KPMG just published the most consequential AI study an accounting firm has ever produced — and read carefully, it's about AI professional skepticism, not AI skill. Three accounting professors at UT Austin's McCombs School of Business spent eight months watching 2,500 KPMG employees and 1.4 million real workplace AI interactions. They identified four behaviors that separate sophisticated AI users from routine ones — frequency, persistence, ambition, and intentionality.
Monday's roundup named the bottleneck. Anthropic and OpenAI announced $11.5B in same-day private-equity-backed enterprise joint ventures and admitted on stage that selling models isn't enough — the constraint is organisational absorption. KPMG's behavioral data shows what that constraint looks like inside an accounting firm.
The headline finding: 5% of KPMG users consistently demonstrated all four. Eight months. Big Four budget. Executive air cover. Five percent.
That's the number doing the work. Not a benchmark to beat — evidence of how hard the prevailing approach actually is.
The four behaviors are AI professional skepticism by another name
Read them back without the AI vocabulary.
Persistence — pushing back on early answers, requiring sources, demanding explanation. That's how you review a junior's workpaper.
Ambition — leading with audience, objective, success criteria, structured constraints. That's how you scope an engagement.
Intentionality — picking the right tool for the task. That's how you choose between a journal entry and a reclass, or between a tax memo and a footnote.
Frequency — using the tool as part of professional rhythm. That's just how a professional muscle works.
KPMG didn't discover a new competency. They re-derived a version of professional skepticism from prompt logs. They just didn't recognise it — the authors wrote for HBR's general business audience, not for accountants.
I wrote about this in Neither of Us Alone — a cross-border tax research session where adversarial dialogue with the AI caught a material legal error neither of us would have found alone. That's not a new skill. It's a partner reviewing a junior's work, applied to AI output. KPMG calls it persistence. Most accountants already have a name for it.
Microsoft's own 2026 Work Trend Index confirms the gap. 86% of AI users say they treat AI output as a starting point and stay responsible for the thinking. KPMG's behavioral data says only 5% actually do. The gap between what people claim and what they deploy is the transformation problem.
Why accountants are pre-equipped to be Frontier Professionals
This makes accountants the workforce most pre-equipped to be sophisticated AI users in any profession. The four behaviors aren't exotic — they're the muscle CAS practitioners use every day on staff workpapers, tax positions, audit opinions, and advisory deliverables.
The reason most CAS staff aren't deploying it is cultural, not technical. No one has named AI output as work product subject to professional review. A draft from a junior gets reviewed. A draft from AI gets accepted. The muscle exists — it just isn't being pointed at the AI.
The 5% problem reads differently for a 10-person firm
Microsoft's Work Trend Index this week asked a different question: not which individuals are ready, but which organisations. Organisational factors — culture, manager support, talent practices — correlate with roughly twice the AI impact of individual mindset and behavior (67% versus 32% in their model).
Microsoft's own line on page 17: the real question isn't whether people have the right skills — it's whether the organisation is built to unlock them. That's the entire CAS argument in their words.
Here's where the data shifts the read for small firms. Microsoft's Frontier Professionals — the 16% of AI users who consistently practice ambitious AI work — are 11% in finance and accounting roles. Accountants are over-represented at the frontier. The muscle exists, and the data agrees it does.
Frontier Professionals also concentrate in larger firms (53% in 500-plus-employee companies), so small CAS practices probably have a lower rate than the headline 16%. Probably is not none. In a 20-person firm, you almost certainly have one Frontier Professional already on your payroll. In a 30-person firm, two or three. Below 10, you have a real chance — and finding and partnering with that person is the highest-leverage move you can make this quarter.
The structural advantages run in your favor too. Microsoft's data names manager modeling as the single highest-leverage practice. Frontier Professionals' managers openly use AI (85% versus 64% for non-Frontier), set quality standards, create experimentation space, and reward ambitious redesign even without immediate results — 20-plus point gaps across all four behaviors. In a Big Four firm, a partner's AI behavior is visible to maybe twenty people. In a 10-person firm, every staff member sees it daily. The intervention that matters most in Microsoft's data is mechanically more powerful at small-firm scale.
KPMG had to fight a 50,000-person organisational system. You don't.
Stop teaching prompt engineering. Start training the muscle.
A year ago, AI training meant prompt patterns and lunch-and-learns on how to phrase a request. That curriculum now teaches the wrong skill. Frontier models forgive sloppy prompts. The differentiator has moved from how you ask to whether you bring the muscle.
Real training now does four things the old approach didn't. It works against firm-specific artifacts, not generic prompts — staff start with the firm's instruction files, skills, and templates. It happens inside real deliverables — a junior learns persistence drafting an actual variance commentary or tax position memo with a coach reviewing, not reading a slide on iteration. It pairs Champions with Operators — champions experiment, operators codify the wins into firm context. The 5% doesn't have to grow to 90% individually; they feed the artifact layer everyone else operates inside.
And it sets standards rather than encouraging behaviors — the part most firms get most wrong.
Require professional skepticism on AI work — don't encourage it
You don't encourage professional skepticism on workpapers. You require it as a condition of practice.
The AI-generated draft is a workpaper. Treat it that way.
Every variance commentary touching AI gets reviewed against the underlying numbers. Every tax memo gets sourced. Every client communication that started with a prompt gets the same scrutiny as one drafted by hand. No special treatment for AI-generated work product.
Microsoft's data shows what happens when firms encourage instead. Only 13% of AI users say they're rewarded for reinventing work without immediate results; only 26% say their leadership is consistently aligned on AI. The rest are nodding politely while returning to old workflows.
Encourage produces 5%. Require produces a different number.
Your champions are already on the payroll
KPMG just published the most consequential AI study an accounting firm has ever produced. The trap is reading it as a story about a new skill the profession needs to acquire. It isn't. It's a story about a muscle you already have, deployed at small-firm scale, with structural advantages running in your favor — if you're willing to lead.
A diagnostic question for the partner reading: can you name the Frontier Professional already on your payroll? If you can, are you giving them the conditions to lead — visible permission, manager modeling, a workpaper standard for AI output, and a reward structure that rewards the work? If you can't, finding them is this quarter's project.
That's what the AI Practice Transformation program is built for. Three weeks, one day per week — we map your firm's highest-judgment work, rebuild it as adversarial AI processes, and give your team a repeatable system. If you're ready to stop encouraging and start requiring, visit theaiaccountant.ai/transformation.