The Case for Expanding Audit Margins Through Shifting Quality Left
Every conversation about margin in audit ends up being a conversation about AI. Draft the memo faster, tie out the schedule faster, get the contract summarized in seconds. The pitch lands because execution is where the hours sit, and hours are what get billed against budget.
But if you ask someone who runs review what actually blows up a budget, the answer is rarely that fieldwork ran long. It's that the partner kicked the file back three weeks later and half of it had to be redone. The expensive problem in an audit is the lag between finishing a piece of work and catching what's wrong with it, and most firms skip straight past it on the way to buying a copilot.
Where the hours actually go
Strip out the marketing and the hours on a typical engagement fall into a familiar shape: roughly half go to execution, about a quarter to review, and the rest to planning, wrap-up, and keeping the file organized. The ratio shifts with the size and risk of the engagement, but the pattern holds.
The number that decides whether you come in on budget is that review quarter, and it doesn't spread evenly across the calendar. It gets compressed into whatever window is left before the deadline, running on work that was finished weeks earlier. The distance between when a piece of work gets done and when someone with authority looks hard at it is where margin gets lost.
The cost of a three-week-old question
A senior finishes a section. It sits. Two and a half weeks later a manager picks it up, finds a hole in the documentation, and sends it back. By then the preparer has rolled onto another engagement, the client contact who could explain the number has gone quiet, and no one can reconstruct from memory why the judgment call was made the way it was.
That has nothing to do with competence. It's what happens to context when the feedback loop runs in weeks instead of hours. The work was fine, or close to fine, when it was fresh. By the time someone qualified to catch the problem gets to it, the person who could have fixed it in a sentence has lost the thread.
Most of the time that costs a few hours of re-explaining and re-tracing. Sometimes it costs far more. If the conclusion was actually wrong rather than thinly supported, no clarifying email saves it. The evidence gets pulled again. The client gets re-opened on a question they thought was closed a month ago. The section gets rebuilt from scratch. That is the engagement starting over, on the clock, after the budget has already been spent once.
The real cost driver is distance
The natural response is to make execution faster and buy more runway before the deadline. It helps a little. It does almost nothing about the real cost driver, which is the distance between doing the work and catching what's wrong with it.
The cheapest version of any defect is the one caught the day it's made.
A documentation gap flagged the same afternoon costs a sentence. The same gap found three weeks later, after the file has moved on and everyone's context has scattered, costs a re-investigation. The defect is identical. What changed is the price of fixing it, because every week of distance is another week of context the team has to pay to rebuild.
None of this is new. Every discipline that runs on review cycles already knows the earlier a defect surfaces, the cheaper it is to resolve. Audit has just never had a good way to act on it, because review has always meant a human reading a finished file days or weeks after the work was done.
Why this comes before the AI conversation
Most of the current AI pitch to firms gets the sequence backwards. The promise is faster execution: more drafts, more first passes, more output, produced quicker than a person could manage. None of that touches the review gap. It arguably widens it, because now more output funnels into the same review capacity, on the same multi-week cadence, shedding the same context before anyone senior sees it.
There's also a new kind of review stacked on top of the old one. Clean, confident prose is easy for an agent to produce. Whether it followed the firm's methodology is a separate question, and answering it is real work. Firms are learning that an agent slick enough to write a polished first draft is not necessarily one that understood how the firm reaches its conclusions, and the babysitting that closes that gap can eat most of the time the drafting saved.
The economics only hold up if you are honest about where inference actually belongs. Much of an audit is deterministic work: recalculate a total, tie a number back to its source, confirm the footing. Those steps don't need a language model, and pushing them through one is slower and more expensive than the formula that has always handled them, with no gain in reliability. Judgment is the narrower part, and that is where inference is worth paying for. Firms that never draw that line end up paying model prices for arithmetic and calling it transformation, and because most AI tooling still bills by the token or the seat, the vendor has little reason to help them draw it. We took the opposite view a while ago and priced on outcomes, so the only incentive is to spend inference where it actually changes the answer.
Faster execution poured into a slow, distant review cycle doesn't compress the timeline. It just moves the same bottleneck earlier.
The lever that matters sits upstream: catch the problem where the work happens, while the preparer still remembers why they did what they did, rather than weeks later when it becomes someone else's puzzle to reconstruct. A documentation hole, an unsupported conclusion, a skipped methodology step, surfaced at the point of work, gets closed with a sentence instead of a rework cycle.
This is the part of the audit that's genuinely broken, with or without AI in the picture. Close the distance between doing and reviewing first, and everything else, including the places AI really does earn its keep, has something solid to build on instead of a faster way to produce more of the same problem.
Margin comes from not doing the work twice
The firms that pull ahead on margin over the next few years will be the ones that shrink the distance between when a quality issue is created and when it's caught, so the three-week-old question never has to be asked at all. Generating more drafts won't get them there. It's a sequencing decision more than a tooling one, and it's worth settling before the rest.