Every practice owner has now been told that AI will change the profession. Few have been shown which Tuesday-morning jobs it can take off their team this year, and which ones it will get wrong with total confidence. That second list matters more, because one confidently wrong answer to a client undoes a year of time saved.
This guide sorts the work of a typical small practice into three piles: jobs AI does well now, jobs it can help with under supervision, and jobs to keep away from it. Then it covers the controls that make any of it safe to use with client data.
What AI does well in a practice today
The common thread is reading and drafting. A language model is very good at turning a messy input into a tidy first draft, and at spotting what is missing from a set of documents. It is poor at judgement and at arithmetic it has not been given the tools to check.
- Drafting client emails. Records chases, deadline reminders, replies to routine questions. A model that can see the client's record writes a usable first draft in seconds, and a person reads it before it goes.
- Summarising HMRC letters and Companies House notices. What the letter says, what it asks for, by when, and which client it belongs to. This is reading, which models do well.
- Turning a pile of papers into a query list. Bank statements, receipts and last year's file in; a list of gaps and questions out, ready for the senior to check.
- Answering staff questions from your own procedures. "How do we handle a client who wants to change year end?" answered from the firm's own manual, with the page it came from. The important part is that it answers from your documents, never from memory.
- First-pass categorisation. Suggesting a nominal code for a bank line, with a confidence level, for a bookkeeper to accept or change.
AI drafts and a person sends. Every piece of work above ends in an approval step. The value is in the drafting and the noticing; the decision stays with somebody who is accountable for it.
Where it helps under supervision
Some jobs sit in between. AI can do a useful part of the work, provided a qualified person reviews every output.
Pre-review of a set of accounts. A model can compare this year's draft with last year's, flag movements above a threshold, and note anything unusual, such as debtors up by half or a director's loan account that has swung overdrawn. It cannot decide whether those movements are right. It saves the reviewer from finding them, and leaves the judgement with the reviewer.
Explaining figures to clients. A plain-English paragraph on why profit moved can be drafted from the numbers. The partner then edits it, because the partner knows the conversation that sits behind it.
Spotting work the client needs. Reading a client's accounts and Companies House record to suggest that they may need a pension conversation, a review of their salary and dividend split, or help with an overdue confirmation statement. The suggestion has to show the figure that raised it, so a person can check the reasoning in seconds.
What to keep away from AI
Three kinds of work should not be handed to a model at all, however good the demo looks.
- Tax rates, thresholds and deadlines from memory. A model trained a year ago will quote last year's figures with complete confidence. If a tool answers tax questions, it must look the answer up in a source you maintain, and show where it came from.
- Anything sent to a client without a person reading it. A wrong reminder is a small cost. A wrong statement about someone's tax position, sent in the firm's name, is a professional problem.
- Filing. Submissions to HMRC and Companies House stay in regulated software, approved by a person, with the client's agreement recorded.
Questions to ask any supplier about client data
A practice holds some of the most sensitive data a small business has. Before any AI tool touches it, get written answers to these:
- Where is the data processed and stored? UK or EU, and named.
- Is it used to train anyone's model? The answer should be no, in the contract as well as on the website.
- Who can see it? Your staff, by role, with a log of every access.
- What happens when you leave? Your data returned in a usable form, and deleted from their systems, with confirmation.
- Is there a data processing agreement? If they cannot produce one, stop there.
Add the tool to your privacy notice and your engagement letter's list of processors. Clients do not object to a firm using good tools. They object to finding out afterwards.
Where to start
Pick one job your team repeats more than a hundred times a year, where a mistake is cheap and easy to spot. Records chasing is the usual winner. Run it for a month with every message approved by a person, count the minutes saved and the drafts that needed changing, and decide from those numbers whether to go further.
Avoid starting with the most impressive demo. Start with the dullest job that eats the most hours. That is where the time comes back, and where the team learns to trust the tool before it goes anywhere near a judgement call.