AI for accountants: what it does, and what you still check

AI for accountants is at its most useful on the dull half of the job. Getting figures off a PDF, tidying a bank feed, drafting the email that chases the missing records, reading a forty page lease to find the three dates that matter.

What it will not do is take responsibility for any of it. That part has not moved, and I do not think it is going to.

The line that went straight through

This one is not from a practice, but it is the same failure, and it is why I am careful.

A landscape construction firm was comparing two supplier quotes. One line had matched Clematis Armandii Apple Blossom against Camellia Japonica, because the cultivar names looked a bit alike.

Completely the wrong plant.

The mistake was in the source data before AI went anywhere near it. A person had typed it. But it went through the comparison, into the summary and out the other end, because nothing in that process was looking for it.

I built that step, and I had not put a check in it. I was too busy being pleased that thirty eight plant lines had come off a PDF in about four seconds.

So that is the thing worth knowing before any of this goes near a set of books. AI does not create most of the errors. It moves the ones that were already there, faster, and it lays them out neatly enough that they look like somebody has thought about them.

Where the profession actually is

A UK government adoption plan published in June 2026 put professional and business services at 43.4 per cent using AI in December 2025, against 31.4 per cent a year earlier. Ahead of most sectors, and moving.

The more useful number comes from the ONS, which found the average UK business using AI runs 1.6 AI technologies, and only one in ten describes its use as extensive.

So this is not a profession that has ignored it. It is a profession where a lot of people have a licence, use it for the odd email, and have not changed a single thing about how a job moves through the office.

That gap is where the value is sitting. It is a skills question rather than a software one, which is inconvenient, because software can be bought on a Tuesday and skills cannot.

What it is genuinely good at in a practice

The Federal Reserve Bank of St Louis asked nearly 14,000 workers what they actually use generative AI for. Reading documents to gather technical information came top at 61.3 per cent, preparing research reports at 60.7 per cent, and analysing data for trends at 57.5 per cent.

Read that back as a description of a Tuesday in a small practice.

The jobs I see land well are the ones where something exists already and needs moving or reading.

Supplier invoices and statements arriving as PDFs, turned into a table you can work with. Bank descriptions that mean nothing on their own, grouped and given a sensible first guess. The email chasing a client for the records you asked for in June, drafted in the practice's tone rather than yours at half past six.

Then the reading jobs. A lease, a loan agreement, a new guidance note. Not to tell you what it means. To tell you where in the forty pages the relevant bits are, so you read four pages instead of forty.

And the writing that goes out to clients. Turning your technically correct answer into something they can follow without ringing you to ask what it meant.

None of that is clever. All of it is time.

The part that does not move

Every one of those jobs ends with somebody qualified looking at it.

Microsoft's 2026 Work Trend Index, which surveyed 20,000 knowledge workers and which is research by a company selling the product, found 86 per cent of AI users treat output as a starting point rather than a final answer. Half named quality control of AI output as a skill that is becoming more important.

That is the job now. Not producing the first version. Deciding whether the first version is right.

Which, for an accountant, is more or less what the qualification was always for.

So, plainly, what it should not be doing. It does not decide a treatment. It does not judge whether something is material. It does not tell a client what to do about it. It does not file anything and it does not sign anything.

Put the check in the process, not in somebody's memory

The plant went through because the check lived in somebody's head, and that day the head was busy. That is not a character flaw. It is what happens in every business in January.

The fix is to make the checking part of the workflow rather than a thing you remember to do afterwards.

Ask for the working. When it pulls figures off a document, have it put the page or the line it took each one from in the next column. A figure you can trace in four seconds is a figure you will actually check.

Give it something to reconcile against. If the invoice total is on the document, have it total its own extraction and tell you when the two disagree. That single instruction catches most of what goes wrong.

And ask it what it is least sure about. Genuinely useful, this one. It will name the three lines it guessed at, and those are the three worth your attention rather than all forty.

Will it replace accountants

The same UK government adoption plan estimates 13.7 per cent of roles in professional and business services are at risk of substitution, and 52.8 per cent likely to be significantly augmented. Roughly four to one in favour of the job changing rather than disappearing. The methodology behind it is not published, so read it as a considered estimate rather than a measurement.

My own view is duller than the headlines. The risk to a small practice is not that software replaces the staff. It is the practice two miles away that gets to the same standard of work in half the time and prices accordingly.

That is a more concrete thing to worry about, and a better reason to start now.

Where the trainees come in, and the bit I cannot answer

Here is the part I find genuinely difficult.

The Stanford Digital Economy Lab, working from payroll records covering millions of employees, found employment of 22 to 25 year olds in AI exposed occupations sitting 19 per cent below where it would have been had it kept pace with less exposed peers, mostly through reduced hiring rather than redundancies. The authors are careful about attributing that to AI, so I will be too.

Meanwhile the canonical study on this, Brynjolfsson and colleagues on 5,179 customer support agents, found the gain was 34 per cent for the least experienced and close to nothing for the most experienced.

Put those two together and the problem shows up. The entry level work is what AI does best, and the entry level work is also how somebody learns to tell when a set of numbers is wrong.

I do not have an answer to that. I am fairly suspicious of anyone who says they do.

Client data, said plainly

The MIT NANDA report, which is labelled preliminary and surveyed 153 leaders, found around 90 per cent of employees using personal AI tools for work while only around 40 per cent of their companies had bought subscriptions.

Small sample, so take the precision with a pinch of salt. The direction is right and every trainer I know sees it.

For a practice, that is the whole risk in one sentence. Not the software. A client's management accounts pasted into a free account on somebody's phone, on a train, with no record that it happened.

Banning the tools does not remove that. It just means you cannot see it.

Three things sort most of it. Pay for the business tier of whichever tool you use, where the terms say your data is not used for training, and check that for your specific tool rather than assuming it. Write one page naming what may and may not be pasted. Then spend twenty minutes showing people, because a policy nobody has read is a document rather than a control.

ICAEW and ACCA have both published guidance for members. Read your own institute's rather than my summary of it.

What it costs

A paid business seat on Copilot, Claude or ChatGPT is roughly £20 to £30 per person per month. If the practice already runs Microsoft 365, part of that is sitting there now.

The guides ranking for this subject quote implementation engagements from five to forty five thousand pounds. That is a real market and it is not the one most small practices are in.

For a practice of two to twenty people, the honest version is a handful of paid seats and somebody teaching the team to use them properly. Our own work is fixed price and starts at £500, and the first project gets chosen on how quickly it pays back rather than how impressive it sounds.

Bespoke builds come later, if at all, and usually only when you can describe the thing you need in one sentence and nothing off the shelf covers it.

What you could do on Monday

Take one client and last month's purchase invoices. Twenty of them will do.

Put them through a paid business account and ask for a table: supplier, invoice date, net, VAT, total, invoice number, plus a column saying which ones it was unsure about and why.

Then check all twenty against the paperwork yourself, and time both halves. How long the extraction took, and how long the checking took.

Those two numbers tell you more than any guide will, this one included. If the checking takes longer than typing would have, you have learned something useful for the price of an hour.

The honest limit

The Department for Science, Innovation and Technology found 75 per cent of UK businesses using AI report improved workforce productivity, while 77 per cent report no change in revenue and only 12 per cent report an increase.

That gap matters more in a practice than almost anywhere else. Time saved comes back in twenty minute pieces spread across a week, and twenty minutes is exactly the size of thing that quietly disappears into the next job.

So it is worth deciding in advance what the saved time is for. Advisory work you keep meaning to start, or getting out on a Thursday at five. Either is a fine answer. Neither happens by accident.

And none of this touches the thing most practices are actually short of, which is the records arriving on time. AI will not fix that. It will just get you through them faster once they land.