Linkedin
Facebook
Email
more
Six months of working with AI in tax has made one thing increasingly clear: the model itself is not the durable asset. The real value lies in the controlled environment around it, where data, tax policies, calculations, permissions and professional review come together.
The six months behind us have been the fastest in the short history of applied AI. Copilots became agents. Frontier models leapfrogged each other on capability and on price, several times over. The word harness moved from engineering forums into boardroom vocabulary. We have been running this experiment in public since March, when we published our first account of building, simulating and testing AI in tax work. The six months since have been the most instructive part of it, not because the technology surprised us, but because of what happened when that technology met real tax departments. Some of what we expected held. Some of it did not survive contact with reality, which is the point of taking positions in public. This article shares what we saw in the field, and the one conclusion we would now place above all others.
This article was written by Mourad Seghir (mseghir@rsmnl.nl) and Mario van den Broek (mvdbroek@rsmnl.nl). Mourad and Mario are consultants with RSM Netherlands with a focus on Tax & AI.
AI & Tax: we have got a story to tell (March 2026). Our white paper documented two years of building, simulating and testing AI in tax work: the five layer working hypothesis, workflow simulations across corporate income tax, transfer pricing, VAT and global employer services, and the assumptions that did not hold. It remains the practical playbook underneath this article and is available via rsm.global/netherlands.
What has six months of AI in tax taught us?
The agent arrived, and it was not the hard part. In March, most tax teams were working with copilots: tools that help one person with one task. Since then, AI systems have learned to receive an objective, plan the necessary steps, operate software tools, retrieve documents and carry a multi step assignment through to completion. Every serious provider now ships agentic capability, and the demonstrations tax teams are shown have become dramatically better almost overnight: an agent that assembles a reconciliation, drafts a Local File update or prepares a first response to an information request is no longer remarkable.
We expected the technical ceiling to hold for longer than it did. Yet production barely moved. An agent that can update a transfer pricing Local File still needs to know which prior year document is final, which policy has been approved, where the current financial data lives, which calculation it must use and when a change requires human approval. None of that knowledge sits in the model. It sits, or fails to sit, in the organisation. Capability, it turned out, was never the missing ingredient.
Why the AI model itself is becoming less important
The model stopped being a decision. Over the same six months, performance, pricing and deployment options across the leading providers changed repeatedly, sometimes within weeks. The cost of a given level of capability fell in steps, the capability ranking changed hands more than once, and the deployment menu widened from a single cloud subscription to a spectrum that includes European hosting and locally deployed models for data sensitive functions. Organisations that designed their tax workflow around a single model or a single chatbot licence have already redesigned it twice since spring.
We assumed model choice would matter more than it does; the market corrected us within a quarter. What survived every model release unchanged was something else entirely: the tax function's own data definitions, its encoded policies, its calculation logic, its permissions and its review gates. The models kept changing underneath, and the work carried on.
AI in tax is becoming a governance and compliance challenge
The regulatory clock kept running. While capability raced ahead, the demand side hardened in parallel. Pillar Two has moved from a data collection exercise into live filing obligations that must be rerun whenever data updates. The EU has fixed the trajectory towards mandatory digital reporting and e invoicing for cross border transactions under VAT in the Digital Age, and several member states are not waiting for 2030.
The EU AI Act's transparency and documentation obligations are phasing into force, which means that AI used in material financial processes must itself be governed, logged and explainable. Tax authorities are adopting the same technology for risk selection, so scrutiny is becoming continuous and data tested rather than periodic. The uncomfortable symmetry of the period is this: the tools available to tax departments and the expectations placed upon them rose together, and both now assume an operating model that most departments do not yet have.
Why AI pilots in tax succeed or fail
The pilots that died, died for organisational reasons. Our March simulations estimated that a controlled model could remove roughly 55 to 75 percent of manual reconciliation effort across corporate income tax, transfer pricing and VAT cycles. Six months of field work has not disproved those numbers; it has qualified them. The reductions materialise where the organisational decisions were taken first, and they do not materialise at all where the pilot was expected to substitute for those decisions. Drafting, the task everyone worried about, turned out to be the easy part. Deciding turned out to be the hard part.
One development cuts both ways. Businesses can now build working applications with far fewer technical barriers, and some of the most useful tools we saw this period were created inside tax teams rather than bought. That is genuine progress, and it confirms that the profession can shape its own tooling. But it has also produced a new generation of ungoverned mini tools: calculators, prompt libraries and data extracts living on personal drives, outside version control and outside review. Shadow IT has a successor, and it is shadow AI. The energy is welcome; the absence of a controlled home for it is the risk.
Why the harness matters more than the model
For tax leadership, these observations converge on one strategic conclusion. The durable asset in an AI enabled tax function is not the model. It is the controlled environment around the model: the agreed data dictionary, the entity and account mappings, the encoded tax policies, the deterministic calculation engines, the permissions, the logging, the evaluation criteria and the review gates that together determine where, how and under which conditions AI may perform tax work.
That environment is the harness. The agent performs the work; the harness decides what the work is allowed to touch, which version of the truth it must use and what evidence it must leave behind. Models are rented, replaceable and improving on someone else's schedule. The harness is owned, and it compounds: every definition agreed, every policy encoded and every review gate designed makes the next process cheaper to automate and easier to defend. Six months of model churn did not devalue a single well built harness. It devalued every workflow built directly on a licence.
Where AI governance and tax control come together
The harness is also where two compliance worlds meet. Tax authorities increasingly expect a reported number to be rebuilt from source to filing, with the data, the logic and the approvals visible on demand. The EU AI Act expects AI used in material processes to be documented, logged and reviewable. These are, in substance, the same requirement approached from two directions, and a properly built harness satisfies both with one investment: the run log that evidences your tax control framework is the same artefact that evidences your AI governance.
Departments that treat AI governance, data management and tax control as separate compliance projects will pay for the same capability three times. Departments that build the harness once will find that defensibility stops being the bottleneck and becomes the by product.
Why AI adoption is now an organisational challenge
The second implication follows directly. The adoption gap is now an organisational fact, not a technology fact, and it will not close through procurement. Budget buys tools; only decisions build operating models. The decisions are unglamorous and specific: who owns this output, which dataset is approved, what does correct mean for this process, where must automation stop. A tax function that cannot answer those four questions will get the same result from the next tool as from the last one, only faster.
And the gap compounds against those who wait, because every quarter of encoded knowledge, agreed definitions and designed review gates makes the leaders cheaper to run and harder to catch, while the talent entering the profession increasingly expects to work this way. We will make the forward claim sharper: within a year, which model do you use will no longer be a question serious tax directors ask, in the same way nobody asks which database their ERP runs on. The questions that will matter are different. Do the numbers come from controlled, deterministic logic? Can the evidence be traced from source to filing? Can the professional who signs reconstruct the answer? Those questions are answered by the harness, not by the model, which is precisely why the harness is where attention and investment belong.
What does AI mean for the role of tax professionals?
The largest impact, finally, is on people rather than systems. As collecting, checking and drafting are progressively delegated, professional attention shifts to interpretation, challenge and ownership. The tax professional of the next few years must be prepared to take responsibility for conclusions they did not personally calculate or draft line by line.
That requires a broader AI literacy than prompting: knowing what data the system used, which steps were deterministic and which were probabilistic, which actions the agent performed and where the designed stopping points sit. Review becomes a skill in its own right, and it must be trained, evidenced and valued as such. Fewer handovers, stronger ownership. That is the trade the profession is being offered, and in our experience the professionals who embrace it do not feel replaced; they feel promoted.
How should tax functions start using AI?
The practical route does not start with a transformation programme, and it does not start with a model selection either. Pick one recurring tax process that causes stress every cycle: an effective tax rate bridge, a VAT reconciliation, a transfer pricing margin review. Redesign it once, end to end, from source data to professional sign off. Agree who owns the data, what the approved policy version is and where automation must stop.
Separate the calculations that must be deterministic from the tasks where AI reasoning genuinely adds value. Then connect the workflow through a harness with appropriate permissions, logging and review gates. The test of success is not whether the system produces an answer; it is whether another reviewer can reconstruct that answer and evidence the approval. Build that once, and you own something no model release can take away. We would be glad to help you decide where to start.
RSM is a thought leader in the field of AI and Tax. We provide frequent insights through training and the sharing of thought leadership, based on our detailed knowledge of industry developments and practical applications gained from working with our clients. RSM helps tax functions connect individual AI developments to the broader changes affecting their operating model. By doing so, we support businesses in distinguishing short term hype from structural change and translating that assessment into coordinated decisions on data, calculations, governance and professional review. If you would like to discuss what building the operating model for AI means for your tax function, please contact one of our consultants.
This is a publication that is part of RSM's AI & Tax thought leadership. On a frequent basis, RSM shares insights on artificial intelligence and taxation. Our consultants constantly follow global AI developments in an ever changing society and translate their impact into practical considerations for internationally active companies.
