First published on LinkedIn, 23 July 2026.
“We don’t know where to start.” A head of tax said it to me over coffee last month, and I’ve heard versions of it all year. The reasons that follow vary: no budget, no time, no spare people, no clear picture of what the technology can actually do, and no idea how to get the board to care. What almost never varies is who’s saying it: tax people running functions that defend complex positions under scrutiny. I want to take those blockers seriously in this article. Deciding what to do with AI isn’t about technology. It’s a judgement skill, and it’s one we already own as a profession.
1. The Borrowed Language of Scarcity
No budget, no time, no people. Think again and notice these are the answers any unscoped project collects, on any subject, at any time. A CFO funds ranked consequences: this deadline, this exposure, this saving, in this order, for these reasons.
The scarcity is also the raw material. When we say “no time”, we already know exactly which processes eat the hours. When we say “no people”, we already know where the weight sits. The complaint list is the candidate list. No discovery project required. And the first ordering doesn’t need a framework, because we mostly carry the facts in our heads: what has a date attached (an e-invoicing mandate, a filing deadline), what costs money when it goes wrong (a tricky audit), what eats the most hours (a manual process stitched across various systems). Put the handful of things we already say out loud in every team meeting down on paper, and we have a list. That list shapes the next budget conversation: the ask stops being “fund AI” and becomes “fund the first item on a list the CFO and Head of Tax can audit”.
2. The Uneven Reach of the Same Model
Not knowing what AI can do is the most rational blocker on the list, because the boundary of AI capability is very tricky to define. Harvard Business School research, since published in Organization Science, described that unevenness through a field experiment with 758 BCG consultants across 18 real tasks, one of them deliberately designed to sit just beyond what the technology could handle. On the ordinary tasks, the consultants with AI produced work rated over 40 per cent higher in quality than the control group. On the engineered task, the consultants with AI were 19 per cent less likely to get the right answer than colleagues with no AI at all. As the authors put it, AI assistance “improves performance for some tasks but worsens it for others”, within the same workflow and at apparently similar difficulty.
The counter-move is to substitute the question. Stop asking what the technology can do in general, and ask instead which of our own specific tasks our own tax people could verify quickly if a machine attempted them. The original question has no stable answer. This substitute question is a piece of professional judgement a tax team could carry out, confidently.
3. The Native Currency of Buy-in
“We can’t get senior buy-in” almost always means the ask arrived in the wrong currency. Boards buy consequences, not technology: a mandate start date that won’t move, an exposure growing quarter on quarter, a position that can’t currently be defended with the data the business holds.
We in tax spend our entire careers converting technical complexity into consequence, deadline and defensibility for non-specialist audiences. The translation is the skill, and we already have it. What kills buy-in is skipping the translation and taking a tool and its expected benefit to the board directly.
So present the ranked consequence and the order in which we intend to retire it, never the technology itself. Buy-in follows the same logic budget does: it gives a board something it can say yes to without taking our subject on faith.
4. The Familiar Discipline of Judgement
Look at the judgement running through all three moves:
- rank by risk and value, with evidence beside every choice;
- favour what can be checked over what merely impresses;
- hold only positions we can defend under challenge, and review before we rely.
None of this was invented for AI. It’s the working discipline of a tax function, applied every day to positions carrying real financial and reputational risk. It is not for tax alone either. Engineering teams building with AI run the same discipline under their own vocabulary, evals and guardrails, where we’d say controls: owner, evidence, review, remediation. The PCRT topical guidance on AI (Jan 2026) also points members the same way: professional judgement and oversight over anything we rely on, whoever built it.
In the past few months I’ve watched several in-house tax teams present AI work they built themselves. What separated them from the functions still at the “where do we start” stage wasn’t engineering talent but the decision to treat “what should we point this at” as a professional judgement: scoped narrowly, evidenced, and reviewed like a tax position they would sign. The building part, given where the tools have got to, turned out easier than expected.
The question we are already qualified to answer
“What deserves to be first, and how would we defend the choice?” That question replaces “what can AI do and where do we start”. We can exercise that judgement today, and the entire apparatus for it, evidence, ranking, defensibility and review, is the one our profession has been practising all along.
So here’s the version I’m taking into summer: which single task in your tax function deserves to go first, and what evidence makes the choice defensible? I’d love to hear how others are answering it. Find me on LinkedIn if you’d like to compare notes.