AI for accountancy practices: what UK firms automate first, and what they should not.

Three regulatory changes are landing on UK practices at once, and each one is a workload rather than a policy. Making Tax Digital for Income Tax began on 6 April 2026 for sole traders and landlords with qualifying income over £50,000 in 2024/25; those over £30,000 join from 6 April 2027 and those over £20,000 from 6 April 2028. Each client in scope must keep digital records, submit quarterly updates through compatible software, and file the return by 31 January. The penalty regime is points-based: a point per missed deadline, a £200 penalty at four points, and £200 for each miss after that.

Companies House identity verification became a legal requirement on 18 November 2025 for new directors and people with significant control, with existing directors verifying at their next confirmation statement, and accountants registered as Authorised Corporate Service Providers doing the verification for clients. And in October 2025 HM Treasury decided that AML supervision of accountancy service providers will move to the FCA, a single professional services supervisor for a population of around 60,000 firms, subject to legislation.

The common thread is volume of routine, deadline-bound contact with clients. That is what AI is good at, and it is why the first things practices automate are not the clever things. A practice with 100 clients entering MTD in April 2027 is taking on 400 new quarterly cycles a year, each of which starts with a request for records and ends with a submission somebody has to check, and the practice has the same people it had last year. This guide is about where those hours go, what the tools that recover them cost, and where the line sits between what a system may do and what an accountant must.

Sage’s 2024 survey of 1,000 accountants and bookkeepers found 92 per cent agreed they spend too much time on manual tasks, 89 per cent said automating would free up time, and only 37 per cent were doing it. Karbon’s 2025 report found 85 per cent of accounting professionals were excited or intrigued by AI while only 37 per cent of firms invested in AI training; its 2026 report says 98 per cent of accounting professionals now report using AI, with data security concerns at 83 per cent. Karbon is a practice management vendor and those are its own figures, from gated reports whose sample sizes are not on the landing pages.

ACCA’s Global Talent Trends 2026 found 82 per cent of respondents confident in their ability to learn and apply AI, while 51 per cent remain worried about its impact on their jobs; its 2025 edition, from over 10,000 people in 175 countries, found only 32 per cent said their organisation was providing opportunities to learn AI skills. ACCA and CA ANZ’s July 2026 survey of 1,600 finance professionals put the top barriers at data quality problems (42 per cent), lack of appropriate skills (42 per cent) and integrating multiple data sources (40 per cent).

The official statistics are more sober. The ONS found around 35 per cent of UK businesses with ten or more employees used AI in June 2026, up from about 12 per cent in late 2023, and only 11 per cent had trained more than half their workforce. DSIT’s adoption research across 3,500 businesses found 60 per cent cited limited AI skills, expertise or knowledge as a barrier. We could not read ICAEW’s own survey pages on the day, and have left its figures out rather than quote them second-hand.

Records capture is already automated in most practices and is the least interesting. Dext, AutoEntry and the capture built into Xero, Sage and QuickBooks read receipts and invoices and post them; the remaining work is exceptions. If your practice still keys purchase invoices by hand, start there, and the tools table below has the prices. The generative features the software vendors now bundle, Xero’s JAX and Sage Copilot among them, sit on top of capture and answer questions about the ledger; they are worth switching on because they cost nothing extra, and they are not a substitute for the three jobs below, which happen between the practice and the client rather than inside the ledger.

Client chasing and records requests are the largest recoverable block of hours in a practice and the one most vendors do least well, because a scheduled reminder is not a chase. A real chase knows the client has sent the bank statements but not the sales ledger, asks for exactly that, notices when it arrives, and escalates to a person when the deadline is close. With MTD adding four quarterly cycles a year per client, this is where the 2026 to 2028 hours will go.

Onboarding is the third, because it now contains AML checks, Companies House verification and HMRC authorisation with their own deadlines; our separate guide covers it step by step. Query triage is the fourth: the "quick question" emails that arrive daily and interrupt fee-earning work. An agent that reads them, answers the ones that are status or process questions from the practice’s own records, drafts a reply for the ones that need a person, and routes the ones that are advice to the right accountant recovers time without ever giving advice itself.

Prices read on the vendors’ UK pages on 14 September 2026, excluding VAT where the page says so. Introductory discounts are noted but the standard price is the one to budget on. Two things stand out. The bookkeeping platforms now include their AI assistants at no extra charge, so the marginal cost of that layer is zero for a practice already paying for the ledger. And none of the practice suites in the table publishes a price for the kind of agent that chases clients across systems, because none of them sells one; that is the gap a bespoke build fills.

Karbon displays dollars. Xero, Sage and QuickBooks show six-month introductory discounts; the standard monthly price is given. Prices change; the vendor page is the authority.

A tool does one thing inside one product. An agent works across the products a practice already pays for, and it is worth paying for when the work is the joins. The test is simple: if a person in your practice spends their day moving information between two screens and writing the same email in slightly different words, that is agent work. If they spend it inside one product doing what the product was built for, buy the product’s own feature instead. Four cases, in the order we would build them.

The build price is the visible number and the running cost is the one that decides whether the agent survives its first year. For a chasing and triage agent the running cost has three parts: the model calls, which read documents and draft messages; the channels, which are email and text messages to clients; and the verification checks at onboarding, which are £2 to £3 each at the published Xama rates. In our own production systems the language model has been a single-digit share of the variable spend on AI features, with messaging the largest line, and a practice should expect the same shape.

What that means in practice is that a proposal quoting the model cost and nothing else is incomplete, and that the right question to a supplier is what the agent will cost per client per month once it is live, including the messages it sends and the checks it runs. A first agent from us comes with that figure in writing before anything is built.

Advice, and anything a client will rely on as advice. HMRC’s standard for agents says agents should work to prevent errors in their clients’ tax calculations, avoid including figures that are unsubstantiated or speculative, take reasonable steps to make sure any third-party input used in acting for a client, including software, provides accurate results, and keep contemporary records of what they advise and when. Nothing in that standard is delegated to a model. An agent can draft; the accountant signs.

The AML risk assessment and the decision that enhanced due diligence applies. The Money Laundering Regulations require a relevant person to make those judgements, and the CCAB guidance approved by HM Treasury is written for people. A system should collect, flag and record; it should not rate. The same logic applies to the sign-off on a set of accounts or a tax return: a system can check that every schedule is present and every figure reconciles, and can tell the reviewer what changed since last year, but the review is the accountant’s and the practice’s name is on it.

Anything the practice cannot explain. Karbon’s own 2026 report puts data security concerns among accounting professionals at 83 per cent, and the practical answer is not to avoid AI but to be able to say, for every automated step, what data went where and who approved the output. If the supplier cannot show you that in writing before the build, the build is not ready.

Take a practice of twelve with 300 clients, of which 100 are in scope for MTD for Income Tax by April 2027. Price the time at the ONS 2025 median gross hourly pay for the roles that do the work: £16.77 for a full-time bookkeeper, payroll manager or wages clerk and £26.02 for a full-time chartered or certified accountant. Employer National Insurance and pension are on top, so the numbers understate.

Chasing by hand: assume 90 minutes of administrative time per client per year across year-end records, sign-offs and follow-ups, plus 30 minutes of an accountant’s time dealing with the queries that chasing generates. That is 450 hours at £16.77 (£7,547) and 150 hours at £26.02 (£3,903), or £11,450 a year. Add the MTD quarterly cycle: 100 clients, four updates, 30 minutes each, 200 hours at £16.77, or £3,354. Total: about £14,800 a year of gross pay on chasing and the questions it generates.

An agent does not remove all of it, because a person still approves escalations and answers the advice questions. Assume it removes two-thirds: about £9,900 a year. Tooling on top: Dext Solo for 300 clients at £2.50 each is £750 a month, £9,000 a year, though most practices already pay for capture in some form, and a practice manager at £32 to £58 per user a month is likewise usually already there. The new spend is the agent at £4,000 to £15,000 once, plus a few pounds a month to run. Payback is five to eighteen months on time alone, and the practice enters the 2027 MTD tier with the quarterly chase already running.

Hours are assumptions, stated so you can substitute your own. Pay rates are ONS ASHE 2025 provisional, Table 14.5a, full-time median gross hourly pay for SOC 4122 and SOC 2421.

First, make sure capture is switched on and used; that is a tool, not a project. Second, automate the chase, because it runs every day and now has statutory deadlines. Third, onboarding, because every new client goes through it and the compliance steps inside it are a checklist a system can run. Fourth, triage, once the practice trusts the agent with the first three. Advice and sign-off never.

Before any of it, train the team. The ONS finds only 11 per cent of businesses have trained more than half their people, and DSIT’s employer survey found the most common barrier was not knowing which training is relevant. A half day that agrees what may be pasted where, and shows each role the two or three things a model does well in their work, costs less than a month of the tools above and removes the risk that somebody has already put a client’s bank statement into a consumer chatbot.

Augustova’s ladder for a practice is an AI adoption audit at £2,500 to £6,000, credited against a build, a first agent at £4,000 to £15,000, and a custom AI system at £8,000 to £25,000 where the practice wants chasing, onboarding and triage together with reporting. Training for the team is £1,200 for a half day and £2,000 for a full day, which is where most practices should start if nobody can yet say what may be pasted into a chatbot. All fixed, excluding VAT, no day rates.

We read the HMRC and Companies House guidance on GOV.UK, HM Treasury’s October 2025 consultation response, the ONS and DSIT statistics, and the ACCA, Karbon and Sage survey pages, on 14 September 2026, and quote the figures as published with base and date where the page gives them. Karbon’s figures are a vendor’s own and their sample sizes are not on the landing pages; they are labelled as such. ICAEW’s survey pages could not be read by our tools on the day and are not cited. Vendor prices were read on the vendors’ UK pricing pages the same day. Pay figures are from the ONS ASHE 2025 provisional Table 14, downloaded and read directly. Hours in the worked example are assumptions. Nothing here is tax or legal advice.