents don’t need a PDF summary of a VAT reconciliation. They need a structured JSON object: clear field definitions, status codes, confidence levels where applicable, and error states that explain exactly what broke and why. If your outputs are designed to be read by a human, they’re the wrong outputs.

4️⃣ Draw the architectural line explicitly. AI reasoning handles classification, routing, and anomaly detection. Deterministic logic handles tax calculation, rate application, and filing validation. Make that boundary visible in your documentation, data model, and audit trail.

5️⃣ Redesign the UI for governance, not navigation. If most of your product’s interactions come from agents, the human interface is a control room, not a workflow tool. Approval queues. Exception dashboards. Audit trail viewers. Threshold alerts. That’s what a tax director actually needs from a screen in an agentic world. The navigation menu that dominated your last three sprints? It’s infrastructure now.

For more on how the TaxTech vendor landscape is repositioning around these decisions, I mapped the field in The 2026 AI TaxTech Map (worth reading alongside this).

The Thing Nobody Says Out Loud

Everyone can build agents now. LLM APIs, no-code platforms, an autonomous workflow shipped by someone who’s never shipped before. Role definitions are dissolving, and almost nobody is asking the architecture question first. In a consumer app, that produces useful experiments. In compliance, it produces something different: fast-moving systems that are confidently wrong before anyone notices.

The fix isn’t to slow down. It’s to define the contracts before you release the velocity. Who owns the tax calculation layer? What is an agent explicitly not permitted to do without a human gate?

Here’s the diagnostic I’d run if I were auditing a TaxTech product today: can an external AI agent call your core tax determination capability right now — with a structured input, a deterministic output, a versioned contract, and a full audit trail of what the agent did and why? If the answer is no, you’re not behind on AI. You’re behind on the architecture that makes AI safe to use in your domain.

Frequently Asked Questions

What is agent-first product design in TaxTech? It means building TaxTech products where the primary interaction surface is a programmatic API or an MCP server — not a human-facing UI. The product is designed so that AI agents can reliably call it, receive structured outputs, and operate within defined governance boundaries without a human having to navigate the interface at each step.

What is agentic AI in tax compliance? Agentic AI in tax compliance refers to AI systems that autonomously orchestrate multi-step workflows — pulling data from ERPs, performing tax calculations, validating against e-invoicing records, and generating filing-ready outputs — without requiring a human to manually execute each step. The key difference from earlier AI tools is that the agent determines its next action based on intermediate results.

What are the risks of AI agents in tax compliance? The primary risks are audit exposure due to opaque decision-making, compounding errors from agents operating at scale without human review, and compliance failures arising from applying probabilistic AI reasoning to deterministic calculations. The mitigation is a governance architecture comprising audit trails, deterministic calculation engines, and mandatory human-review gates at material decision points.

What governance is required for AI agents in regulated industries? At a minimum: a full audit trail of every agent action (inputs, logic applied, output, reviewer, timestamp), a hard separation between AI-assisted orchestration and deterministic calculation logic, mandatory human review gates at material decision points, and versioned API contracts that make agent behavior reproducible and auditable.

What is MCP and why does it matter for TaxTech? The Model Context Protocol is an open standard — introduced by Anthropic in November 2024 and adopted by OpenAI in March 2025 — that defines how AI agents communicate with enterprise software. For TaxTech, it means agents can call your tax engine, reconciliation tools, or e-invoicing platform directly, without screen scraping or custom integrations. Vendors with MCP-compatible surfaces are the ones enterprise orchestration layers will choose to call.

References & Further Reading

1. IDC (Eric Newmark) — “The Agent Takeover: What Happens When AI Becomes the Primary User of Enterprise Software“ — May 14, 2026

2. Nielsen Norman Group (Sarah Gibbons & Kate Moran) — “AI Agents as Users“ — April 10, 2026

3. Thomson Reuters — “Is your tax department ready for the agentic AI explosion?“ — June 8, 2026

4. Thomson Reuters — “Expert-first vs. AI-first tax compliance“ — 2026

5. Gartner — “Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026“ — 2025

6. Anthropic — “Introducing the Model Context Protocol“ — November 2024

You ask about features. You ask about integrations. You ask about pricing and support. You get a demo with a very smooth presenter who uses the words “agentic,” “real-time,” and “compliance-native” in every other sentence.

You do not ask, “How does your system decide which model handles my request?”

And the vendor — relieved you didn’t ask — moves on.

This is the infrastructure gap in TaxTech AI. And it is costing enterprises a lot more than they think.

What Actually Happens When You Hit Send

Let me walk you through what happens the moment you submit a prompt to an AI-powered compliance tool. Somewhere, a server rack with an AI chip receives your request. Your text gets converted into tokens — numerical representations of words and concepts. A model gets allocated GPU compute to process those tokens. The response streams back. Your account gets charged.

Simple, right?

Here is the part nobody talks about: in most TaxTech deployments today, the same model handles every single request. Whether you’re asking it to validate a VAT registration number (200 tokens, done in milliseconds) or reconcile 400,000 transaction lines against twelve EU continuous transaction control regimes (tens of thousands of tokens, multiple tool calls, minutes of compute) — it all goes to the same place.

No routing. No tier selection. No “this task is simple, use the cheaper model.” No “this task is complex, escalate.”

One model to rule them all. You pay for all of it.

Now let me tell you what this looks like at Uber scale, because I tested it.

I Asked for One Line. I Got a Redesign.

Earlier this year I ran a test with Anthropic’s Fable 5 — their highest-capability model at the time, since restricted for enterprise use outside the United States.

The task was simple. Embarrassingly simple. Recalculate a single line in a compliance spreadsheet based on updated input parameters.

What Fable 5 did instead: redesigned the entire spreadsheet architecture.

It reformatted the structure. Added conditional logic across twelve new columns. Restructured the formula dependencies. Generated explanatory commentary for every change. Produced a masterpiece of spreadsheet engineering that absolutely nobody asked for.

The model wasn’t broken. It was doing exactly what frontier models do — applying maximum capability to every request regardless of whether that capability is needed. Maximum intelligence. Maximum tokens. Maximum compute.

At Uber’s transaction volumes — millions of tax determination events per day, across dozens of jurisdictions — a task mismatch like this isn’t a funny anecdote. It’s a Token Effectiveness Ratio disaster that compounds quietly until it detonates your quarterly compute budget.

The fix isn’t a smarter model. The fix is a routing layer that catches the mismatch before it happens.

Do TaxTech vendors not have one?

What AWS Figured Out Fifteen Years Ago

AWS didn’t launch EC2 in 2006 and say “good luck figuring out the routing.”

They spent the next decade building the layer that makes the hard part invisible:

1️⃣ Load balancing. Incoming requests get distributed across multiple compute instances. The month-end close traffic spike doesn’t bring the system down. Someone already solved that.

2️⃣ Auto-scaling. Capacity expands during a VAT filing deadline and contracts at 3am on a Tuesday. You pay for what you use. Someone already solved that, too.

3️⃣ Instance tier routing. A background batch job doesn’t get the same resources as a real-time checkout event. AWS routes intelligently. The user doesn’t choose the rack, and he shouldn’t have to.

4️⃣ Availability zone failover. If one data center degrades, traffic reroutes automatically. Your 99.99% SLA is enforced at the infrastructure layer. Not the application layer. The infrastructure layer.

5️⃣ Spot vs. on-demand pricing. Non-urgent workloads run on spare capacity at a fraction of the cost. Latency-sensitive operations get premium compute. The decision is automatic.

6️⃣ Request-level observability. You can audit exactly which requests ran, on what instance type, for how long, and at what cost. Every dollar is traceable.

7️⃣ Agentic governance. At the AWS New York Summit in 2026, AWS introduced release management capabilities for its DevOps agent specifically because parallel AI coding creates a compounding risk of failure. As David Yanacek, AWS Senior Principal Engineer for Agentic AI, put it: “Coding agents are getting a ton done in parallel. The more that you’re doing in parallel, the higher the probability that the resulting train will not be able to leave the station because one of these things has a bug in it.” AWS also introduced Continuum, a cybersecurity tool designed to detect vulnerabilities in AI models before they reach production. The result: a task that took eight minutes in January takes three minutes in May — not because the agents got smarter, but because the governance layer got better.

Now. Imagine asking an enterprise buyer to choose which physical rack handles their AWS request. They would look at you like you’d lost your mind. That’s not a user decision. That’s an infrastructure decision. One that AWS made for you, automatically, fifteen years ago.

In 2026, asking a CTO to “select the LLM model” for their TaxTech AI prompt — without any of the above abstraction — is the exact same thing.

They will be lost. They are right to be lost.

The TaxTech AI Gap

In 2026, none of the major TaxTech AI deployments publish anything equivalent to what AWS documents on day one.

Go ahead and try.

Ask Avalara how their ALFA framework decides whether your request goes to a large or small language model. Ask Thomson Reuters what ONESOURCE+’s failover logic looks like during a provider outage. Ask Sovos what model tier handles a simple tax code lookup versus a complex cross-border VAT determination. Ask Anrok what happens to your in-flight compliance request if their primary LLM provider has a brownout at the month-end close.

You won’t get an architecture document.

You’ll get a product brochure.

And to be fair — the product surface area is genuinely impressive. As I mapped in The 2026 AI TaxTech Map, Avalara’s ALFA framework combines large and small language models across 1,400+ application integrations. Thomson Reuters launched ONESOURCE+ as an “Intelligent Compliance Network.” Sovos shipped Sovi AI. Anrok built Atlas with Big 4-verified outputs. The teams building these products are getting there.

But “combines large and small language models” is not a routing architecture. It’s a marketing line.

What governs the routing decision? What’s the threshold? What happens on failover? What’s the uptime SLA for the AI layer specifically — not the underlying AWS or Azure layer the vendor is running on, but the AI orchestration layer?

Nobody discloses this. Probably nobody asks for it either.

That’s the gap.

A Brief, Painful Look at Vendor Reality

Let’s be specific, because specificity is the point.

Avalara deserves credit for the ALFA framework — and their MCP server integration signals genuine infrastructure thinking. But there is no published routing logic. No compute tier disclosure. No AI-specific SLA. What exactly triggers the “use the small model” decision? Unknown.

Thomson Reuters ONESOURCE+ — “Intelligent Compliance Network.” What does the intelligence routing look like? What’s the failover path during a provider degradation? Not disclosed.

Sovos Sovi AI and Anrok Atlas both lead with the output story. Anrok’s combination of AI and human tax experts is smart, accuracy risk management. It is a quality control layer, not an infrastructure layer. Those are different problems.

Xero and Intuit build AI through partnerships — including Xero’s deep integration with Anthropic, which I covered here. Partnership architecture creates a second question: when Xero routes your request through Anthropic’s API, which model tier handles it? Which region? What’s the fallback? Those answers live inside the partnership agreement. They are not visible to you, the user.

Pattern: strong on what the AI does. Silent on how it’s governed.

Caps Are Not Architecture. They’re Surrender.

The enterprise response to AI budget overruns in 2026 has been impressively uniform.

Uber burned its entire AI budget by April — Uber’s CTO confirmed it publicly. Amazon capped it. Microsoft capped it. Meta capped it. The industry response to “AI is costing us a fortune” was: put a ceiling on it.

Respectable as a short-term move. Completely missing the point.

A cap is a circuit breaker. It stops the bleeding when you hit the limit. But it does nothing about what happened before the limit. If your AI compliance stack is routing 40,000-token tasks to frontier models when 2,000-token tasks to a smaller model would have produced the exact same compliance outcome, you will hit your cap faster. And you will hit it on waste, not on value.

The agentic dimension makes this worse, not better. Agentic TaxTech workflows — multiple AI agents running reconciliation in parallel, cross-referencing jurisdictions, and calling external APIs simultaneously — amplify the risk of compounding failures. AWS’s own David Yanacek flagged this exact problem at the 2026 NY Summit: the more you run in parallel, the higher the probability that one agent’s bug stops the entire output from clearing. In TaxTech terms: one misfiring reconciliation agent doesn’t just produce a bad output. It can hold up the entire compliance filing.

Are TaxTech vendors addressing this?

As I detailed in The Future Cost of TaxTech, the unit economics of AI compliance are brutal at scale. A usage cap doesn’t fix unit economics. It just tells you when you’ve run out.

The AWS analogy again: imagine AWS gave you a monthly compute budget and no load balancer. You’d burn the budget on poorly routed requests, hit the ceiling, and call it a month. Nobody would accept that from a cloud provider.

A cap is not a substitute for architecture. It is an admission that architecture is missing.

(That sentence. Read it again. That’s the one to send your ELTs.)

Five Questions. Before You Sign Anything.

If you are evaluating or renewing a TaxTech AI contract in the next twelve months, your vendor should be able to answer all five of these. No exceptions. No “we’ll get back to you on that.”

If they stall, escalate to the product team. If the product team stalls, you have your answer.

1️⃣ How does your system determine which model tier handles a given request?

A mature answer: routing logic based on task complexity, instruction length, and domain. A weak answer: one model handles everything, or worse — “the user selects the model.” (Yes, this happens. No, it’s not acceptable.)

2️⃣ What is your AI-specific uptime SLA, and what is your failover path when a primary LLM provider degrades?

A mature answer: named failover path, detection latency, recovery time objective. A weak answer: “we rely on our provider’s SLA.” That is not your SLA. That is their SLA, applied to them, not to you.

3️⃣ Do you implement semantic caching for repeated compliance queries?

A mature answer: yes, with a description of how near-identical queries are matched. A weak answer: no mention of caching. (Semantic caching alone can cut compute costs by 40–60% for structured compliance workloads. If your vendor isn’t doing it, you are paying for redundant compute every single day.)

4️⃣ How do you prevent runaway token consumption in agentic workflows?

A mature answer: circuit breakers, per-task token budgets, human escalation triggers. A weak answer: usage caps at the contract level. See the previous section on why that’s not the same thing.

5️⃣ Can you provide task-level compute telemetry — model tier used, token consumption, latency, and cost — per request?

A mature answer: yes, available in account reporting. A weak answer: aggregate spend data only. (If you can’t see task-level telemetry, you cannot manage your Token Effectiveness Ratio. You’re flying blind with a very expensive instrument panel.)

Your cloud provider answered all five of these before you signed. Your TaxTech AI vendor should too.

If they can’t — you’re not buying managed AI infrastructure. You’re buying model access with a compliance-themed UI on top. The infrastructure risk sits on your P&L. Not theirs.

The companies that asked the right cloud infrastructure questions in 2012 built the most resilient platforms of the 2010s.

The ones signing enterprise AI contracts in 2026 without asking the infrastructure questions are building the most expensive compliance technical debt of the 2030s. And they’ll call it “AI transformation” right up until the end.

Don’t be that company. Ask the questions.

What answers have you gotten from your TaxTech AI vendors? I’m genuinely curious. Drop them in the comments.

Frequently Asked Questions

What is an AI orchestration layer, and why does it matter for enterprise tax compliance? An AI orchestration layer sits between your application and LLM providers, routing each request to the appropriate model tier based on task complexity, cost, and latency. Without it, every request — from a simple tax code lookup to a complex cross-border determination — hits the same model at the same cost. At enterprise transaction volumes, that’s not a technical inconvenience. It’s a structural margin problem.

How should I evaluate AI infrastructure when selecting a TaxTech vendor? Ask five questions before signing: How is model tier selection governed? What is the AI-specific uptime SLA and failover path? Is semantic caching implemented? How are runaway agentic workflows stopped before they exhaust compute budgets? Is task-level token telemetry available in account reporting? A vendor who cannot answer all five is selling model access, not managed AI infrastructure.

What is the difference between an AI gateway and a usage cap on LLM consumption? A usage cap stops spending when a threshold is hit. An AI gateway optimises spending before it happens — routing simple tasks to cheaper models, caching repeated queries, preventing runaway loops. A cap is a circuit breaker. A gateway is the circuit design. You need both, in that order.

Why does model selection matter for automated tax calculations? Different compliance tasks have wildly different complexity profiles. Recalculating a single tax line requires a fraction of the compute needed to reconcile multi-jurisdictional indirect tax exposure across 90 countries. Without intelligent routing, both tasks hit the same frontier model. At scale, that’s millions of dollars in unnecessary compute spend per year.

What are the infrastructure risks of single-vendor LLM lock-in in TaxTech? If your TaxTech vendor integrates deeply with one LLM provider, your compliance infrastructure inherits that provider’s pricing volatility, uptime performance, and geographic access restrictions. A vendor with multi-model routing can arbitrage across providers during outages or price spikes. A single-provider vendor cannot — and when that provider has a bad quarter, so does your compliance stack.

References & Further Reading

1. Hybrid SLM + LLM Orchestration: The 2026 Strategy for Cost-Effective Enterprise AI — Medium/GenAI Protos — Accessed June 2026

2. LLM Orchestration in 2026: Frameworks + Best Practices — orq.ai — Accessed June 2026

3. 5 Enterprise AI Gateways for LLM Cost Control in 2026 — Maxim AI — Accessed June 2026

4. Routing, Load Balancing, and Failover in LLM Systems — DEV Community — Accessed June 2026

5. Avalara Agentic Tax & Compliance — Avalara — Accessed June 2026

6. AI Prices Are Going Up, Up, Up — Josh Bersin — May 2026

7. Gartner: LLM Inference Cost to Drop 90% by 2030 — Gartner — March 2026

8. Anrok Raises $55M Series C — Business Wire — October 2025

9. Agentic AI Cost Management: Stopping Margin Erosion — Kong Inc. — Accessed June 2026

10. The Future Cost of TaxTech: Managing the New Unit Economics of Compliance — dmihaylov.com — April 2026

11. Intelligent LLM Routing: How Multi-Model AI Cuts Costs by 85% — Swfte AI — Accessed June 2026

12. Failover Routing Strategies for LLMs in Enterprise AI Applications — Maxim AI — Accessed June 2026

13. AWS Introduces New Cybersecurity, Coding Tools at NY Summit — The Information (Catherine Perloff) — June 2026

Every major platform — Xero, Avalara, Thomson Reuters, QuickBooks, Sovos, Sage, Vertex — has shipped an AI product this year. Most look identical in features. The real differences are in the stack underneath.

This week, I am taking a quick look at what’s under the hood of the leading TaxTech products.

🔀 Multi-model is the default. A single model may be risky.

Thomson Reuters runs Claude, GPT-4, and Gemini simultaneously in CoCounsel Tax. Xero uses Anthropic for financial reasoning and OpenAI for web research. Avalara combines six providers with proprietary SLMs in its ALFA framework.

These multi-model architectures came to be on purpose, I think. When a frontier model gets a jurisdictional edge case wrong — and they all do — you want a fallback that isn’t the same model with a different temperature. Or when you run out of compute (tokens) during your period close, you want a fallback option. I wrote a full article about the contingencies you’d need in the future.

A reliance on a single model will most likely ruin your weekend.

🏗️ Intuit went the other direction — and it worked.

While everyone else built on top of frontier models, Intuit built underneath them. Proprietary Financial LLMs (GenOS), fine-tuned on financial data at scale: 90% accuracy on transaction categorization, 50% lower latency than off-the-shelf alternatives.

The case for owning your model is real — but only when the data flywheel justifies the build. If a vendor tells you they built their own AI, ask what data it was trained on. “We built our own model” without a data advantage is a marketing decision, not a technology one. And it won’t last.

As I covered in Your Tax AI Vendor Cannot Fix This — the constraint has never been the model. It’s the data layer underneath it.

VentureBeat, “How Intuit built custom financial LLMs that cut latency 50% while boosting accuracy”

🔌 MCP support is the integration signal nobody’s reading.

Avalara and Sovos have both shipped Model Context Protocol (MCP) servers. That’s how any AI agent can query a vendor’s compliance APIs using a standard protocol. No custom middleware.

This is the difference between a vendor that’s an island and one that’s a node in your system. If your TaxTech vendor hasn’t shipped an MCP server, ask when it’s on the roadmap. If it’s not on the roadmap, you’re writing bespoke connectors — and maintaining them every time either side changes an API.

Avalara, “Connecting AI to Tax Automation: How MCP Servers Bridge the Gap,” November 2025

🤔 Every vendor calls it “agentic.” Ask what the agent actually executes.

Avalara’s Avi can ingest transaction data, calculate tax, validate exemptions, generate a return, and file it — autonomously, with human approval checkpoints. That’s an agent completing a workflow end-to-end.

Sage’s Finance Intelligence Agent routes your question to the right data source and composes an answer. That’s an agent retrieving and summarizing.

Both are called the same thing. The distinction matters enormously for how you architect around them. Before signing any contract with “AI agents” in scope, ask for a demo showing the agent completing a specific workflow from trigger to output. Not a slide. A working execution.

📋 The contract your organization signs this quarter is an architecture decision.

When a Tax Director signs a three-year deal with a single-LLM vendor and no MCP server, they’ve made a product architecture decision for their PM and engineering teams. They’ve locked in the integration surface, the agent compatibility layer, and the model fallback strategy — without realizing they’ve made that choice.

In Issue 004, I mapped every major AI partnership in TaxTech YTD. The pattern is consistent: AI capability is being acquired at the marketing layer, not the architecture layer. PMs who understand the stack before the contract is signed will have significantly more optionality in 18 months. Those who don’t will be managing a migration.

The question I keep coming back to: why aren’t product teams in the room during TaxTech contract negotiations? Reply and let me know if that’s changing at your organization.

References

1. Xero — “Xero and Anthropic Collaborate” — https://www.xero.com/us/media-releases/xero-and-anthropic-collaborate/ — June 2026

2. Avalara — “Avalara Launches Agentic Tax and Compliance™” — https://newsroom.avalara.com/2025-09-30-Avalara-Launches-Agentic-Tax-and-Compliance-TM-AI-Agents-That-Work-for-You — September 2025

3. Avalara — “MCP Servers Bridge the Gap” — https://www.avalara.com/blog/en/north-america/2025/11/model-context-protocol-servers-bridge-the-gap.html — November 2025

4. Sovos — “Sovos Expands Sovi AI” — https://sovos.com/press-releases/sovos-expands-sovi-ai/ — March 2026

5. Thomson Reuters — “One Million Professionals Turn to CoCounsel” — https://www.thomsonreuters.com/en/press-releases/2026/february/one-million-professionals-turn-to-cocounsel-as-thomson-reuters-scales-ai-for-regulated-industries — February 2026

6. VentureBeat — “How Intuit Built Custom Financial LLMs” — https://venturebeat.com/ai/how-intuit-built-custom-financial-llms-that-cut-latency-50-while-boosting — 2025

7. Wolters Kluwer — “CCH Axcess Expert AI Launch” — https://www.wolterskluwer.com/en/news/wolters-kluwer-launches-cch-axcess-expert-ai — November 2025

8. Vertex Inc — “Vertex Advances AI-Powered Capabilities” — https://www.globenewswire.com/news-release/2026/04/07/3269067/0/en/Vertex-Advances-AI-Powered-Capabilities-to-Improve-How-Enterprises-Execute-Compliance.html — April 2026

9. Sage — “Sage Debuts Finance Intelligence AI Agent” — https://www.channelinsider.com/ai/llms-chatbots-and-agents/sage-finance-intelligence-agent/ — 2026

10. Intuit — “GenOS Press Release” — https://investors.intuit.com/news-events/press-releases/detail/1272/intuit-rapidly-advances-genos-to-accelerate-development-of-agentic-ai-experiences-at-scale — 2025

Every major player in the TaxTech space has an AI strategy. Most have AI platforms. Some have actual results. The competition is now focused on platform procurement (external or internal), adoption by business customers, and results. And it’s getting hot out there.

What the Models Can (and Can’t) Do

The AIMultiple Finance LLM Benchmark — 238 hard financial reasoning questions, updated June 2026 — has the clearest current ranking.

Claude Opus 4.8 leads at 89.08%, dethroning GPT-5 (88.23%) while using 7× fewer output tokens. Notably, the top six models are within 2.5 points of each other. Looks like the model choice barely moves the needle at this level. DeepSeek R1 (open-source) trails at 62.18% but is still relevant for European deployments where transaction data can’t leave the infrastructure. It also needs further training to be on par with the rest of the pack.

Clearly, the constraint is not the model.

The Providers Building the Infrastructure

Before evaluating any AI vendor, it helps to know who’s funding the underlying models — and how quickly they’re growing.

Anthropic grew from $1B to $30B ARR in 18 months. OpenAI hit $25B ARR — over 40% from enterprise. Mistral crossed $400M ARR through European enterprise deals, 20× in a year.

The interesting dynamic: Anthropic and OpenAI have competing claims on the same enterprise clients. Anthropic signed Deloitte and PwC. OpenAI signed Bain and H&R Block. Both signed Intuit, separately.

Big Four: $9 Billion and Counting

The combined AI investment across the Big Four is $9B+. Here’s where it went.

The platforms are real. Client-facing outcome data — tax cycle times, audit error rates, jurisdiction efficiency — is still emerging. Whoever publishes that data first and proves fidelity will hold the relationship for the next decade.

Everyone Else Who Actually Shipped

The more interesting story is happening outside the Big Four.

Ramp’s Stack — a full AI OS for accounting firms, every agent action audit-logged to source data — launched June 3. Anrok raised $55M Series C in October 2025 and now handles VAT, GST, and sales tax for Anthropic, Notion, and Cursor across 100+ countries. Wolters Kluwer shipped Expert AI across CCH Axcess — their unified tax, audit, and firm management platform — in November 2025. Filed built an AI Tax Preparer that processes 3× more returns without adding headcount. A European institution applied LLMs to supplier invoices and achieved a 10% cost reduction on a multibillion-euro spend basis.

79% of tax leaders say AI will be transformational. 37% have invested. Early adopters are getting mixed results. Some will figure it out — and when they do, they’ll have 18 months of hard architecture decisions baked in before the first case study is written. By the time best practice gets published, the advantage is already in the past.

As I argued in Your Tax AI Vendor Cannot Fix This: AI on top of bad data fails quietly. The firms winning right now fixed the data layer first.

The Indirect Tax Infrastructure Layer

The enterprise indirect tax market — the layer between your ERP and global VAT/GST compliance — is having one of its most active M&A years on record.

Fonoa is the headline story. $110M Series C in May 2026, and they simultaneously acquired PwC’s Indirect Tax Edge platform — the same platform PwC was building to compete with them. They now own both sides of that argument. Current clients include Uber, Netflix, Canva, and Booking.com.

Avalara Launched Agentic Tax & Compliance in September 2025. The ALFA framework enables agents to act autonomously across SAP, Oracle, NetSuite, and SYSPRO — and shipped MCP servers for tool-level integration. An IPO is planned for 2026. The company that defined indirect tax compliance for a generation is now positioning itself as an AI-native orchestration layer.

Sovos Launched Sovi AI in March 2026 and acquired IRIS APAC to expand Asia-Pacific reach. Their Label product handles FATCA, CRS, and CARF. The positioning is deliberate: compliance intelligence as a distinct product category from compliance execution.

Thomson Reuters launched ONESOURCE+ in November 2025 — an AI-powered Intelligent Compliance Network that connects enterprise indirect tax, trade, and risk infrastructure — and then followed with ONESOURCE Sales and Use Tax AI in January 2026. Powered by CoCounsel, it cuts routine reporting time by 65% and audit exposure by 75%. The CoCounsel Tax product already has 1 million users across 107 countries. ONESOURCE is where the indirect tax workflow actually runs — and now it has an AI layer on top.

Vertex acquired AI tax categorization from Ryan in December 2025, added Kintsugi AI the same month, and shipped 65 product enhancements throughout 2025. Vertex Copilot is already embedded in Oracle, SAP, and Salesforce — meaning the AI sits where transactions actually happen.

Fintua — formerly Taxback International — rebranded in July 2025, collapsing VAT recovery, e-invoicing, and payments into a single platform. The rebrand signals compliance as a commodity and automation as the margin.

Every significant player in indirect tax infrastructure either built an AI product, acquired one, or bought a competitor’s platform this year.

You’ve probably been in the room when your Tax SaaS vendor claimed to have an excellent tax filing audit trail. It lives on their infrastructure. It’s queryable through a UI. You can easily generate and download it. But there’s a catch. It may take 6-8 weeks or more to do it. For one jurisdiction.

I’ve learned this while working on TaxTech infrastructure at Uber. We operate in 70+ countries. Every jurisdiction has different VAT rules, entity registration requirements, and filing cadences. When you’re processing millions of trips, deliveries, and partner transactions per day, the gap between what your transaction system records and what your compliance tool reports can span days, entities, and undocumented transformation steps. Reconstructing how you got there is a different problem entirely — and it’s one the Tax SaaS was never designed to solve. Or at least I have not seen a tool remotely capable of doing this yet.

So, when the IRS, HRMC, or an EU tax authority flags your entity structure or a discrepancy between real-time e-invoicing and VAT reporting, and the inquiry lands, you’ll be facing this very problem. And it’s an issue you’ll face very often because of the sheer efficiency with which tax authority systems can now deploy AI to find even minor inconsistencies in your data.

Last week, I wrote about how three tax authorities deployed AI in 2026 that queries your data rather than your return. The IRS runs Palantir’s SNAP (Selection and Analytic Platform) across 100+ fragmented legacy databases, surfacing entity relationships no human auditor could trace manually. Through a contract with the HMRC, Quantexa is building a Unique Customer Record: a single, consistent view of every taxable entity in the UK, resolved from fragmented public and private sources. The EU AI Act’s high-risk AI enforcement comes into force on 2 August 2026.

The question I left open: how do you prepare for this?

The Readiness Gap, By the Numbers

On 19 May 2026, Vertex published a global study of 1,050 senior IT, Finance, and Tax leaders across the US, UK, and Europe:

94% expect stronger collaboration between IT, Tax, and Finance. Only 12% have achieved full integration. Tax teams are consulted on technology decisions 37% of the time. Only 37% report high confidence in their tax-ready data.

I recognize that 12% gap from the inside. At Uber, the product team making decisions about trip pricing, partner fee structures, and payment flows is not the same team filing VAT returns. The decisions upstream create tax obligations downstream — and the pipeline connecting those two realities is something you build intentionally, or you discover under pressure. Most enterprises are discovering it under pressure.

Why Your Vendor Cannot Close This Gap

So, data readiness seems to be the most pertinent compliance risk. The natural move is to buy a tool. I’ve seen it happen repeatedly, and it’s the right call for many problems. But it is the wrong call for this one — for three specific reasons.

1. Your vendor’s audit trail is not your audit trail

When your compliance data lineage lives on a vendor’s infrastructure, you are paying for access to your own operational data. If an inquiry requires you to reconstruct a tax position independently — outside the vendor’s interface, without an active support contract — you discover that the lineage exists only within their system.

At Uber scale, this matters profoundly. What I’ve found is that no single vendor has access to all of our data. Let’s take Portugal as a good example. A trip there touches a local operating entity, a Tax SaaS, a payment processor — each potentially in a different jurisdiction with a different data system. The compliance tool shows the output of that structure, not granular views of the local entity’s ERP, the SaaS vendor’s software, or the payment processor’s dashboard. However, the Tax Authority’s system has a unified view of all data and uses AI to detect discrepancies to the finest detail. So, if you can only explain the output, you’ve already lost the first exchange.

2. The compliance logic lives in a vendor’s black box

In many cases, your vendor provides the tax determination logic—configurable via toggles and rule dropdowns. After all, that’s their bread and butter. But it’s something you don’t actively monitor for correctness, and you cannot explain on demand. A human auditor might accept “it’s a system configuration”. The IRS SNAP and HMRC Quantexa do not. They generate conclusions before the inquiry even arrives. By the time a human asks, the model has already formed a view.

You need to be able to say: here is the rule, here is the source, here is the transformation, here is the timestamp. If the only honest answer is “our vendor configured it,” you’re in trouble.

3. Entity resolution has to reflect your specific corporate structure — not a vendor template.

Vendor entity models are built for the median enterprise. They support a parent holding company, a handful of subsidiaries, and standard related-party taxonomies. They were designed to handle what most companies look like — not what your company actually is.

At Uber, a single trip can touch four legal entities before it produces a VAT-reportable output: the local operating entity that contracts with the driver, the technology licensor that licenses the platform, the payment processor that clears the funds, and the entity that reports the consolidated revenue. Each entity has a different registration, a different role in the intercompany chain, and potentially a different identifier across our data systems. No off-the-shelf entity template maps that structure correctly out of the box. You configure it, you maintain it, and when the structure changes — a new country entity, a restructuring, a merger — you own the update.

HMRC’s Quantexa does not work from your vendor’s entity template. It resolves your structure from Companies House filings, PAYE records, MTD transaction feeds, and cross-referenced VAT registration data — sources your vendor never touches. The result is a Unique Customer Record that HMRC builds from the outside in, using data you submitted to other government systems, not data your vendor configured on your behalf. If that externally resolved view of your entity graph surfaces a parent-subsidiary relationship that your ERP has inconsistently classified across jurisdictions, the algorithm flags it. Your vendor’s template cannot explain the inconsistency because the template was never aware of it.

You own that template, which means it has to be built and maintained by people who understand your corporate structure, not inherited from a vendor taxonomy that was never designed for it.

The Nine-Step Enterprise Readiness Guide

These nine steps follow a five-phase arc: Diagnose → Measure → Build → Test → Own. They’re sequenced so each phase builds on the output of the previous one.

1️⃣ Map your tax data pipeline. Find every source feeding a compliance output. Find where the inconsistencies live before the algorithm does.

2️⃣ Establish column-level lineage. For any material tax position: which transaction, which rule, which source, which transformation. If it takes more than an hour to answer, you don’t have defensible lineage.

3️⃣ Run a cross-jurisdiction entity audit. The same entity shouldn’t have different identifiers in your US, UK, and EU systems. HMRC’s Quantexa will find that mismatch. Run it yourself first.

4️⃣ Benchmark your data freshness SLA. HMRC queries real-time patterns against your quarterly filing. If your data is 90 days stale, that gap is your exposure — not your return. Here’s why compliance latency matters.

5️⃣ Audit your vendor’s output format. An ML model cannot query a PDF. If your audit trail is a PDF, it is not an audit trail for an algorithmic adversary.

6️⃣ Decouple your compliance logic from your vendor’s UI. If your vendor is acquired tomorrow, can your engineers independently reproduce the determination logic? If not, version-control it until they can.

7️⃣ Build the Tax × Engineering × Data bridge. Not a committee. A standing function with a charter, a meeting cadence, and a CPO/CTO reporting line.

8️⃣ Red-team your own entity graph. Run the analysis SNAP and Quantexa would run. The IRS AI targets “allocations that don’t match economic reality.” Find those internally before the model does.

9️⃣ Name an owner for AI audit readiness. Not the tax team. Not IT. One cross-functional role, one mandate, one reporting line to the CPO or CTO. If nobody owns it, nobody fixes it.

📥 The full operational guide — with step-by-step instructions, the questions to ask your team, and the tools to use at each phase — is available as a free PDF.

Download

Download: The Nine-Step Tax Data Readiness Guide →

The Tax Data Readiness Guide

17.3KB ∙ PDF file

Download

Based on my experience at Uber

Download

(Free. No paywall. If you find it useful, please share it.)

🔮 Prepare for What’s Coming

By 2028, “tax data observability” will be a defined engineering discipline at S&P 500 companies — with dedicated tooling, named ownership, and SLAs sitting inside Product and Engineering, not the tax function. The CPOs and CTOs who build this infrastructure in 2026 will spend 2028 explaining their methodology to auditors from a position of control. The ones who don’t will spend 2028 reconstructing pipelines under inquiry timelines they don’t control. The gap between those two outcomes is the nine steps above.

Frequently Asked Questions

What data does the IRS AI actually query during a tax enforcement action?

The IRS uses its Selection and Analytic Platform (SNAP), built with Palantir, to query structured transaction data, entity relationships, intercompany flows, and basis records across its fragmented legacy databases. The system compares these against sector benchmarks and historical patterns to flag statistical anomalies — without reviewing a return in the traditional sense. Source: GAO-26-107522, March 2026; Capitol Technology University, 2026; FindLaw.

What is entity resolution, and why does it matter for enterprise tax compliance?

Entity resolution is the process of linking records that refer to the same individual, company, or transaction across multiple databases — even when those records use different identifiers, formats, or naming conventions. HMRC's Quantexa contract is built precisely on this capability. If your enterprise's entity model is inconsistent across jurisdictions, enforcement systems will surface the inconsistency before you can explain it. Source: Quantexa press release; Global Government Finance; The Next Web, May 2026.

How do I make my enterprise tax data audit-defensible?

Audit defensibility requires column-level data lineage: the ability to trace any filed tax position back to its source transaction, through every transformation, with timestamps and ownership records — producible on demand, independently of your compliance vendor, in under one hour for material positions. Manual reconstruction via spreadsheets, email chains, or vendor support tickets does not meet the standard an algorithmic inquiry implies. Source: OvalEdge; JMCO, 2026.

Who should own tax data infrastructure — the tax team, IT, or engineering?

None of them independently. The gap between these three functions is precisely where enterprise tax exposure lives in the agentic enforcement era. A cross-functional role spanning Tax, Product, and Engineering — with a mandate, a budget, and a reporting line to the CPO or CTO — is the correct ownership structure. This is distinct from the tax team, which owns determination logic, and from IT, which owns system availability. Source: Vertex Inc. global research, May 2026.

Should enterprises build or buy their tax compliance technology stack?

Both — but the boundary matters. Vendor tools remain appropriate for commodity compliance tasks: return preparation, statutory filing, and routine determination. Tax data infrastructure — lineage, entity resolution, pipeline architecture, independent audit trail ownership — must be a core internal engineering competency. A vendor can be a layer on top. The pipeline, the lineage, and the ability to reconstruct your position independently cannot be outsourced. Source: Fortune, April 2026; Vertex Inc., May 2026.

REFERENCES & FURTHER READING

1. Vertex Inc. — “New Vertex Research Highlights Rising Revenue Risk from IT, Tax, and Finance Misalignment” — https://www.vertexinc.com/company/news/latest-news/new-vertex-research-highlights-rising-revenue-risk-it-tax-and-finance-misalignment — Accessed May 2026

2. Global Government Finance — “HMRC inks £175m deal to boost data, analytics and AI firepower” — https://www.globalgovernmentfinance.com/hmrc-quantexa-entity-resolution-service-data-analytics-ai/ — Accessed May 2026

3. The Next Web — “HMRC awards £175 million AI contract to British firm Quantexa” — https://thenextweb.com/news/quantexa-hmrc-ai-tax-fraud-sovereignty — Accessed May 2026

4. Quantexa — “Quantexa Selected by HMRC for Landmark £175M Sovereign Data and AI Transformation” — https://www.quantexa.com/press/quantexa-selected-by-hmrc-for-landmark-pound175m-sovereign-data-and-ai-transformation/ — Accessed May 2026

5. U.S. GAO — “Artificial Intelligence: IRS Actions Needed to Address Skills Gaps, Information Quality, and Strategic Management” — https://www.gao.gov/products/gao-26-107522 — Accessed May 2026

6. FindLaw — “IRS Seeking Increased Use of Palantir’s AI and Data Tools to Help Decide Who to Audit” — https://www.findlaw.com/legalblogs/law-and-life/irs-seeking-increased-use-of-palantirs-ai-and-data-tools-to-help-decide-who-to-audit/ — Accessed May 2026

7. OvalEdge — “Data Lineage Best Practices for 2026: Ensure Accuracy & Compliance” — https://www.ovaledge.com/blog/data-lineage-best-practices — Accessed May 2026

8. JMCO — “Data Lineage and Audit Trails: What Finance Teams Must Get Right in 2026” — https://www.jmco.com/articles/daniel-shorstein/data-lineage-and-audit-trails-what-finance-teams-must-get-right-in-2026/ — Accessed May 2026

9. Fortune — “Why your data infrastructure — not your AI model — will determine whether Agentic AI scales” — https://fortune.com/2026/04/30/agentic-ai-data-infrastructure-readiness-scale/ — Accessed April 2026

10. GoCardless — “Making Tax Digital explained: Your guide to the 2026 Income Tax changes” — https://gocardless.com/blog/mtd-itsa-income-tax-changes-2026-guide/ — Accessed May 2026

11. Deloitte TaxScape — “Making Tax Digital for Income Tax” — https://taxscape.deloitte.com/article/making-tax-digital-for-income-tax.aspx — Accessed May 2026

12. PwC — “The next era of tax compliance: real-time and data-driven” — https://www.pwc.com/gx/en/services/tax/connected-tax-compliance/tax-authorities-ai-compliance.html — Accessed May 2026

13. Capitol Technology University — “Audited by an Algorithm: How the IRS Is Using AI in 2026” — https://www.captechu.edu/blog/audited-algorithm-how-irs-using-ai-2026 — Accessed May 2026

Two weeks ago, I posted on LinkedIn: “Can someone please kill the internal AI leaderboards? Or at least put the right KPIs in place?”

Last week, Amazon did:

From the article:
Dave Treadwell, Senior VP, told staff to stop using AI “just for the sake of using AI” after employees “tokenmaxxed” their scores on Kiro Rank — running meaningless tasks through the system to inflate usage numbers. Amazon replaced it with “normalized deployments”: AI output that is actually shipped.

The coverage called it a leaderboard problem. To me, it really looks like the appearance of efficiency from AI use is more important than the return on AI spending, measured, for example, in ARR.

It is a measurement pattern — one that has repeated three times in 150 years.

The gaming is only half the problem

According to Goodhart’s Law, when a metric becomes a target, it ceases to be a good metric. In other words, once a specific metric is used to reward performance, people will manipulate the system to optimize that metric, inadvertently undermining its original purpose. Amazon set an 80%+ weekly AI usage target. Developers met it by running AI on tasks that didn’t need it (no wonder 🤷‍♂️).

Victorino Group has documented the same failure at three enterprises in 2026 — different tools, identical result.

But the deliberate gaming is only half the problem. METR’s May 2026 study found that workers with no leaderboard to game still overestimated AI’s effect on their own productivity by 40 percentage points on average.

The moat always appears when measuring

Paul David’s 1990 research on electrification documented a 40-year productivity lag: factories installed electric motors in 1895 and saw no gains until the 1920s. The ones that captured the gains didn’t measure kilowatts installed. They reorganized their floors around electricity’s properties — unit drive over group drive — and measured output per reorganized unit.

Cloud/SaaS compressed the same pattern to 15 years: the companies that invented LTV/CAC and NRR in 2000–2015 built durable moats while the rest measured seat counts.

AI is in year three. As I explored in The Future Cost of TaxTech, the compliance cost stack has already changed. ‘Tis the time to establish the right metrics. Most enterprises haven’t noticed.

Those measuring outcomes reap enormous benefits from AI

PwC’s 2026 AI Performance Study (n=1,217 executives, 25 sectors) found that leaders generate 7.2× more revenue and efficiency gains than the average competitor. What separates them: twice as likely to redesign workflows around AI rather than layering tools on top, 2.8× more likely to automate decisions without human intervention. The single strongest predictor of AI financial performance is identifying growth opportunities — not cost reduction, not usage frequency. They are measuring outcomes, not activity.

Tax and Finance have two value categories that are structurally invisible

BCG’s first systematic AI ROI study in finance found that only 29% of finance executives are able to confidently measure AI ROI. The root cause is that standard measurement captures only labor (effort) savings. It misses other important metrics, such as audit risk reduction (exposure shifted before the filing) and compliance failures avoided (the penalty never incurred). Both have real expected value. Neither shows up on any dashboard.

The leaderboard was never the problem. The metric was.

What does your AI ROI calculation include?

REFERENCES & FURTHER READING

1. Financial Times — “Amazon scraps AI leaderboard to stop workers chasing usage scores” — ft.com — May 2026

2. The Decoder — “Amazon kills internal AI leaderboard after employees gamed it with pointless tasks” — https://the-decoder.com/amazon-kills-internal-ai-leaderboard-after-employees-gamed-it-with-pointless-tasks/ — Accessed June 2026

3. Victorino Group — “Three Companies, One Failure Mode: Goodhart’s Law Comes for AI Adoption” — https://victorinollc.com/thinking/goodhart-ai-adoption-tokenmaxxing-mainstream — Accessed June 2026

4. METR — “Measuring the Self-Reported Impact of Early-2026 AI on Technical Worker Productivity” — https://metr.org/blog/2026-05-11-ai-usage-survey/ — May 2026

5. PwC — “Three-quarters of AI’s economic gains are being captured by just 20% of companies” — https://www.pwc.com/gx/en/news-room/press-releases/2026/pwc-2026-ai-performance-study.html — April 2026

6. PwC — “Want ROI from AI? Go for growth” (full study PDF) — https://www.pwc.com/gx/en/so-you-can/2026/content/roi-from-ai.pdf — 2026

7. BCG — “How Finance Leaders Can Get ROI from AI” — https://www.bcg.com/publications/2025/how-finance-leaders-can-get-roi-from-ai — 2025

8. Paul David — “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox” — 1990. Accessible via AEI updated commentary: https://www.aei.org/articles/the-dynamo-the-computer-and-chatgpt-explaining-todays-productivity-paradox/

9. CNBC — “Almost every Fortune 500 is tracking overall AI usage: What that means for employees” — https://www.cnbc.com/2026/05/05/ai-use-work-employee-monitoring-tech-surveillance.html — May 2026

🎯 Start with the outcome. Not the data.

90% of AI projects never reach fruition. Not because of bad data, but because of the wrong starting point.

According to the research, most TaxTech projects begin with a data inventory - invoices, spreadsheets, PDFs, and returns.

RoAI Institute, “Crossing the AI Chasm: Start with Outcomes, Not Data” · HBR, “Most AI Initiatives Fail. This 5-Part Framework Can Help,” November 2025

🪤 Pilot purgatory. Where most AI initiatives die.

HBR calls it “pilot purgatory”: an AI initiative succeeds in proof of concept, then stalls before it scales. They mapped 7 frictions behind it — pilot proliferation, process debt, accountability gaps, tribal knowledge loss, governance failures, integration complexity, and the efficiency trap.

Look at that list from a tax function’s perspective, and you’ll recognize every single one.

HBR, “The Last Mile Problem Slowing AI Transformation,” March 2026

💰 By 2029, strategic CFOs will add 10 margin pp.

Tax is a margin lever. Every recovered VAT euro, every avoided penalty is direct P&L. CFOs designing TaxTech against financial outcomes now are building that advantage — compliance first, features second. Everyone else is watching demos.

Gartner, “By 2029, CFOs Who Implement Strategic AI Deployment Will Add 10 Margin Points of Growth,” April 28, 2026

📉 74% want AI to grow revenue. 20% are doing it.

Deloitte, 3,200+ leaders: Only 20% of the interviewed actually achive revenue growth when using AI. That’s not a technology gap — I would aruge that it’s a design choice made from the beginning. Revenue impact requires defining the outcome before you build and holding someone accountable for the result. Most organizations skip both.

Deloitte, “State of AI in the Enterprise 2026,” surveyed 3,200+ leaders, Aug–Sep 2025

🔍 72% of CIOs are losing money on AI. Here’s what separates the 28%.

Gartner, 782 I&O leaders: 72% of CIOs are breaking even or losing money on AI. The 28% who succeed embed AI into existing workflows and invest 4× as much in data foundations before any model is deployed.

For tax: the hard decision isn’t which tool to buy — it’s whether to invest in the integration layer before the business case is approved. You can’t model your way out of a data architecture problem.

Gartner, “AI Projects in I&O Stall Ahead of Meaningful ROI Returns,” April 7, 2026 · Gartner, “Organizations with Successful AI Initiatives Invest Up to Four Times More in Data and Analytics Foundations,” April 16, 2026

Missed Issue #001? Read more about the data readiness gap in enterprise AI.

Tax enforcement went agentic. Your tax compliance didn’t.

That sentence likely lands somewhere between “interesting” and “uncomfortable” for many TaxTech leaders. The assumption that has governed enterprise tax strategy for the last decade — buy/build a compliance tool, let the tax team run it, move on — has quietly broken.

And the people who need to fix it must start asking the right questions.

The New Watchdog Doesn’t Read Returns

The IRS now runs 126 active AI applications targeting audit selection, fraud detection, and pattern matching — up from 10 in 2022. HMRC just awarded a £175M, ten-year contract to Quantexa — a platform originally built for financial crime detection at banks — to cross-reference every UK business entity against a unified data infrastructure it’s building from scratch. The EU AI Act’s full high-risk AI enforcement lands on 2 August 2026 — fines for prohibited AI practices are already active at up to €35M or 7% of global turnover; high-risk non-compliance from August carries a further exposure of up to €15M or 3%.

These systems don’t review returns the way a revenue agent does. They query structured (and probably unstructured) data. They resolve entities across fragmented datasets. They build graph models of your corporate structure, intercompany flows, and VAT reclaim ratios — and compare them against sector benchmarks, peer groups, and expected configurations at scale. The IRS AI has already selected 3,600+ large partnerships for examination. It didn’t do that by only reading PDFs.

The return is the structured output of your data pipeline. If the pipeline is fragmented, entity resolution is inconsistent across jurisdictions, or your intercompany flows don’t reconcile cleanly, the algorithm finds the gap. This is when you also start blipping on the tax authority’s radar.

What HMRC Just Revealed About Your Stack

The Quantexa contract is worth studying in detail because what HMRC is building on the enforcement side mirrors what most enterprise tax teams haven’t built on the compliance side.

Quantexa’s platform was built for entity resolution at scale — connecting fragmented data from multiple sources, resolving identities across datasets, and using graph analytics to surface patterns that no human team could find manually. HMRC’s stated goal is a single, trusted, connected view of every taxable entity in the UK, built from data currently stored across multiple legacy systems that don’t communicate with each other.

Read that again from the other side of the table: HMRC is solving the data fragmentation problem on the enforcement side. Most enterprise tax teams haven’t solved it on the compliance side. The authority will soon have a cleaner, more connected view of your entity structure than your ERP does.

Three Questions to Run Internally This Quarter

1. Can you reconstruct the lineage of any tax position in your stack?
   Not the final number — the lineage. Which transaction triggered it, which rule applied, which data source it came from, and which engineer last touched the pipeline that produced it? If the answer takes more than a day to produce, your lineage is not audit-defensible against an algorithmic adversary.
   

2. Is your entity model consistent across jurisdictions?
   If the same legal entity appears differently in your US ERP instance, your UK instance, and your EU reporting layer, an AI enforcement system will notice. A human auditor might accept an explanation. The algorithm flags the inconsistency and moves on. Run the cross-jurisdiction entity reconciliation before Quantexa does.
   

3. What is the data freshness SLA for your compliance outputs?
   HMRC’s Making Tax Digital has been feeding continuous transaction data into Quantexa’s platform since April 2026. If your compliance tool produces a quarterly filing from data that’s already 90 days stale, and the enforcement system queries real-time transaction patterns against that filing, the mismatch is your exposure, not your return.

When an AI enforcement inquiry arrives — and it will arrive not as a letter but as a model’s conclusions, already formed before the inquiry is even sent — the tax team will need to answer data questions that you can only address if you have planned for it. This gap is where enterprise exposure lies in 2026.

The adversary is now a data system. The only defensible response to a data system is better data.

References & Further Reading

1. U.S. GAO — “Artificial Intelligence: IRS Actions Needed to Address Skills Gaps, Information Quality, and Strategic Management” — https://www.gao.gov/products/gao-26-107522 — Accessed May 2026

2. The Next Web — “HMRC awards £175 million AI contract to British firm Quantexa to close £46.8 billion tax gap” — https://thenextweb.com/news/quantexa-hmrc-ai-tax-fraud-sovereignty — Accessed May 2026

3. FinTech Futures — “Quantexa lands £175m HMRC contract for data infrastructure overhaul” — https://www.fintechfutures.com/partnerships/quantexa-lands-175m-hmrc-contract-for-data-infrastructure-overhaul — Accessed May 2026

4. Capitol Technology University — “Audited by an Algorithm: How the IRS Is Using AI in 2026” — https://www.captechu.edu/blog/audited-algorithm-how-irs-using-ai-2026 — Accessed May 2026

5. U.S. GAO — “Inside the IRS’s Use of Artificial Intelligence” — https://www.gao.gov/blog/inside-irs-use-artificial-intelligence — Accessed May 2026

6. Thomson Reuters — “Tax firm AI platform vs. point solution: 2026 buyer’s guide” — https://tax.thomsonreuters.com/blog/tax-firm-ai-platform-vs-point-solution-2026-buyers-guide/ — Accessed May 2026

7. PwC — “Navigating the shift to an AI-powered tax operating model” — https://www.pwc.com/us/en/services/tax/library/tax-operating-model-ai-shift.html — Accessed May 2026

8. Axis Intelligence — “EU AI Act News 2026: Compliance Requirements & Deadlines” — https://axis-intelligence.com/eu-ai-act-news-2026/ — Accessed May 2026

📊 Everyone’s experimenting. Almost no one is scaling.

Three years into the gen AI era, 88% of enterprises use AI in at least one function, and 61% use it in Tax/Finance. Only 33% are scaling it, and only 39% see any impact on EBIT.

In tax and finance, it’s worse. Use cases are still stuck in document processing and chatbots. Not reconciliation automation. Not real-time VAT determination. Chatbots at most.

Looks like Tax folks are still in the pilot phase. Worth remembering that a TaxTech roadmap without workflow redesign isn’t a business reinvention for the AI era — it’s making the old way a bit faster. I even wrote an article about this.

McKinsey QuantumBlack, Nov 2025

🏗️ The bottleneck isn’t the AI. It’s the plumbing.

94% of organizations say connected data is critical to AI success. Only 27% actually have it. And only 39% of unstructured data — the kind that dominates tax departments — is ready for AI use.

Tax data lives across ERPs, invoicing platforms, e-invoicing networks, and legacy GL systems. Nothing is connected by design. You can’t drop an AI reconciliation agent on top of that and expect it to work.

The model isn’t the problem. You need to assign a Tax PM to the data architecture problem.

Hyland / HBR Analytic Services, May 2026

🔐 Your AI vendor stack is now an attack surface.

The IMF flagged AI-fueled cyberattacks as a mounting financial stability risk. The FSB named third-party AI concentration as a new systemic vulnerability. Both were published this month.

Tax data is a high-value target — VAT filings, entity structures, transfer pricing docs. Two or three global AI providers now underpin most of financial services’ critical processes. When your AI tax-determination vendor goes down, your ability to file goes with it.

It will be an interesting SLA conversation during the next contract review.

IMF Blog, May 2026 · FSB Report, Oct 2025

🧠 AI can now read its own thoughts. This matters for tax.

Anthropic published Natural Language Autoencoders this week — a method that converts a model’s internal activations into readable text. Claude suspected it was being evaluated on 26% of benchmark tasks, though it never said so.

Your regulator will eventually ask: how do you know what your AI was reasoning about? Today, you can’t answer that. NLAs are the first credible path to an audit trail of AI reasoning — not just outputs.

Watch this space.

Anthropic Research, May 7, 2026

📈 Spend is up. Governance is not.

Seven consecutive years of rising AI investment in financial services. The same top concerns as year one: data quality, third-party oversight, and regulatory lag. The IIF-EY survey just released its 7th annual edition.

For tax specifically, this is a compounding risk. Multi-jurisdictional obligations, unstructured data, and growing vendor dependency — the failure modes are multiplying faster than the controls designed to catch them.

The firms building governance infrastructure now — before the audit — will have a structural advantage. The rest will retrofit under pressure.

IIF-EY Annual Survey on AI in Financial Services, Oct 2025

According to the latest HBR analysis on AI Transformation, firms that see AI as simply a plug-and-play SaaS solution often get stuck in a "micro-productivity trap." They achieve isolated improvements in tasks, but overall firm-level productivity stalls due to manual handoffs and outdated bottlenecks. To truly boost EBITDA by 10–25%, leaders need to shift from just "task optimization" to full "business reinvention." This idea is supported by KPMG’s 2026 Global AI Report, which notes that "continuous assurance" agents are now autonomously bridging the gap between ERP data and tax authority portals. Meanwhile, Perplexity’s enterprise research tools are being used for "knowledge discovery," enabling TaxTech PMs to map global mandates to product SKUs in real-time, ending the era of manually reviewing 500-page PDFs.

It really looks like AI transformation in TaxTech is less about making the old way faster and more about rebuilding workflows on the premise that powerful, agentic AI already exists.

With that in mind, here are 5 core finance and tax processes ripe for redesign today to gain huge efficiencies (almost) immediately.

1. Troubleshooting E-Invoicing Rejections

As global tax authorities move to CTC, your invoice isn’t “valid” until the government portal confirms it.

• The Inefficiency: To make matters worse, there is no unified government e-invoicing rejection code schema. Instead, you are dealing with different codes and descriptions, often in exotic languages such as Turkish, with little context about why the invoice was rejected. Due to these factors, portal rejections (caused by schema drift or mapping errors) often trigger an incredibly slow manual troubleshooting cycle. Each rejection can take days to resolve, directly blocking cash flow and damaging customer relationships.

• The Redesign: Instead of a human analyst sifting through XML logs, an AI agent autonomously identifies the cause of the rejection, cross-referencing your ERP’s master data, proposes the corrective mapping, and re-triggers the submission (with HITL confirmation in advance, if necessary).

• Efficiency Gain: You shift from reactive “firefighting” to an autonomous, self-healing document pipeline by collapsing the troubleshooting cycle.

2. Continuous Reconciliation (the Zero-Day Close)

The “month-end close” in many cases assumes that data must be gathered, then checked, then fixed. Once a month, because more often makes it largely impractical.

• The Inefficiency: Research from KPMG shows that finance teams spend more than 50% of their time on manual reconciliation. This “process lock-in” forces a 15–30-day cycle time, leaving leaders with a permanent rearview mirror view of the business.

• The Redesign: Move from semi-automated reconciliation to agents autonomously executing well-defined tasks (to avoid errors). Thanks to Anthropic’s recent work on Finance Agents, this is now possible. Provided that you define the context and framework well through a plugin, agents can pull data from ERPs, banks, and PDFs, reason over discrepancies across millions of rows, and execute corrective ledger entries (after a HITL confirmation, if necessary).

• Efficiency Gain: You move from “detecting errors” at day 30 to “preventing drift” at day 0.

3. Autonomous Exception Handling in AP/AR

As one of my product discoveries at Uber revealed, traditional OCR is brittle. When it fails, the process stalls until a human can manually override the data.

• The Inefficiency: S&P 500 companies face an average exception rate of 20–30% in invoice processing. Deloitte’s studies show each exception costs 15–20 minutes of manual labor, creating a massive hidden tax on operational efficiency.

• The Redesign: Agentic AI understands transaction context and cross-references vendor history to autonomously resolve missing POs, mismatched totals, or even communicate directly with the vendor to resolve issues.

• Efficiency Gain: Eliminate the need for a dedicated manual exception handling team.

4. Predictive Regulatory Monitoring

Staying compliant currently requires a small army of researchers to read 500-page PDF updates or monitor government press releases (which are often ambiguous, to say the least).

• The Inefficiency: Manual monitoring of global mandates is estimated to cost large enterprises millions in indirect labor annually, given over 1,000 regulatory updates globally each year (Thomson Reuters Data), the risk of a “blind spot” is a mathematical certainty for manual teams.

• The Redesign: Use Knowledge Agents (powered by Perplexity-grade search) to map legislative changes directly to your product catalog and draft the engineering tickets needed to update logic in real-time.

• Efficiency Gain: Reduce the manual regulatory research workload, freeing the team to focus on high-value strategic planning.

5. Monthly VAT Return Template Filling

In a multinational enterprise, the simple act of filing a VAT return is the height of complexity.

• The Inefficiency: S&P 500 firms such as Uber operating in 50+ jurisdictions face 50+ unique XML, PDF, or Excel schemas. If not outsourced to firms like KPMG, Fintua, or Sovos, internal tax analysts will spend hundreds of hours manually mapping thousands of internal tax codes to specific “boxes” on government forms. But do you really need to pay the enormous amount of money for this?

• The Redesign: Deploy Schema-Mapping Agents. These agents don’t just “fill a form”; they reason about the meaning of the jurisdictional template and the intent of the underlying transaction data. The agent can autonomously establish and maintain data lineage from the ERP to the government form.

• Efficiency Gain: Reduce time spent on return preparation, enabling the team to focus on more strategic work.

The Risks: Avoid the Danger Zone

Of course, these suggested transformations require rigorous guardrails to prevent AI-driven liabilities. Here are some to keep in mind:

• The Hallucination in the Ledger: We want to avoid errors in tax filings. Therefore, agents must operate within deterministic guardrails in which AI proposes the logic, but a human-in-the-loop confirms the result.
  

• Accountability & Decision Trails: Tax authorities require “explainability,” as I’ve written in a previous article. If an agent makes an autonomous $10M tax determination, “the AI did it” is not a defense. You need a clear Accountability Matrix and “Traceability Logs” that record exactly which agent made which decision, based on which specific legislative data point, including the HITL that signed off on the decision.
  

• Regulatory Drift: Agents trained on stale data are liabilities. You need a Continuous Evaluation Suite to autonomously evaluate system outputs against current desired behaviors. Examples include:

  • Tax Code Correctness: Testing the agent against 10,000 historical transactions for which the ground truth is known.

  • Schema Compliance: Validating AI-generated XMLs against the latest government XSD schemas.

  • Adversarial Testing: Attempting to force the agent to apply an invalid tax exemption to test for “hallucination resistance.”

  There are also commercially available frameworks from companies including LatticeFlow, LangSmith, Giskard, and Ragas.

• Privacy & Data Leakage: Sharing PII or sensitive transaction data with third-party agentic AI providers poses a significant risk. You must implement private-tenant deployments or “clean room” data processing layers to ensure that proprietary financial data isn’t used to train public models.
  

• Reliability & Uptime: In a CTC world, an AI outage is a revenue outage. You must address provider lock-in and volatility in model pricing. Reliable uptime requires strategies such as multi-model fallbacks, so that if Agent A (Anthropic) is down or its price spikes, Agent B (OpenAI) takes over.

The Roadmap: Moving from AI Experiment to AI Transformation

To build on the examples above, you need an approach to escape the “micro-productivity trap”. The HBR transformation framework is a good example:

1. Narrow Possibilities Strategically: Don’t sprinkle AI everywhere. Pick 4-5 critical domains (e.g., Software Dev, Tax Determination) and focus on them.

2. Reimagine the Workflow: Do not automate your current process. Take an “outcome-oriented” approach. If you were building a tax department today with 2026 technology, what would it look like? Start there.

3. Engage the Frontline Leaders: Your best “TaxTech PMs” are the people currently doing the manual work. They know where the bottlenecks are.

4. Measure What Matters: Stop tracking “efficiency.” Track what matters and use the dashboard I suggested in a previous article.

References & Sources

• Harvard Business Review: How to Move from AI Experimentation to AI Transformation (Accessed: May 11, 2026)

• Anthropic: Finance Agents: The Next Frontier of Tool Use (Accessed: May 11, 2026)

• PwC US: Tax Function of the Future: Evolution of Data and Analytics (Accessed: May 11, 2026)

• KPMG International: 2026 Global AI in Finance Report (Accessed: May 11, 2026)

• OpenAI: OpenAI for Enterprise: Agentic Workflows (Accessed: May 11, 2026)

• Deloitte: Finance Transformation: Modernizing the Core (Accessed: May 11, 2026)

• Thomson Reuters: Top 4 Tax Technology Trends for 2026 (Accessed: May 11, 2026)

• LatticeFlow AI: AI Model Robustness and Data Diagnostics (Accessed: May 11, 2026)

• LangChain: LangSmith: Trace, Test, and Optimize Agents (Accessed: May 11, 2026)

• Giskard: The First Open-Source Quality Platform for AI (Accessed: May 11, 2026)

• Exploding Gradients: Ragas: Evaluation Framework for RAG Pipelines (Accessed: May 11, 2026)

In the boardrooms of the S&P 500 and high-growth scale-ups, the “common wisdom” is that AI is primarily a productivity booster within specific workflows—a narrative fueled by marketing promises of an “era of builders” where everyone can create at scale. This is not wrong. But there is a twist. In 2026, we have transitioned from highly predictable, established SaaS pricing models to a “tokenized economy,” in which data is the input and output, and compute is the fuel, metered like a utility—a shift that makes real-time AI infrastructure stability as critical to the P&L as the tax logic being executed.

But here is the catch. You cannot manage a 2026 cost structure with a 2016 KPI dashboard. Here is the ground truth for the new KPI framework needed to navigate the tokenized economy today.

I. Latest Trends in AI Tech: The Tokenization of the S&P 500 Cost Structure

1. Enterprise Pricing: The Metered Reality

A look at the pricing models of the top 4 AI platform providers reveals that the era of the “flat-fee seat” is dead. We have entered the era of “utility AI”, in which compute is metered like electricity1:

It’s clear that for the S&P 500—companies managing 100 million to more than 1 billion transactions per year—AI costs are a direct hit to the margins. This is also what Uber’s CTO, Praveen Neppalli Naga, recently confirmed. In this interview, he revealed that the company exhausted its entire 2026 Claude budget by April due to deep agentic integration. Uber's R&D expenses rose 9% to $3.4 billion in 2025, and the company expects that figure to keep climbing—suggesting AI may be as much a cost driver as a productivity lever.

2. The Stability Layer: Reliability Tracking

At the S&P 500 scale, downtime isn’t an inconvenience—it’s a breach of compliance. The gold standard is “four nines” (99.99%), which allows for only 52 minutes of downtime annually. Let’s look at how the AI platform providers are doing in 2026:

None of the major LLM providers currently guarantees 99.99% uptime during peak global cycles. So, any taxtech architecture must have a failover logic to hot-swap providers in milliseconds when a “brownout” occurs.

3. Ownership Disruptions: Related Parties and Consolidation

In 2026, the AI world is one of related parties and imminent consolidation. Here are only the recent signals:

• The SpaceX-Cursor Vertical Integration: SpaceX has secured a $60 billion option to acquire Cursor to fill xAI’s infrastructure gap.

  This signals that the infrastructure used to build AI is as valuable as, if not more valuable than, the models themselves. If your infrastructure is tied to a platform that gets acquired, the impact may not be convenient.

• The OpenAI-Microsoft “Risk Factor”: In recent draft IPO filings, Sam Altman officially labeled Microsoft a “risk factor” due to OpenAI’s deep dependence on their compute.

  When such partnerships signal strategic friction, you must prepare for a change in compute infrastructure that could spike your costs overnight.

• The Anthropic-Amazon “compute debt”: Anthropic and Amazon have committed $100 billion to AWS custom Trainium chips.

  This creates a new vendor lock-in. If you are integrated with Claude, you are now effectively an AWS-dependent entity, regardless of your primary cloud strategy.

In summary, the shifting landscape introduces risks such as:

In other words, if your tax data is latent or your vendor stack is fragile, your AI becomes an expensive way to generate errors at the speed of light—transforming a potential efficiency engine into a high-frequency audit liability.

So, what’s the right course of action to mitigate those risks?

To mitigate these risks, the traditional fintech/taxtech KPI dashboard is no longer sufficient. We are moving toward dynamic, metered infrastructure with significant cost implications — a shift that requires a new set of KPIs to protect the bottom line. Here is how I would stay ahead:

1. No Infinite Spending Loops: Agentic AI can enter recursive loops that burn $100,000 in tokens before you finish your morning coffee. You need hard “circuit breakers” at the API level to stop runaway costs.
   

2. Get a Grasp of ROI: Stop using AI to only write memos or summarize meeting notes. Use it to automate the 4,000 manual monthly reconciliations that keep your accounting team up at night. The highest returns are found in escaping the manual loop, not executive dashboards or piles of memos.
   

3. Own Your Logic Library: If you bake your proprietary tax logic into a specific vendor’s prompt, as Xero did, you are a hostage to their uptime and pricing. Build for model agnosticism so you can migrate from OpenAI to Gemini in under 24 hours.
   

4. Audit for Infrastructure, Not Just Model Performance: With SpaceX securing a $60 billion option to acquire Cursor, the infrastructure that builds the code is now as valuable as the AI itself. Ensure your vendors are built on hyperscale foundations (Google/AWS/Azure) that can actually handle your transaction volume.

II. The S&P 500 TaxTech Dashboard

For the Chief Tax Officer or VP of Engineering, the greatest risk in 2026 isn’t a single audit; it’s invisible inefficiency. At the S&P 500 scale, you cannot manage what you do not measure. Therefore, we need KPIs that move beyond “monitoring” to “ROI orchestration,” i.e., actively managing the unit economics of compliance. Your dashboard should be more than a status report—it should be a decision engine that incorporates AI efficiency.

Here's my proposal for a scalable TaxTech Dashboard at the S&P 500 level:

Along with the traditional metrics such as Effective Tax Rate, Compliance Spend, and % of Manual Journal Entries, TaxTech leadership must follow some new KPIs:

1. AI Pricing Drift (Delta %)

With the end of the “flat-fee era,” the AI costs are now as volatile as energy prices. This KPI tracks the per-token spend variance. For example, if weekly drift exceeds 5%, the dashboard triggers an automated “model swap,” routing low-complexity tasks to more cost-effective tiers.

2. AI Decision Accuracy (%)

In 2026, manual troubleshooting is the silent killer of scale. This KPI tracks the percentage of tax triages handled by AI without human intervention. If it dips below 98%, for example, the dashboard should automatically increase the “human-in-the-loop” sampling rate to prevent systemic audit risks.

3. Token Effectiveness Ratio (TER / 1M tokens)

This KPI identifies exactly how many tokens are converted into defensible compliance outcomes versus the compute waste generated by recursive loops like API validation ping-pong (endless retries against failing portals), consensus conflict (agents arguing over logic until quotas are exhausted), or hallucination spirals (expensive internal reasoning used to resolve contradictory data). By tracking TER, budget owners can detect these “inference leaks” before they decimate quarterly results, ensuring the system maintains a defensible 1:1 ratio rather than slipping into the 1:1000 efficiency abyss that defines unmanaged infrastructure.

4. AI Latency & Uptime

In the fintech world, 5 seconds is an eternity. This metric tracks the milliseconds between a transaction and its tax-defensible status. Any dip below 99.99% reliability triggers a failover to a secondary sovereign cloud instance.

5. CAPEX Token Allocation

In the tokenized economy of 2026, CAPEX Token Allocation (ASU 2025-06) is extremely important. Following the 2025 FASB update regarding AI software capitalization, organizations must now surgically differentiate between tokens consumed for “creation”—the R&D and asset-building phases that qualify as capitalizable investments—and those used for “maintenance”, which hit the books as immediate operational spend. By monitoring this metric through real-time financial telemetry, budget owners can meet specific asset capitalization targets, thereby demonstrating to auditors that their AI spend is building a long-term intangible asset rather than merely subsidizing an expensive, ongoing maintenance loop.

Strategic Forecast

Over the next 24 months, the most successful FinTax leaders will shed their traditional “Head of Tax” skins and re-emerge as VPs of Compliance Infrastructure. In this new reality, the silo between the tax department and the platform engineering team will vanish entirely, replaced by a shared mandate: the real-time management of compliance telemetry. The standard for success will shift from “getting it done” to algorithmic efficiency. Expect to see:

• The Rise of the TER Audit: The Token Effectiveness Ratio (TER) will shift from a niche DevOps metric to a primary audited financial disclosure. The market will begin to value AI-enabled firms by their compute-to-compliance efficiency, with a ratio approaching $TER = 1:1$ (one successful outcome per unit of spent compute) representing a competitive moat, while anything higher signals an unmanaged inference leak.

• Drift-Triggered Rerouting: Pricing drift will become the new “interest rate risk.” Companies will deploy automated arbitrage bots that “hot-swap” compliance logic across LLMs (OpenAI, Anthropic, or Gemini) in milliseconds, based on real-time price-per-token volatility and latency performance.

• The End of the Monthly Close: As AI decision accuracy stabilizes above 99%, the concept of a monthly close will become a legacy term. We are moving toward “continuous compliance”, where tax filings are an automated, real-time byproduct of a data stream, and audit readiness is a constant state rather than a quarterly fire drill.

If you aren’t metering your way to ROI today, you aren’t just falling behind—you are accruing massive technical and fiscal debt that no amount of human “productivity” will be able to pay off.

FAQs: The 2026 Tokenized Economy

What is the “Tokenized Economy” in 2026?

The tokenized economy represents a shift from flat-fee SaaS pricing to a metered “utility AI” model. In this framework, data acts as the input/output, while compute is treated like electricity—metered and billed based on consumption. This makes real-time infrastructure stability a critical factor in a company’s P&L.

Why is the “flat-fee seat” model dying for enterprise AI?

As companies like Uber integrate deep agentic workflows, the volume of transactions (often exceeding 1 billion annually) makes per-user pricing unsustainable for providers. Utility AI allows providers to bill for the actual compute consumed, ensuring margins are protected as agentic complexity scales.

What is the Token Effectiveness Ratio (TER)?

The Token Effectiveness Ratio (TER) is a KPI that measures the ratio of AI tokens converted into defensible outcomes to the amount of waste. A 1:1 ratio is the gold standard, while higher ratios indicate “inference leaks” caused by recursive loops, hallucinations, or redundant API validation.

How does ASU 2025-06 affect AI spending?

Under the ASU 2025-06 FASB update, organizations must surgically differentiate between tokens used for “creation” (capitalizable R&D) and “maintenance” (operational spend). CAPEX Token Allocation allows companies to build long-term intangible assets on their balance sheets rather than just incurring ongoing expenses.

What are the main risks of AI infrastructure fragility?

The primary risks include pricing drift (volatility in token costs), infrastructure fragility (service outages during peak periods), and model latency. In 2026, a 5-second delay in AI processing can lead to blocked checkouts or failures in real-time tax compliance reporting.

How can companies achieve “model agnosticism”?

To avoid vendor lock-in and mitigate downtime, firms must build a proprietary logic library independent of any single LLM. This allows them to “hot-swap” between providers such as OpenAI, Gemini, or Anthropic in under 24 hours, based on real-time pricing and performance telemetry.

What is “Continuous Compliance”?

Continuous compliance is the 2026 standard, where tax filings are a real-time byproduct of automated data streams rather than a monthly manual “close.” As AI decision accuracy stabilizes above 98%, audit readiness becomes a constant state, eliminating the need for traditional quarterly fire drills.

References & Sources

• Anthropic Pricing: Anthropic Shifts Enterprise Billing to Per-Token Pricing (Implicator.ai, April 14, 2026).

• Uber AI Budget: Uber CTO Shows Claude Code Can Blow AI Budgets (The Information, April 10, 2026).

• SpaceX-Cursor M&A: SpaceX Secures Right to Acquire Cursor for USD 60 Billion (LatestLy, March 28, 2026).

• OpenAI: Official API Pricing Guide (Accessed April 22, 2026).

• Google Gemini: Vertex AI Pricing Overview (Accessed April 22, 2026).

• Vertex Resource Library: AI and Indirect Tax Trends (Accessed April 22, 2026).

• PwC Tax Tech: Unlocking Value in Tax Technology (Accessed April 22, 2026).

• Platform M&A & Infrastructure: Musk’s SpaceX bets $60 billion on Cursor to fix xAI’s coding gap (The Decoder, April 2026).

• Accounting Standards (ASU 2025-06): Accounting Best Practices for AI Use in Software Development (Broadcom/Clarity, April 2026).

• Unit Economics & Pricing: Anthropic Shifts Enterprise Billing to Per-Token Pricing (Implicator.ai, April 14, 2026).

• Performance & ROI: Want ROI from AI? Go for growth (PwC Global, April 2026).

• System Reliability Benchmarks: Real-Time Availability & Status Monitor (OpenAI, April 2026) and Anthropic System Status (Anthropic, April 2026).

• Sovereign AI & Compliance Risk: The Sovereign AI Boom and Trillion-Dollar Predictions (Motley Fool, April 2026) and Legal Analysis of AI Compliance Risks (Northman & Sterling, March 2026).

Some S&P 500 CFOs and Chief Tax Officers may take the obvious path, treating AI as a procurement exercise: buy the license, schedule mandatory training, set KPIs for usage, and wait for ROI. After all, this approach worked for other challenges, such as e-invoicing or the adoption of last year’s new accounting/tax software.

It won’t come. At Uber, I’ve seen that you cannot bridge the 2026 “fluency gap” with a three-month seminar. Nor with mandatory usage metrics.

The Market Mandate for AI-Native Fin/TaxTech Talent

The industry standard has shifted. My analysis of more than 100 current openings across the S&P 500, high-growth scale-ups such as OpenAI, Anthropic, and Lovable, the Big Four, and traditional SaaS vendors such as Xero, Intuit, and Brex reveals that leading firms are bypassing the “training phase” and hiring for AI fluency as a day-one prerequisite.

We are seeing a structural shift in the “Tax/FinTech Lead” role. In other words, there are new capabilities and requirements being sought for high-stakes transformations:

• OpenAI is recruiting a Tax Technology Lead to design tax as a “Production System,” integrating compliance directly into core product infrastructure.

• PwC has shifted from “upskilling” to hiring AI-Native Engineering Leads who must demonstrate daily proficiency with AI coding assistants.

• Intuit is aggressively hiring for AI Workforce Transformation to lead the shift toward AI-first developer velocity.

• Anthropic and Lovable are seeking AI-native operators who use agentic workflows to handle real-time Finance & BI decision velocity.

• EY and Grant Thornton are targeting Tax Automation & AI Managers to build “AI-enabled solutions” for global VAT and indirect tax reporting.

So, what do you need to do?

The FinTax AI Fluency & Hiring Rubric

To move your organization from “pro forma” adoption to “front-line” efficiency, you need to update your hiring rubric. Here is Zapier’s hiring rubric, designed to bring AI-native talent in-house with every hiring decision. It’s a good example.

Here is what a Fin/TaxTech hiring rubric may look like, along with examples of the expected behavior, potential gain, and hiring filter questions to ask:

• Stage 1: The Consumer (Passive Use)

  • Behavior: Using public LLMs to summarize tax memos.

  • Potential Gain: 5-10% productivity lift on administrative tasks and gross time savings of ~5 hours per week.

  • Hiring Filter: Ask: "Explain the difference between an LLM and a database." If they view AI as a "search engine" rather than an "execution engine," they are liable for hallucinations.

• Stage 2: The Auditor (Problem Decomposition)

  • Behavior: Reviewing AI-generated reconciliations. They understand the “why” behind the logic.

  • Potential Gain: 20-30% reduction in manual review time and a 15-25% reduction in manual errors.

  • Hiring Filter: Ask them to decompose a complex tax problem into prompts. If they can’t break down the workflow, they can’t create working AI flows.

• Stage 3: The Orchestrator (Agentic Design)

  • Behavior: Designing workflows where AI agents handle the “first mile” (data ingestion) and “last mile” (filing).

  • Potential Gain: Teams hiring orchestrators see a 3.5x higher ROI than those stuck in “training mode.”

  • Hiring Filter: Ask: "Walk me through a multi-agent workflow you've designed to automate a high-volume tax workstream. How did you structure the 'human-in-the-loop' intervention points?"

• Stage 4: The Architect (Tax Product Lead)

  • Behavior: Building a reusable logic library for 100+ jurisdictions.

  • Potential Gain: 40-60% cycle-time reduction on global filings and the ability to launch in new markets in days, not months.

  • Hiring Filter: Ask: "How would you structure a global tax engine to ensure compliance logic is a 'production system' integrated into the product roadmap, rather than a stand-alone 'to-do' list for the tax team?"

Summary

To scale a FinTech or TaxTech function in 2026, you must stop treating AI as a procurement project and start treating it as a hiring mandate.

References & Comprehensive Job Analysis (100+ Openings)

• Deloitte, “The State of AI in the Enterprise 2026: Strategy for the S&P 500”, April 2026. deloitte.com/state-of-ai-2026

• DualEntry, "AI in Accounting: The Complete 2026 Guide", April 2026.

• Market Growth Reports, "Artificial Intelligence in Accounting Market Size and Trends Research [2035]", February 2026.

• Thomson Reuters, “2026 Corporate Tax Technology Report: The Rise of Agentic Tax Tech”, March 2026. tax.thomsonreuters.com/corporate-tax-technology-2026

• Zapier, “Raising the AI Fluency Bar in Hiring”, March 2026. zapier.com/blog/raising-ai-fluency-bar-in-hiring

• PwC, “From Compliance to Intelligence: The 2026 Global Tax Outlook”, February 2026. pwc.com/global-tax-outlook-2026

• Microsoft Official Blog: New IDC study shows AI’s real-world ROI (November 1, 2023)

• Direct Link to the IDC InfoBrief: The Business Opportunity of AI

• OpenAI Careers, Tax Technology Lead, April 2026. openai.com/careers/tax-technology-lead

• PwC Careers, AI-Native Engineering Lead, April 2026. jobs.us.pwc.com/ai-native-lead

• Intuit Careers, AI Workforce Transformation, March 2026. jobs.intuit.com/ai-workforce

• Lovable Careers, Finance & BizOps Strategic Partnerships, April 2026. lovable.dev/careers/finance-bizops

• Anthropic Careers, Finance Analytics & BI, April 2026. pitchmeai.com/anthropic-finance

• EY Careers, Tax Technology and Operations, February 2026. ey.com/tax-tech-jobs

• KPMG Careers, Tax Transformation Specialist, 2026. kpmg.com/tax-transformation

I have been watching the rapid rollout of AI assistants across the tax/accounting/finance tech stack—from Brex Assist and Xero’s JAX to Intuit Assist. These tools are incredible at turning fragmented, manual processes into centralized, automated engines.

But amidst these rapid developments, I've been asking the same question: if your AI “hallucinates” a €10M tax deduction based on a misinterpretation of a cross-border VAT treaty, who signs the penalty check?

So, I decided to analyse the consequences focusing on two aspects:

1. What are the current guidances on using AI for tax reporting from Tax Authorities around the world? (note: of course, my research is not tax advice and is based on publicly available sources. Consult tax professionals for further guidance.)

2. How should product manager, engineers, and tax managers approach incorporating these guidances when integrating AI into critical systems?

Global Research Summary: AI Liability & Tax Compliance (2026)

To answer the first question, I dove into publicly available, country-specific guidances. Here are some observations from across the globe based on publicly available articles.

1. The UK & EU: The “Liability Hammer”

• United Kingdom: The FRC’s March 2026 guidance is the global gold standard—audit partners and directors are 100% accountable for AI outputs. The precedent set in Gary Elden v HMRC (2026) proves that “unchecked AI use” is now legally categorized as professional incompetence. As the regulator says: “You can’t blame it on the box.”

• European Union: As of August 2026, the EU AI Act is fully enforceable. Penalties for “high-risk” financial AI errors can hit €35M or 7% of global turnover.

• France & Poland: France’s 2026 e-invoicing mandate and Poland’s KSeF system have removed the buffer. Taxpayers—not their software providers—remain solely liable for every data point transmitted.

2. APAC: Governance of “Agentic” AI

• Singapore: Budget 2026 introduces a 400% tax deduction for AI expenditure, but with strings attached. Singapore has launched the world’s first Model Governance Framework for Agentic AI, emphasizing that “meaningful human accountability” must remain central to deployment.

• Japan: The 2026 Tax Reform incentivizes AI R&D but tightens intra-group transaction documentation. The NTA is moving toward Tax Administration 2.0, using AI to automatically match NTA data to identify taxpayer errors.

• New Zealand: Inland Revenue (IRD) has confirmed that “human in the loop” oversight is mandatory for all AI workflows.

• Australia: The ATO now requires an AI Transparency Statement from all entities that use AI in tax-affecting decisions.

3. LATAM: Transparency & “Risk-Based” Compliance

• Brazil: The new Taxpayer Protection Code (TPC) introduces the “habitual tax debtor” status. If your AI creates unjustified non-payment, you lose tax benefits.

• Mexico: The 2026 Master Plan targets improper deductions identified by SAT’s AI. If your system cannot prove the “materiality” of a transaction, the deduction is denied.

• Argentina: Law No. 11,089 mandates that any AI used for tax settlement must be subject to external, independent audits.

4. North America: The Enforcement Surge

• USA (IRS): The IRS has expanded its AI use cases to 126, with a focus on large partnerships. They use AI to catch the “hallucinations” in your filings. Without contemporaneous support, your AI-driven deduction is a red flag.

• Canada: The CRA 2026-27 Plan treats AI errors the same as manual ones.

Based on this research, here is a summary of where global authorities sit on the “AI Sentiment Spectrum”:

Guide to the Spectrum:

• Zone 1: Innovation-First

  • Core Philosophy: AI is viewed as a national competitive advantage. Regulators prioritize adoption and “safe harbor” provisions for early movers.

  • Examples: In Singapore, Budget 2026 offers massive tax deductions (up to 400%) for AI implementation, provided companies follow the Model Governance Framework for Agentic AI.

  • The Trade-off: High transparency is expected, but the focus is on “Cooperative Compliance” rather than punitive measures.

• Zone 2: Balanced Scrutiny

  • Core Philosophy: “Trust, but verify.” These authorities use AI aggressively for their own enforcement (e.g., the IRS’s 126 use cases) and expect taxpayers to maintain a high bar of documentation.

  • Examples: Australia (ATO) and the USA have effectively mandated “Contemporaneous Support.” If an AI picks a deduction, you must be able to surface the underlying data trail immediately.

  • The Trade-off: No special penalties for AI errors yet, but standard penalties are applied strictly with a lower tolerance for “clerical error” excuses.

• Zone 3: Liability-First

  • Core Philosophy: “The AI black box is no defense”. Algorithmic mistakes are treated as human professional negligence.

  • Examples: The EU AI Act (enforced Aug 2026) imposes fines up to 7% of global turnover. In the UK, the FRC’s stance is that a hallucination is proof of a lack of professional skepticism.

  • The Trade-off: Speed is secondary to Auditability. In these regions, a “Black Box” AI system is essentially a ticking legal time bomb for a CFO.

So, what do we do to prepare for this seismic shift? In the upcoming sections, I am giving some thoughts on how to prepare.

1️⃣ The Tech Maturity Gap: Breaking the “Tech Stack Prison”

In my experience building at S&P 500 scale, the biggest risk isn’t the AI itself—it’s the Maturity Gap. According to the Thomson Reuters 2026 Corporate Tax Department Technology Report, the term describes the relative advancement of a company’s tech stack through the different stages of technological maturity, which refers to the typical steps that a corporate tax department travels through on its technological journey toward a more proactive approach to tax work:

According to the report, 55% of the surveyed companies are stuck in Phase 2: Reactive. This stage includes:

• Using tax department databases, some 3rd party software, with some auto mated feeds, but not connected to enterprise data or departments across the company.

• Tech budget shared with multiple departments

• Hired a tech consultant or outsourced tech resources

• No technology training provided

• Needs investment in process automation/tech solutions for most time-consuming activities (tax data management, transfer pricing, FATCA/CRS/AEOI )

• More robust metrics needed to measure succes

In essence, these companies have an outdated tech stack which makes any scalable AI usage impossible. To escape this “tech stack prison,” they must move up the Technological Maturity Curve to get to Optimized & Predictive (Stages 4-5). In those, the tax logic is a standalone platform capability. This is the only way to scale sustainably.

2️⃣ The Audit Trap: Defining “Reasonable Care” in a Neural Network

How do you prove “reasonable care” when your tax logic is buried in a neural network?

In a traditional audit, you show your work via spreadsheets or law citations. In an AI world, regulators are shifting the definition of “care” from what you filed to how you governed the system that filed it.

As the ACCA’s 2026 Topical Guidance notes: “Members should avoid overstating the accuracy of the data output... Undertaking appropriate due diligence on the output provided by an AI tool can help safeguard against the risks.”

The “A.C.E.S.” Framework for Defensible AI

To get a step closer to complying with the updated reasonable care definition, I suggest the A.C.E.S. Framework—a four-pillar approach to ensuring any AI is audit-ready and legally defensible:

3️⃣ The Penalty Reality: Moving to “Rules as Code”

If you are sourcing a system today, move beyond the “AI” hype. Tax requirements shouldn’t be a one-off ticket; they need to be a reusable logic library.

As organizations like the OECD and OPSI have championed in their Rules as Code (RaC) frameworks, the goal is to transform legislation into machine-consumable code. To incorporate this at your enterprise, I suggest the you consider the following:

1. Rule Atomization: Decouple tax logic from your ERP. Store rules in a modular “Source of Truth” (using engines like OpenFisca or Drools).

2. API-First Consumption: Expose these rules via a Tax Service Layer. Any system—whether it’s a checkout page or a GenAI assistant—must call this central library for a deterministic check.

3. Governance Wrappers: Wrap AI probabilistic insights within human-defined, deterministic rules to ensure an immutable audit trail.

In case you source your tax reporting software from a 3rd party, make sure those are in place before signing a contract.

🎯 The Bottom Line: Strategic Mandates for Defensible Tax Reporting Using AI

The era of “black box” tax automation is officially over. To reach the next level of global digital transformation and mitigate the €35M risk, senior leadership must adopt the following three mandates:

I. Accountability: Human-Centric Liability

“You cannot blame it on the box.” Regulatory mandates in the UK (FRC) and the EU (AI Act) have made it clear: liability for AI hallucinations sits squarely with human partners and company directors. Neither the software vendor nor the AI model can hold the burden of proof. Organizations must move from passive consumption of AI outputs to active, documented oversight.

II. Architecture: Transitioning to the “Predictive” Stage

The “Tech Stack Prison” is one of the greatest barrier to AI-driven efficiency. Companies stuck in Stage 2 (Reactive) lack the data integration required to validate AI insights. To scale sustainably, the tax logic must be decoupled from the ERP and implemented as a standalone platformcapability. This shift allows for the validation of probabilistic AI guesses against hard-coded legal rules in real-time.

III. Defense: Re-Defining “Reasonable Care”

In a traditional audit, care is shown through spreadsheets. In an AI-native world, auditors will inspect your governance. Institutionalizing AI governance through a framework such as A.C.E.S. or another is the only way to satisfy the “Reasonable Care” requirements of modern tax authorities. Without a framework that allows you to “replay” the logic of a 2024 transaction in a 2026 audit window, your AI-driven deductions will be treated as red flags.

References & Sources

• OECD OPSI: Rules as Code: A Case Study in Public Sector Innovation (2024-2026)

• Thomson Reuters Institute: 2026 Corporate Tax Department Technology Report (March 2026)

• Financial Reporting Council (UK): Guidance on AI Accountability (March 2026)

• Ross Martin Tax: Gary Elden v HMRC (2026) Case Summary (Jan 2026)

• K&L Gates: Singapore’s New Model AI Governance Framework for Agentic AI (Feb 2026)

• Alvarez & Marsal: Singapore Budget 2026 Spotlight (Feb 2026)

• EY Japan: 2026 Japan Tax Reform Highlights (Feb 2026)

• IMF/NTA Japan: Tax Administration 2.0 Digital Transformation (2026)

• Inland Revenue NZ: OIA: Use of AI at IR (Jan 2026)

• ATO Australia: AI Transparency Statement (2026)

• BDO Global: Brazil Taxpayer Protection Code 2026 (March 2026)

• Olshan Law: IRS AI Use in Tax Enforcement (March 2026)

• Baker McKenzie: EU AI Act Compliance Radar (Accessed April 2026)

• OECD: Due Diligence Guidance for Responsible AI (Feb 2026)

• ACCA Global: Application of PCRT to AI Ethics (Jan 2026)

Imagine that a Tax, Accounting, and Finance team consists of 100 people. 80 of them are doing work a script could do in 4 seconds. Why?

Because that team is trapped in what I would call a “Tech Stack Prison.” According to the newly released March 2026 study by Thomson Reuters, tax teams are trapped extracting data from fragmented ERPs, shared drives, and disparate CRMs. Tax teams that eliminate these structural, manual bottlenecks reduce preparation time by 50%—reclaiming more than 10,000 hours annually. The report also highlights that human error at scale is the silent killer of enterprise agility. A single decimal-point error, multiplied across millions of global transactions, doesn’t just result in an overpayment; it also triggers audits and blocks new market expansion. A single decimal-point error, multiplied by a quarter, creates massive financial liability.

If you lead Finance or Tax at an S&P 500 company or a hyper-growth scale-up like Anthropic or OpenAI, your team already knows the law. They understand the regulations natively. What they probably lack is the technical architecture to execute that knowledge at scale.

This is where the “Compliance Brain” needs the “Product Heart.”

To prevent compliance from blocking top-line growth, Finance and Tax leaders must stop thinking like back-office administrators and start operating like Product Managers. Here are five architectural strategies to break your team out of the tech stack prison:

1️⃣ Escape the Manual Loop via Automated Pipelines

Don’t drown your top talent in arduous reconciliations. The modern standard is to partner with your product and engineering teams to replace fragmented ERP data extraction with automated API pipelines. By feeding a centralized data warehouse, you establish a single source of truth. You turn a chaotic, manual data-gathering exercise into a centralized, automated engine that provides auditability in days, not months.

2️⃣ Eradicate the Spreadsheet Version Nightmare

A comprehensive Bloomberg Tax survey found that 92% of corporate tax professionals view their workpapers as critical. Yet, most rely on legacy spreadsheets that lack cell-level tracking and robust access controls. Moving out of the tech stack prison means adopting systems with embedded native controls. You need automated roll-forwards, version control, and cell-level audit trails so that “version_final_v4.xlsx” never compromises your compliance posture again.

3️⃣ Data Workings at S&P 500 Scale

What I learned while scaling global operations at Uber is that compliance isn’t just about calculating the tax rate; it’s about the data trail. Today, AI can identify anomalies in millions of transactions and generate reporting at a massive scale. But modern tax and finance tech shouldn’t just flag an error; it should simulate the fix upstream.

Shift your system from reactive error-catching to proactive, systemic resolution.

4️⃣ Move from Historical Reporting to Predictive Exposure

As outlined in EY’s Global Tax and e-Invoicing strategies, governments are shifting to near real-time compliance (like the EU’s ViDA or LATAM’s continuous transaction controls). They want your data before the transaction is finalized. Your tech stack must evolve from historical reporting to predictive analytics. AI should be leveraged to identify tax exposure and nexus risks before they materialize, allowing you to build compliance roadmaps that accelerate—rather than delay—business expansion.

5️⃣ Tax as a Product, Not a Project

Tax requirements shouldn’t be a one-off IT ticket every time a regulation changes. You need to build a reusable logic library. When your company decides to expand into its next market, you shouldn’t have to rebuild your tech stack from scratch. A modular, product-led architecture allows you to enter 100+ jurisdictions seamlessly because the foundational logic is already centralized and adaptable.

For modern enterprises, compliance is merely the baseline. Efficiency, real-time insight, defensible data trails, and enabling business outcomes are the true prizes.

Take a hard look at your current infrastructure: Is your tax logic a standalone “to-do” list, or is it baked directly into your core product roadmap?

Sources

For those looking to dive deeper into the architectural shifts and data driving this transformation, I highly recommend reviewing the foundational research:

• Thomson Reuters: 2026 Corporate Tax Department Technology Report (Primary data source on AI integration timelines and the 10,000-hour reclamation metric).

• EY (Ernst & Young): Global Tax e-Invoicing & Digital Reporting Solutions (Context on the shift toward near real-time, predictive tax authority interactions).

• Bloomberg Tax: Corporate Tax Department Benchmark Report (Data regarding the 92% reliance on legacy workpapers and spreadsheet vulnerabilities).

It looks like in a world governed by binary tax code and rigid regulatory frameworks, we are still burning through thousands of human hours on manual reconciliations.

The Structural Collapse of the Pyramid

For decades, Tax departments were built as pyramids: a massive base of junior staff handling the “routine banality” of data entry. That model is dead.

As PwC recently highlighted, we are shifting to a diamond-shaped talent structure. The base of manual preparation is shrinking. What is expanding is the center: a core of Financial AI Agent Orchestrators. In this new model, the "middle" expands with skilled professionals who don't prepare tax returns but orchestrate the AI agents that do.

The Shift: From “Worker” to “Orchestrator”

In a deterministic world, the “What” (the law) is fixed. The “How” (the implementation) is where the S&P 500s lose millions in efficiency. Here is how the high-growth tax team must evolve:

• From Data Entry to Logic Design: Instead of fixing errors in a spreadsheet, the Orchestrator designs the logic gates that prevent the error from ever hitting the ERP.

• From Reporting to “AI Ops”: We don’t need people to generate reports. We need professionals who can troubleshoot the automated engine, ensuring AI agents don’t hallucinate jurisdictional logic.

• From Analysts to “Product Managers for Tax”: The most valuable hires won’t be tax specialists; they will be “multidisciplinary units” who can translate a regulatory update into a technical roadmap.

The 2026 Mandate for Tax Leaders

If your team is still spending 80% of their time on work that a script could do in 4 seconds, you are building a massive “human-heavy debt”. The world is moving to Continuous Transaction Controls. You cannot “human” your way through a 5-second API validation window.

Better start planning your tech stack and team skills now. The alternative is a “Compliance Brain” that eventually chokes your “Business Heart.”

Sources:

• Thomson Reuters: 2026 Corporate Tax Technology Report

• Deloitte: Tax Insights and Perspectives

• PwC: Tax Technology Transformation

• EY: Tax Technology & Data Strategy

According to the Thomson Reuters Institute’s 2026 Corporate Tax Department Technology Report, 61% of tax departments surveyed are ignoring the only metric that actually matters to the C-suite: business outcomes. Instead, they are still measuring themselves by “Tax Internal” yardsticks. The survey shows they prioritize time savings (66%) and accuracy (55%):

To avoid basing my conclusions on a single report, I had a quick look at what the “Big Four” think about this. The consensus is that measuring “adoption” or “accuracy” alone is no longer sufficient; tax departments must now prove their impact on the enterprise P&L and growth speed.

Main Conclusions:

• PwC (AI-Driven Value): Predicts a shift from “ground-up” crowdsourced AI to “top-down” enterprise strategies. Metrics are moving toward “proof points” that link AI-driven tax workflows to market differentiation and P&L impact.

• KPMG (Value Driver Trees): Recommends identifying “functional levers” that impact the broader organization, such as optimizing indirect tax cash flow and minimizing irrecoverable taxes to boost working capital.

• EY (Controversy & Policy): Focuses on “Tax Risk & Controversy” as a business metric. They suggest that “Time to Resolution” and “Data Readiness for Pillar Two” are critical KPIs to prevent tax from becoming a bottleneck during global expansion.

• Deloitte (Resource Mix): Suggests that the “Optimal Resource Mix” (balancing AI, outsourcing, and in-house talent) should be measured by its ability to free up the tax team for “high-value, strategic activities” rather than just cost reduction.

Don’t get me wrong—accuracy is a baseline requirement. It’s the “Compliance Brain” doing its job. But if your tax tech stack only solves for accuracy, you aren’t a strategic partner; you’re an insurance policy.

To move from being a cost center to a value driver, we need the “Product Heart.”

The “Tech Stack Prison” of Internal Metrics

At the S&P 500 scale, “efficiency” is a generic goal. If you save 1,000 hours but your tax logic blocks a new market launch for three months, the net value to the business is negative.

When tax tech is built as a standalone “to-do” list rather than being baked into the product roadmap, it becomes a blocker. We focus on the data trail for the audit we might have in three years, while the business prioritises the checkout completion rate today.

The Pivot: Adopting “Business-First” OKRs

Without a doubt, Tax Departments must bridge the gap between “tax internal” metrics and “business outcomes.” But what can those be?

To align with the “Product Heart” philosophy, these OKRs are categorized by Compliance Hygiene (Efficiency) and Strategic Enablement (Business Outcomes).

Category 1: Business Enablement & Speed (The “Product Heart”)

• Objective: Accelerate global expansion and product agility through tax automation.

  • KR 1: Reduce “Tax-to-Market” Latency (the time to configure tax logic for a new jurisdiction) from X weeks to < 48 hours.

  • KR 2: Achieve a 0% Tax Latency impact on checkout completion rates by moving tax calculation logic to the “edge” or a real-time engine.

  • KR 3: Increase the “Tax Incentive Capture Rate” (e.g., R&D credits, OB3 Act benefits) by 15% through automated data extraction.

Category 2: Cash Flow & Financial Performance

• Objective: Maximize free cash flow and EBITDA through strategic tax optimization.

  • KR 4: Increase Free Cash Flow by $X million through real-time VAT/GST recovery automation.

  • KR 5: Maintain an Effective Tax Rate (ETR) variance of < 1% between forecast and actuals, enabling more predictable capital allocation.

  • KR 6: Reduce Irrecoverable Tax as a percentage of total spend by X% via improved vendor onboarding logic.

Category 3: Compliance Hygiene & AI Maturity (The “Compliance Brain”)

• Objective: Transition to an “agentic” tax operating model to eliminate manual loops.

  • KR 7: Automate 95% of standard tax reconciliations using AI agents, reducing manual FTE hours by X%.

  • KR 8: Achieve Audit-Ready Status within 24 hours for any jurisdiction via a centralized “Tax Data Lake.”

  • KR 9: Maintain a 0% Audit Penalty rate on all “real-time reporting” mandates (e.g., EU e-invoicing).

If you want a seat at the table where the CEO decides on the next $1B investment, your tax tech needs to speak the language of the business.

Sources:

• 2026 Corporate Tax Department Technology Report | Thomson Reuters

• Tax changes: A strategic look ahead to 2026 | Thomson Reuters Institute

• Tax KPIs & Benchmarks | KPI Depot

• PwC 2026 AI Business Predictions: Delivering Measurable Outcomes

• KPMG: Key Performance Indicators Driving Indirect Tax Value

• EY 2026 Global Tax Policy and Controversy Outlook

• Deloitte Tax Transformation Trends 2025: Rising to Meet the Moment

• PwC Value Creation: Linking Tax Operations to Enterprise Growth

In the world of S&P 500 scale and high-growth scale-ups, the line between “Tax Compliance” and “Revenue Operations” has officially vanished.

We’ve entered the era of Continuous Transaction Controls (CTC). In this world, tax determination isn’t a monthly reconciliation task; it is a live dependency baked into your checkout flow.

The Hidden Risk: Compliance as a Single Point of Failure

The stakes in 2026 are binary. If the API to a tax authority hangs for 5 seconds, your checkout hangs for 5 seconds. In the mobility world, 5 seconds is an eternity—it is the difference between a completed ride and a frustrated user switching to a competitor.

If your data is flawed or your integration is “brittle,” your transactions won’t just be misreported; they will fail to complete entirely. You haven’t just added a compliance step; you’ve added a single point of failure to your revenue stream.

Based on my experience at Uber, here is how you move to a resilient, product-led tax infrastructure:

1️⃣ Robust, Real-Time Calculation: Your tax engine must handle complex logic—location-based rules, exemptions, and thresholds—in milliseconds. If your calculation can’t keep pace with a high-concurrency order service, your architecture is already broken.

2️⃣ Asynchronous Orchestration: Does the transaction move forward while the tax validation runs in a prioritized queue? Or is the user staring at a loading spinner? Build your user flows with contingency in mind. At Uber, we prioritize the user experience while ensuring the tax logic catches up in parallel.

3️⃣ Continuous Data Quality (DQ) Checks: Don’t wait for an audit to find a schema drift. Schedule automated DQ checks to catch mapping errors before the data ever hits a tax authority. At Uber, data points are checked continuously against external databases to ensure success before the API call.

4️⃣ The “Offline Mode” Fallback: What happens when a government server goes dark? You need fallback logic that preserves data integrity without killing the UX. Invest in smart retry mechanisms. At Uber, we invest a lot in smart retry mechanisms. Nothing else will scale in this case.

5️⃣ Agentic Troubleshooting: Manual troubleshooting is the silent killer of the S&P 500. When a document fails, you need an AI Agent that can trace the error, suggest the fix, and re-trigger the submission. This is a deterministic task—perfect for AI agents, not for your high-priced tax team.

The Takeaway for Leaders: Your goal is to build tax technology that never blocks business operations. Is your tax logic a standalone project, or is it baked into your product roadmap?

Stop treating Tax as a project. Start treating it as a Product.

Sources:

• OECD Guidance: Digital Continuous Transactional Reporting (DCTR) for VAT

• Fonoa Blog: Real-Time Reporting and the Shift to CTC

• Deloitte Insights: Strategic Moves: Navigating the Age of Agentic AI

• Vertex Resources: Overcoming SAP S/4HANA & E-Invoicing Challenges

• VATupdate: Daily Global VAT & CTC News