93% of Finance Professionals Don't Trust Their Own AI Output — The Verification Habit That Closes the Gap
93% of Finance Professionals Don't Trust Their Own AI Output — The Verification Habit That Closes the Gap
Distrust isn't the problem. Distrust without a verification step is.
ACCA and Chartered Accountants Australia and New Zealand surveyed 1,600 finance professionals for their July 2026 report, Enabling Finance Insight. The headline finding travelled widely: 93 per cent expressed concern about the integrity and verifiability of AI-generated insights, with 67 per cent reporting a high level of concern.
It got read as a story about AI not being ready. I read it differently. Ninety-three per cent of a profession being sceptical about a new tool is not a failure — it's the profession working as designed. Accountants are trained to distrust unverified numbers. That instinct is the asset.
The problem is what sits underneath it. In the same survey, 72 per cent of respondents reported only basic generative-AI proficiency or none at all — 34 per cent with zero skills, 38 per cent with basic skills only. So the picture is a profession that doesn't trust AI output and largely doesn't have the skills to check it either. That combination doesn't produce caution. It produces one of two things: people who avoid the tools entirely, or people who use them and accept whatever comes out because they have no practical way to test it.
|
93%
Concerned about the integrity and verifiability of AI-generated insights — including 67% with a high level of concern.
|
72%
Have basic generative-AI skills or none at all (34% none, 38% basic) — the gap that turns healthy scepticism into paralysis or blind acceptance.
|
Concern is not a control
Every finance function I've worked with has a review step for human-prepared work. A journal gets reviewed. A reconciliation gets signed. A management reporting pack goes through someone before it goes to a board. Nobody argues about whether that's necessary.
Very few have written down the equivalent step for AI-assisted work. The output gets eyeballed, it looks plausible, and it goes in. "I don't fully trust it" is doing all the work that a documented control should be doing — and unwritten scepticism degrades fast under deadline pressure. On the third day of a month-end close, at 7pm, plausible-looking output gets waved through.
The fix is not more sophisticated AI. It's a verification step that takes ninety seconds and happens every time.
The before half: control what goes in
Most of what gets called an AI hallucination in finance work is actually an input problem. The model was asked a question about data it was never given, and it filled the gap.
Two habits fix the bulk of it.
Give it the source, don't ask it to recall. "What's the current super guarantee rate?" invites a recalled answer that may be a year out of date. "Using the attached ATO page, tell me the rate applying from 1 July 2026" is a reading comprehension task with a checkable answer. The second question is almost always the one you actually want.
Ask for the working, not just the answer. If you request the calculation steps and the specific figures used alongside the result, verification becomes a matter of checking three numbers rather than re-performing the analysis. This costs nothing at the prompt stage and saves the entire review.
The after half: three questions, every time
This is the habit itself. Three questions, applied to any AI-assisted output before it goes anywhere:
1. Does it tie to something? Pick one number in the output and trace it back to the source data. Not all of them — one. If that one ties, the extraction worked. If it doesn't, stop and don't debug the rest until you know why.
2. What did it leave out? This is the failure mode people miss. AI output is rarely wrong in an obvious way; it's incomplete in a quiet way. A variance analysis that covers eleven cost centres when you have twelve reads perfectly well. Check the count, the date range, and the total — completeness before correctness.
3. Would I sign this? The genuinely useful test. If your name goes on it and someone asks you to explain a line in it next week, can you? If the honest answer is no, the output isn't finished regardless of how good it looks.
Ninety seconds. Every time. That's the whole control.
Not everything needs the same level of checking
Applying maximum scrutiny to everything is how verification habits die — it's too slow, so it gets dropped entirely. Tier it instead.
Full verification — anything that goes external or drives a decision. Board and management reporting pack commentary, funder acquittals, anything touching payroll or a statutory lodgement, any number a director will rely on. Every figure traced.
Spot check — internal working documents, first-draft analysis, summarising a long document you'll read anyway. The three questions, one traced number, move on.
Light touch — reformatting, drafting an email, restructuring text you wrote. There's no factual claim being generated, so there's nothing to verify beyond reading it.
The mistake I see most often is treating a management reporting pack commentary as tier three because it "reads fine." It's tier one. It's the thing the board actually reads.
The skills gap is the part worth acting on
Of the two findings, the 72 per cent is the one that should bother finance leaders more than the 93 per cent. Scepticism is free. The ability to verify is not — it's a learned skill, and by the survey's own numbers most of the profession hasn't learned it yet.
The encouraging part of the same report is that more than 60 per cent of finance teams have increased their use of real-time operational data over the past two years. The raw material for better insight is arriving. What's missing is the confidence to act on what comes out the other end.
That's a training problem with an unusually short payback. Nobody needs to become a data scientist. They need to know how to structure a request so the answer is checkable, and they need three questions they ask every single time. An hour of practice covers both.
Is your team using AI without a documented check on the output?
Unwritten scepticism isn't a control, and it won't survive a month-end. PFL provides senior-level outsourced finance, management reporting, and AI automation for Australian NFP, NDIS, and SME organisations — including building the review step in properly.
Talk to PFL →
Comments
Post a Comment