Where AI actually saves time in an accounting firm (and where it doesn't)

Ask a partner where AI should start and you will almost always hear about the thing that irritated them most recently — chasing a client for a missing invoice, re-keying a trust distribution, rewriting the same engagement letter for the ninth time this month. Irritation is a useful signal that something is broken. It is a poor signal for what to automate first.

The tasks that pay back fastest have a specific shape. They are high in volume, low in judgement, and they already have a right answer that somebody can check in seconds. The tasks that burn budget look superficially similar but fail on one of those three dimensions — usually the last one.

The four questions worth asking first

Before scoping any automation, we run the candidate workflow through four questions. A workflow needs a confident yes to all four before it earns a place in the first phase of work.

1. How many times a year does this happen?

Frequency beats duration. A twenty-minute task done four hundred times a year is a far better candidate than a two-day task done twice. Frequency also means you accumulate examples quickly, which is how an automation gets better at the edge cases rather than getting quietly abandoned the first time it stumbles.

The number most firms cannot answer here is the honest one. Practice management data usually records the engagement, not the task inside it. Before you build anything, it is worth two weeks of someone simply tallying what actually consumes the team's time.

2. Is the input structured enough to be recognised?

Modern models handle messy inputs far better than the previous generation of OCR tools — a photographed receipt, a scanned bank statement, a PDF that is really an image. But there is still a difference between messy and ambiguous. A supplier invoice is messy: the fields are always there, they just move around. A client emailing "same as last year but not the car thing" is ambiguous, and no amount of model quality fixes that.

3. Can a human verify the output faster than they could produce it?

This is the question that separates real savings from theatre, and it is the one most often skipped.

If checking the machine's work takes as long as doing the work, you have not automated anything. You have added a step.

Extracting figures from twelve source documents into a workpaper passes easily — a reviewer can eyeball a populated schedule against the source in a fraction of the time it took to type it. Drafting a nuanced piece of tax advice fails: the reviewer has to reason through the whole position anyway to know whether the draft is right, so the draft saves typing and nothing else.

4. What happens when it is wrong?

Every automation will be wrong sometimes. The question is whether being wrong is recoverable. A mis-categorised transaction surfaces at reconciliation and costs a minute. An incorrect figure that flows unchecked into a lodged return costs considerably more than the automation ever saved.

This is not an argument against automating high-stakes work. It is an argument for putting the review gate in the right place, which we have written about separately in designing AI workflows your reviewers will trust.

---

Where the payback usually is

Across the engagements we have scoped, the same handful of workflows come out on top. None of them are glamorous. That is rather the point.

Where it usually isn't

Three categories come up repeatedly in scoping conversations and rarely survive the four questions.

• Technical advisory and complex positions. The reasoning is the work. A draft does not reduce the reviewer's burden, because they must reconstruct the reasoning to trust it.
• Anything that depends on a conversation you have not had yet. If the blocker is that the client has not told you something, automation moves the bottleneck rather than removing it. Automating the chase is valuable; automating the answer is not possible.
• One-off cleanups. Migrating a decade of records or untangling a single messy file feels like an ideal AI job. It is usually cheaper to do it once by hand than to build, test and validate something you will use once.

The compounding effect nobody scopes for

There is a second-order benefit that rarely appears in the business case and often turns out to be the larger one. When document extraction is automated, the extracted data becomes structured and queryable for the first time. Work that was previously impossible — not slow, impossible — becomes routine: comparing this year's deductions against last across an entire client base, flagging clients whose circumstances have shifted, spotting the advisory conversations worth having.

Firms typically buy an automation to save hours and end up valuing it for the visibility it creates. That is worth keeping in mind when you sequence the work: the workflows that produce reusable structured data are worth more than their hour-savings alone suggest.

Start narrower than feels worthwhile

The most common failure mode we see is not a technical one. It is scoping phase one as "automate tax return preparation" instead of "automate the workpaper population step for individual returns for salary-and-investment clients". The narrow version ships in weeks, produces evidence you can argue with, and gives the team a concrete thing to react to. The broad version produces a six-month project and a lot of opinions.

Pick the one workflow that answers yes four times. Get it live. Then let what you learn — not the original plan — decide what comes second.