Hello Diyan. Thank you for publishing no BS TaxTech as there is a lot of BS out there.
Your effort/ROI matrix is the part worth sitting with longer than the connector counts. VAT reconciliation at 9.0 ROI, cash application at 9.3 — those numbers only clear if the underlying system was built to be called by an agent in the first place. Most of what's getting built right now is the opposite: a legacy tax engine with a chat UI stapled on, then months of integration work teaching an LLM to navigate a workflow designed for a human clicking through screens. That's where your $60K–$250K+ enterprise tier really comes from — the translation layer, not the tax logic.
That's the problem we started https://orchestrate.tax to solve. The orchestration layer is API-first and MCP-callable from day one — the same operations a controller triggers from a dashboard are what an agent calls directly, same audit trail, same governance, no translation layer in between. XMPT (exemption certificates) and TAXDAI (indirect tax intelligence) sit underneath as agent-callable services rather than siloed apps to reverse-engineer.
Practically, that collapses your build tiers rather than adding a fourth — a VAT reconciliation workflow becomes composing existing primitives instead of a $100K custom build.
One pushback: build-vs-buy still treats "buy" as one category. Buying a platform whose primitives were built for agents and buying a seat license for a platform now marketing itself as agentic aren't the same purchase, even if the sales deck reads the same. Worth a column in the next version of this map.
@Michael Roytman It’s be easier to understand and discuss if we don’t use AI-generated comments. I may be wrong, but yours really looks like Ai-generated. Just a thought
no worries. Or maybe we are "nearing the time" that no one and nothing can be trusted with very few humans actually perform analytical research, empirical analysis, thoughtful strategy, and pragmatic building.
In any respect, I would love to continue the conversation, will dive deeper into your other publications, including the models and agents by the indirect tax vendors (with which I am intimately familiar). Most of what I had seen is wishful thinking and buzzword speak. We are building indirect tax agentic infrastructure platform, need real use cases to implement, and invite real experts to critique, challenge and evaluate what is possible.
Interesting analysis. The emphasis on starting with high-volume workflows like reconciliation and cash application stands out. For finance teams, the real opportunity isn't simply adopting AI, but integrating it into governed processes where efficiency, accuracy, and measurable business outcomes can be demonstrated.
Appreciate that, David. "Governed" is doing a lot of work in that sentence. The teams actually shipping this treat the audit trail as part of the product spec, not a compliance patch bolted on after the pilot works. If an agent can't explain why it matched a payment, it should not go live.
Curious how you're thinking about "governed" on your end — role-based approval gates before an agent can act, or something closer to real-time monitoring of every decision it makes?
I see it as some of the former (role-based approval gates) coupled with the latter (across-the-board real-time monitoring). Like people, you don't need agents to come to you for approval for everything they do, but you do need to monitor their body of work.
That's a good way to put it. The analogy that's stuck with me from the reconciliation work is the junior-analyst one — day one, every entry gets reviewed; six months in, you're spot-checking. Most orgs never build that trust curve for agents, it's either approve-everything forever or nothing.
Real question for me is what "spot-checking" looks like for an agent — sampling a percentage of transactions, or something closer to anomaly detection flagging the weird ones on its own?
I'm starting to get out of my depth here, but a combination of both seems the best course of action. After all, you're already using AI, so why not have one routine looking for anomalies and another sampling a percentage of transactions? If you only check one way, you're more likely to miss Agentic weaknesses.
Hello Diyan. Thank you for publishing no BS TaxTech as there is a lot of BS out there.
Your effort/ROI matrix is the part worth sitting with longer than the connector counts. VAT reconciliation at 9.0 ROI, cash application at 9.3 — those numbers only clear if the underlying system was built to be called by an agent in the first place. Most of what's getting built right now is the opposite: a legacy tax engine with a chat UI stapled on, then months of integration work teaching an LLM to navigate a workflow designed for a human clicking through screens. That's where your $60K–$250K+ enterprise tier really comes from — the translation layer, not the tax logic.
That's the problem we started https://orchestrate.tax to solve. The orchestration layer is API-first and MCP-callable from day one — the same operations a controller triggers from a dashboard are what an agent calls directly, same audit trail, same governance, no translation layer in between. XMPT (exemption certificates) and TAXDAI (indirect tax intelligence) sit underneath as agent-callable services rather than siloed apps to reverse-engineer.
Practically, that collapses your build tiers rather than adding a fourth — a VAT reconciliation workflow becomes composing existing primitives instead of a $100K custom build.
One pushback: build-vs-buy still treats "buy" as one category. Buying a platform whose primitives were built for agents and buying a seat license for a platform now marketing itself as agentic aren't the same purchase, even if the sales deck reads the same. Worth a column in the next version of this map.
@Michael Roytman It’s be easier to understand and discuss if we don’t use AI-generated comments. I may be wrong, but yours really looks like Ai-generated. Just a thought
Ha, now I sound and write like AI. That is reassuring. I analyzed your materials pretty diligently as it was a well thought out post.
What topics would you like me to dive in deeper?
:) I was wrong. Impressive. Maybe we are nearing the time when machine generated text is starting to be indistinguishable from himan generated.
no worries. Or maybe we are "nearing the time" that no one and nothing can be trusted with very few humans actually perform analytical research, empirical analysis, thoughtful strategy, and pragmatic building.
In any respect, I would love to continue the conversation, will dive deeper into your other publications, including the models and agents by the indirect tax vendors (with which I am intimately familiar). Most of what I had seen is wishful thinking and buzzword speak. We are building indirect tax agentic infrastructure platform, need real use cases to implement, and invite real experts to critique, challenge and evaluate what is possible.
Interesting analysis. The emphasis on starting with high-volume workflows like reconciliation and cash application stands out. For finance teams, the real opportunity isn't simply adopting AI, but integrating it into governed processes where efficiency, accuracy, and measurable business outcomes can be demonstrated.
Appreciate that, David. "Governed" is doing a lot of work in that sentence. The teams actually shipping this treat the audit trail as part of the product spec, not a compliance patch bolted on after the pilot works. If an agent can't explain why it matched a payment, it should not go live.
Curious how you're thinking about "governed" on your end — role-based approval gates before an agent can act, or something closer to real-time monitoring of every decision it makes?
I see it as some of the former (role-based approval gates) coupled with the latter (across-the-board real-time monitoring). Like people, you don't need agents to come to you for approval for everything they do, but you do need to monitor their body of work.
That's a good way to put it. The analogy that's stuck with me from the reconciliation work is the junior-analyst one — day one, every entry gets reviewed; six months in, you're spot-checking. Most orgs never build that trust curve for agents, it's either approve-everything forever or nothing.
Real question for me is what "spot-checking" looks like for an agent — sampling a percentage of transactions, or something closer to anomaly detection flagging the weird ones on its own?
I'm starting to get out of my depth here, but a combination of both seems the best course of action. After all, you're already using AI, so why not have one routine looking for anomalies and another sampling a percentage of transactions? If you only check one way, you're more likely to miss Agentic weaknesses.