Where Pilots Die in Audit Firms
The Graveyard
Audit firms have now run enough AI pilots for the failure patterns to be recognisable, and most deaths happen before the technology gets a chance to disappoint anyone. The commonest grave is client data access nobody cleared: a pilot planned around real engagement files discovers, months in, that engagement terms, confidentiality duties and sometimes the client's own regulators stand between the data and the tool, and nobody had asked. Next to it lies methodology approval that arrives after the busy season: a tool cleared for use in April is a tool nobody will risk learning in the one period the firm actually works, so by the next busy season the licence has lapsed and the champion has moved on. Both deaths share a cause. The pilot was planned as a technology project, and in an audit firm the binding constraints were never technical in the first place.
- Most pilots die before the technology gets a chance to disappoint anyone
- Client data access nobody cleared: engagement terms and confidentiality duties discovered months in
- Methodology approval landing after busy season means a tool nobody will risk learning until next year
- The shared cause: pilots planned as technology projects when the binding constraints were never technical
The Quieter Deaths
Two quieter failures kill the pilots that survive the first year. The first is inspection fear. A partner weighing a new tool is weighing an asymmetry: the upside is saved hours, the downside is standing in front of an inspector defending a novel way of working with no precedent to point at. Rational partners retreat to the old way, and a firm that has not decided — at firm level — how it will explain tool reliance to its regulator has silently decided that its partners will not rely. The second is transfer failure: a tool tuned on one engagement's documents meets another engagement's chaos — different systems, different formats, a client whose contracts live in scanned images — and the accuracy that justified the rollout quietly does not come with it. The pilot's results were real. They were results about the pilot.
- The partner's asymmetry: saved hours against defending a novel approach to an inspector without precedent
- A firm that has not decided how to explain tool reliance has decided its partners will not rely
- A tool tuned on one engagement's documents rarely survives another engagement's formats and chaos
- Pilot results are results about the pilot — transfer to the next engagement is a separate claim
What the Survivors Had
The pilots that became practice share three unglamorous features. They had a named owner — one person answerable for the tool's behaviour, its approval status and its results, which sounds like bureaucracy until you watch an unowned pilot dissolve the first time it produces a strange output and everyone assumes someone else is looking into it. They had a bounded first use: one procedure, one document type, one engagement whose partner volunteered, chosen so that failure would be cheap, contained and informative rather than a firm-wide embarrassment. And they had methodology and quality sign-off before rollout, not after — the people who answer for the firm's system of quality management involved while the answer could still be no. None of this is technical, which is the point: the survivors treated adoption as an extension of how the firm already governs its work, not as an exception to it.
- A named owner: one person answerable for the tool's behaviour, approval status and results
- A bounded first use: one procedure, one document type, one volunteer engagement, cheap to fail
- Methodology and quality sign-off before rollout, while the answer could still be no
- Nothing on the list is technical — survivors extended the firm's existing governance rather than working around it
Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.