A partner at a mid-tier Irish practice told me recently that her team had started using AI tools to draft audit documentation. She was broadly comfortable with it. What made her pause was a question from one of her seniors: “Who’s reviewing the AI’s reasoning, not just its output?” She didn’t have a clean answer. Neither do most firms right now.
The high-profile AI hallucination incidents that emerged at EY and KPMG in 2026 were not, at their core, technology failures. They were review failures. The AI produced plausible-looking work. Experienced professionals signed off on it without scrutinising the underlying logic. The errors only surfaced later – in some cases, significantly later. That gap between output and scrutiny is where the risk lives, and it’s a gap that traditional review processes were simply never designed to close.
The review process most firms use was built for human error
Standard review frameworks in practice are good at catching what they were designed to catch: miscalculations, formatting inconsistencies, missed disclosures, logic gaps that a trained second pair of eyes would notice. They work on the assumption that the person who prepared the work understood what they were doing, even if imperfectly.
AI breaks that assumption. A language model producing a tax memo or an audit narrative doesn’t “understand” anything. It generates text that is statistically likely to be correct given its training data. When it’s wrong, it is often wrong in ways that look entirely right – fluent, confident, internally consistent. That’s precisely what makes it hard to catch using a review framework built around human cognitive patterns.
The firms caught out in 2026 weren’t using AI carelessly. They had policies. They had sign-off procedures. What they didn’t have was a review methodology that was specifically designed to interrogate AI-generated content – to go behind the output and ask whether the reasoning, the sources, and the conclusions actually hold up.
What “fit for purpose” actually means in 2026
Redesigning your review process for an AI-assisted environment isn’t about adding a checkbox that says “AI used: yes/no.” It’s about changing what reviewers are trained to look for and how they are given the time and mandate to look for it.
There are a few things that stand out from conversations with practice leaders who are taking this seriously. First, reviewers need to go back to source. If an AI tool has cited a piece of legislation, a case, or a figure, the reviewer’s job is not to check whether it sounds right – it’s to verify that it’s actually there. That sounds obvious. Under time pressure, with a deadline looming and a document that reads perfectly well, it rarely happens automatically.
Second, the pressure gradient matters. Junior staff are often the ones working most directly with AI tools, and they are also the least likely to push back when something doesn’t seem right. If a manager reviews AI-assisted work from a trainee and that work looks polished, there is a social and time pressure dynamic that makes deep scrutiny harder. Firms need to name that pressure explicitly and build review frameworks that account for it – not assume that good intentions are enough.
Third, and this is the one Irish practice leaders are least comfortable discussing openly: the billable hour model creates a structural disincentive to thorough review of AI output. If AI speeds up preparation but review time is held constant, firms recapture margin. If review time increases to match the new risk profile, that margin narrows. How firms resolve that tension is a strategic decision, not just a quality control one.
The staffing dimension that doesn’t get mentioned enough
Here’s something relevant to where I sit, working with practices on hiring: the AI review gap is partly a talent and capacity problem.
Effective scrutiny of AI-assisted work requires people who understand both the technical content and the limitations of the tools. That’s a particular kind of seniority. Managers and senior managers with three to six years of practice experience, strong technical grounding, and the judgment to know when something that looks right isn’t – those are the people Ireland’s practice market is shortest of right now. They’re also the cohort most actively being recruited by industry and multinationals.
Firms that are investing in AI tools without simultaneously investing in retaining and developing the people capable of reviewing AI output properly are not reducing risk. They’re redistributing it in ways that may not be visible until something goes wrong.
What you can do now
Before your next team meeting on AI policy, ask a more specific question: not “do we have a review process?” but “does our review process actually test the reasoning behind AI output, not just its appearance?” If the honest answer is uncertain, that’s where to start.
The firms that come through this period well won’t be the ones that avoided AI. They’ll be the ones that thought hardest about what human judgment still has to do – and protected the capacity to do it.
