Professional services firms have a specific problem with AI. The advice written for e-commerce and marketing teams does not survive contact with a regulated practice, because the failure mode is not a bad headline, it is a wrong number on a filing or a citation that does not exist. What follows is a practical guide for UK accountants and solicitors who want the drafting speed without the exposure.
Is AI for accountants in the UK actually worth the setup time?
The honest version is that AI does not make you a better accountant or a better solicitor. It makes the typing around the work disappear. A small practice partner can lose a large share of the working week to emails, file notes, query lists, engagement letter variations and explanatory covering notes. None of that is billable in any satisfying way, and all of it is templated in structure while varying in detail. That is the exact shape of task where a language model earns its licence fee.
Adoption in the sector is real but uneven. In the Solicitors Regulation Authority's survey of SRA-regulated firms, just over a third, 37 per cent, said they were currently using legal technology, with a further 24 per cent planning to do so (SRA, Technology and Innovation in Legal Services). Most firms already run cloud storage and video meetings without thinking about it, so the infrastructure question is largely settled. The drafting layer is where the gap sits.
What can AI safely draft in an accountancy practice?
The reliable list in a small UK practice looks like this. Client query lists built from a bank statement export, where the model flags transactions that look personal, unusual or missing a receipt, and you decide which ones actually need asking about. Covering notes that explain a set of accounts in plain English for a director who does not read management accounts. Chaser sequences for outstanding records, which is a genuinely miserable job to do by hand and a well understood one to automate, covered in more detail in this guide to automated invoice and record chasing for UK firms. File notes written up from a client call, which is the single highest return task in most practices because it turns a fifteen minute write-up into a two minute review.
What changes the quality is feeding the model your own materials. A generic prompt produces generic output that reads like a template because it is one. A prompt that includes three of your existing client letters, your firm's tone rules and the specific client's situation produces something you edit rather than rewrite. That distinction is the whole difference between a practice that keeps using AI and one that abandons it in week three. There is a fuller walkthrough of the accounts side in this piece on running small business accounts with Claude.
What can AI safely draft in a solicitors' firm?
Legal drafting splits cleanly into two categories, and the split is not by document type. It is by whether the output contains an assertion about the outside world. A client update explaining what happened at a hearing contains assertions you already know to be true. A skeleton argument contains assertions about authorities, and those need checking one by one against the actual reports.
The SRA's own risk analysis names accuracy and bias as the problems to plan around, noting that these can cause AI to produce incorrect and possibly harmful results, either through hallucinations or amplification of bias in the training data (SRA, Risk Outlook report on AI in the legal market, published November 2023). The same report put AI use among the largest solicitors' firms at around three quarters. Large firms got there first because they could afford a verification layer. Small firms need a cheaper version of the same discipline.
What should never be handed to AI in a regulated practice?
There is a harder line than most guidance admits. It is not about the sensitivity of the task, it is about whether an error surfaces before or after it causes damage. A badly drafted client email surfaces immediately, because the client replies confused and you fix it. A wrong figure in a submission surfaces months later, in correspondence with HMRC, after interest has accrued.
The consequences sit with the professional. In Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank, decided on 6 June 2025, the Divisional Court considered material placed before the court containing case citations that did not exist, and set out that AI is a tool carrying risks as well as opportunities, whose use must take place with an appropriate degree of oversight and within a regulatory framework that ensures compliance with established professional and ethical standards (Courts and Tribunals Judiciary, Ayinde v Haringey). That is not an argument against using AI. It is an argument for knowing exactly which parts of a draft you have personally verified.
How does Making Tax Digital change the case for AI in 2026?
Making Tax Digital for Income Tax went live on 6 April 2026 for sole traders and landlords with qualifying income above £50,000, and HMRC put the size of that first group at around 780,000 people (GOV.UK, One year until Making Tax Digital for Income Tax launches). The same announcement sets out the phasing: qualifying income above £30,000 is scheduled to join from April 2027, with the threshold dropping to £20,000 from April 2028.
For a practice with a hundred affected clients, that is a structural change to the calendar rather than a software change. Four reminder cycles, four query rounds, four sets of confirmations. Practices absorbing that with the same headcount are almost all doing it by templating the correspondence, which is what makes 2026 a different adoption argument than 2024 was. The related pressure is on record keeping quality at the client end, which is covered in this look at AI bookkeeping for UK small businesses.
How do you keep client confidentiality when using AI?
The SRA's analysis flags client confidentiality as a distinct risk, covering both exposure to third parties and keeping sensitive information secure inside firms and with system providers. In practice, four controls cover most of it.
Buy business or enterprise licences rather than letting staff use personal free accounts, because the data handling terms differ materially. Define a short list of what never gets pasted, typically anything identifying a client in a matter involving allegations, plus bank details and identity documents. Prefer pseudonymised inputs where the model does not need the name to do the task, since a chronology works perfectly well with "the claimant" throughout. Log which matters have used AI assistance, so that if a question ever arises you can answer it from a record rather than from memory.
What does a first month of adoption look like in a small practice?
The sequence that works in firms of two to twenty people is deliberately narrow. Week one, choose the task and collect five good examples of the output you already produce. Week two, build a prompt that includes those examples, your tone rules and the specific inputs, then run it on live work while still doing the job the old way. Week three, stop doing it the old way but keep a named reviewer on every output. Week four, count how long it now takes and decide whether to add a second task.
The parallel running in week two is the step everyone wants to skip and the one that determines whether the rollout survives, because it produces the evidence that convinces the sceptical partner. Documenting the resulting process properly matters too, and there is a practical method in this guide to AI-assisted business process documentation.
How much time does drafting with AI really save?
Be suspicious of any blanket productivity percentage quoted for professional services, because the variance between firms is larger than the average. What is consistent is the shape of the saving. Tasks with a stable structure and variable detail, such as file notes, chasers and client updates, compress substantially. Tasks requiring judgement compress barely at all, and adding a verification step to them can make the total time worse.
That is why the measurement has to be per task rather than per person. Time three file notes written manually, then three written with a template and reviewed. If the second set is not meaningfully faster after the first fortnight, the prompt is wrong, not the technology. Meeting write-ups in particular tend to be the clearest early win, and this walkthrough of AI meeting notes for small businesses covers the mechanics.
Which tasks should you automate first, and in what order?
A workable order for a UK practice runs: internal file notes, then client chasers and reminders, then query lists built from records, then explanatory covering notes, then engagement letter variations from your own template, then first drafts of routine correspondence. Everything beyond that involves a professional position and belongs in the verified category.
Resist the temptation to start with the impressive task. The demo that gets a partner excited is usually the one with the highest error cost, which means the first mistake lands somewhere expensive and the experiment ends. Starting with the boring task means the first mistake lands in an internal note that somebody corrects in thirty seconds.
What are the honest trade-offs?
Review fatigue is the underrated one. A draft that is mostly correct and confidently written is harder to review carefully than a rough draft, because your attention slides over the fluent parts. The colour split described earlier exists precisely to force attention back onto the assertions that matter.
The junior training gap needs a deliberate answer. If trainees and juniors stop writing first drafts, they lose the reps that built judgement, so some firms have them draft manually first and then compare against the AI version, which turns a lost exercise into a critique exercise. That is more work, not less, for the first year of someone's career, and it is a cost worth paying.
The setup cost is simply real. Encoding your house style, precedents and tone into prompt templates takes days, not hours, and it is the work that separates output you edit from output you rewrite. For the broader picture of where this fits alongside other paperwork automation, see automating paperwork with Claude AI, and the hub covering Claude AI for UK small businesses collects the workflows by task.
The practices getting value from this in 2026 are not the ones with the best tools. They are the ones that picked one dull, high volume task, encoded their own documents into it, and put a named human on every output before it left the building.
AI for Accountants and Solicitors — FAQ
Is it safe to use AI for accountants in the UK under professional rules?
Yes, provided the human stays accountable for the output. The Solicitors Regulation Authority is explicit that solicitors remain accountable to clients for the service provided whether or not external AI is used, and the same logic applies to ICAEW and ACCA members signing off accounts. Safe use means three things in practice. First, a written policy naming which tools staff may use and on what data. Second, a review step before anything leaves the building, with the reviewer named on the file. Third, a paid business or team plan with training switched off, rather than a free consumer account. If a tool cannot tell you where client data is processed and whether it trains on your inputs, it does not belong near a client matter.
What should an accountancy practice automate first?
Start with the work that is high volume, low judgement and internally facing, because mistakes there are cheap and get caught fast. Typical first candidates are drafting client emails and chasers, turning meeting notes into file notes, summarising bank statements into a query list for the client, drafting engagement letter variations from a firm template, and writing the plain English explanation that sits above a set of accounts. Leave anything that touches a filing, a tax position or a signature until the drafting workflows have been running cleanly for a month or two. The order matters more than the tool. Practices that start with the filing get scared off. Practices that start with the chaser email keep going.
Will AI replace bookkeepers and paralegals?
Not on current evidence, though the shape of both jobs changes. The Solicitors Regulation Authority found that firms could use AI to complete administrative tasks more efficiently and free up staff capacity for more complex work, which is a redeployment argument rather than a replacement one. The pattern seen most often in small UK practices is that the junior person stops producing first drafts and starts reviewing them, which means the skill that gets valued shifts from typing speed to judgement about what is wrong. The awkward part is training. If juniors never write the first draft, they need a deliberate substitute for the learning that used to come from writing it.
How much does this cost to set up in a small practice?
The licence cost is usually the small part. A business AI subscription for a handful of fee earners is a modest monthly figure, and most firms already pay for practice management and cloud storage. The real cost is the internal time to build prompt templates that match your house style, write the AI policy, and run a review period where every output is checked and corrected. Budget that as a project rather than a purchase. Firms that skip the template building step get generic output, decide AI does not work for professional services, and stop. Firms that spend a fortnight encoding their own precedents and tone get drafts that need light editing rather than rewriting.
What happens if AI gets something wrong in a client matter?
The professional consequences land on the human, not the software. In Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank, decided on 6 June 2025, the Divisional Court dealt with material put before the court containing case citations that did not exist, and observed that AI use must take place with an appropriate degree of oversight and within a regulatory framework that ensures compliance with established professional and ethical standards. The practical control is verification at the point of citation or figure. Every case reference gets opened. Every number gets traced to a source document. Anything the model asserts that you cannot check gets deleted rather than softened.
Does Making Tax Digital make AI more useful or less?
More useful, because the reporting rhythm changed. Making Tax Digital for Income Tax went live on 6 April 2026 for sole traders and landlords with qualifying income above £50,000, and HMRC put the size of that first group at around 780,000 people. Quarterly updates replace one annual scramble, which multiplies the number of client conversations a practice has each year even though the underlying work per client barely changes. That is chase emails, query lists, reminders and explanations, four times over. Those are exactly the drafting tasks that respond well to templated AI output with a human check, and they are the ones that quietly consume fee earner hours otherwise.
