Buyers pricing an accountancy practice on the staff AI might replace are reading the wrong line: the value, and the risk, sit in the client records.
The investment case fits on one slide. Accountancy is fragmented, recurring and sticky. Clients tend to stay, fees arrive monthly, and much of the work looks repetitive. Buy enough small practices, run them on the same AI, and the cost of serving each client should fall while the fees stay put.
I run a UK practice that uses AI every working day, and I think half of that slide is right. The other half is where buyers will lose money.
What AI already does well
Start with the good news, because there is plenty of it.
Sorting transactions is the obvious win. A small online seller can produce thousands of lines a quarter: sales, refunds, platform fees, postage, stock. Deciding the VAT treatment of each line used to take a large part of a bookkeeper’s week. A model does a first pass in minutes, and a person checks the exceptions rather than every line, provided someone has told the model the rules of the business it is looking at.
Matching bank lines to receipts and invoices is the second. It is pattern work: amounts, dates, names that are nearly the same. A machine is better at tedious matching than a tired person at the end of a quarter.
The third is drafting client email. Much of what an accountancy firm sends is routine: a deadline, a request for a document, confirmation that something has been filed. A draft that a qualified person checks and sends is quicker than a blank screen.
None of this is speculative. It is how a well-run practice already works. If the thesis were simply “AI makes the repetitive work cheaper”, I would sign it.
Where it stalls
We tested the hardest of those three directly. We took the drafts our AI produced for real client emails and compared them with the replies our accountants actually sent to the same people.
The style was close. The gap was content. Most of what our accountants wrote did not come from the client’s email at all. It came from the records: what the ledger showed, what had just been filed, where the job had got to, what had been agreed in earlier correspondence.
A model cannot write “your return went in on Tuesday” if nothing it can see records that the return went in on Tuesday. Better prompts do not fix that. A bigger model does not fix it either. The records were the limit, not the model.
That finding travels well beyond email. In an accountancy firm, the hard part of almost every answer is knowing the current state of that client’s affairs. The intelligence is now cheap. The context is not.
What that means for a buyer
If the records are the ceiling, the records are what you are buying. Valuing a practice on the headcount AI might save assumes the model can see what the staff can see. In plenty of small firms it cannot, because the knowledge lives in an inbox, a spreadsheet on one person’s desktop, or one long-serving manager’s head.
So diligence the data, not the demo. Four questions I would ask of any practice before pricing in an AI margin:
- Is each client’s position held in a system, or in people? If the senior manager left tomorrow, could anyone say where every job stood?
- How complete are the fields that matter? A practice management system where half the records do not say which service the client pays for is a list of names, not a dataset.
- Are the clients’ own books digital end to end? For VAT this is already the law. Making Tax Digital requires VAT-registered businesses to keep records digitally and file using software, and HMRC’s notice says that manually transferring data between programs “is not acceptable”. A practice whose clients still hand over a box of receipts each quarter is a long way from the margin on the slide.
- Is the correspondence tied to the client record? Earlier emails are where the promises live. A model that cannot read them will eventually contradict them.
Clean data compounds. Every quarter a practice records properly is a quarter the next model can use. Messy data compounds as well, in the other direction.
A roll-up that bolts ten messy practices together does not get one mess. It gets ten different ones.
AI moves the work, not the liability
There is a second line on the slide worth correcting: the idea that review is a cost that fades as the models improve.
Two obligations sit underneath every return an accountancy firm files, and both sit with people. The client stays responsible for what goes in under their name. HMRC’s factsheet on inaccuracy penalties is plain that if you ask someone else, such as an adviser, to act for you, “you must do everything you reasonably can to prevent any inaccuracies.” The adviser, separately, is held to HMRC’s standard for agents, which expects agents to “avoid including figures in returns or claims which are unsubstantiated or speculative.”
Neither obligation transfers to software. A model can propose a figure. It cannot be held to account for one, and it cannot sit across from an HMRC officer and explain why it treated a line the way it did. A qualified person still has to look, and has to be able to say why.
So the review layer is not overhead waiting to be automated away. It is the product. What AI changes is how the reviewer spends the day: less keying, more judgement.
In our practice that is the trade we are making, and it is a good one. But it moves the work. It does not move the liability, and a business plan that assumes otherwise is priced on a cost that will not disappear.
The practices worth buying are the ones that already know where every client stands and can show it from the records rather than from somebody’s memory. Point the model at those and the slide starts to come true.
