AI for accountancy firms

AI is most useful to a UK accountancy firm in three places: preparing workpapers, drafting client communications, and chasing records and deadlines. In each case the pattern is the same. The AI does the repetitive first pass, a qualified person reviews and signs off, and the firm keeps full control of client data. You do not need to replace your practice software or retrain your team to get started. Most firms begin with one bottleneck, prove the time saving on real jobs, then extend from there.

"MTD multiplies the deadlines, not the headcount. Firms that automate the chasing keep their people for judgement. The rest will hire just to stand still."

Dean Cookson, founder, Operosus

Why should accountants take AI seriously now?

Two pressures are arriving at once.

The first is regulatory. From 6 April 2026, sole traders and landlords with annual income from self-employment and property over £50,000 must keep digital records and submit quarterly updates under Making Tax Digital for Income Tax. For most practices that means more filing events per client per year, with the same headcount. The work that expands is exactly the work AI handles well: collecting records, categorising transactions, chasing the clients who have not sent anything.

The second is competitive. ONS data from late September 2025 shows 23% of UK businesses now use some form of AI, up three percentage points in a single quarter. Your clients are among them, and increasingly they expect their advisers to be too. The full adoption picture, with sources, is in our UK small business AI statistics table.

The profession itself is not resistant. Research by Ipsos for Chartered Accountants Worldwide, covering 2,718 chartered accountants across 48 countries, found 85% are willing to use AI given the opportunity, and 83% of those aged 18 to 24 already use AI tools at least once a week. The gap is not willingness. It is knowing which tasks to point it at, and how to do that without putting client data at risk.

Ainslie van Onselen, Chair of Chartered Accountants Worldwide, put the boundary well in that research: "AI is a tool for innovation, not a substitute for human expertise."

What can AI actually do in a practice?

The honest answer is narrower than the hype and more useful than the scepticism. AI is good at reading documents, drafting text, spotting things that look wrong, and following up persistently. It is not good at judgement, and it should never be the last set of eyes on anything that carries your firm's name.

The common thread: AI produces a draft or a flag, never a final answer.

How does AI help with workpapers?

Workpaper preparation is mostly transcription and tie-out, which is why it is the highest-value starting point for most firms.

A practical AI-assisted workpaper flow looks like this:

1. Ingest. Bank statements, supplier invoices and receipts go in as PDFs or photos. AI extraction tools pull dates, amounts, counterparties and VAT into structured data, including from scanned and photographed documents that traditional OCR struggles with.
2. Categorise. The AI proposes nominal codes based on the client's history. Where it is unsure, it says so rather than guessing, and those items go in a query list.
3. Reconcile and flag. Totals are tied to control accounts automatically. Anything that does not agree, or looks unusual against prior periods, is flagged with a short plain-English note.
4. Draft the schedule. Lead schedules and supporting notes are drafted in your template, with every figure traceable back to its source document.
5. Review. A qualified person works through the flags and the draft. The review is faster because the routine ninety per cent is already done and the exceptions are surfaced, not buried.

The discipline that makes this safe is traceability. Every number the AI produces should link back to the document it came from, so the reviewer is checking evidence, not trusting a black box.

Which workpaper tasks should you not give to AI?

Anything that involves judgement on treatment: revenue recognition decisions, provisions, going concern, anything close to a materiality threshold. AI can assemble the facts that inform those calls. The call itself is the job.

Can AI write client communications without sounding generic?

Yes, but only if it writes from your material rather than from the open internet.

Generic AI output is easy to spot: over-formal, faintly American, padded with filler. The fix is a pattern we have used repeatedly at Operosus, including in Bidwell, our tool that drafts tender responses for UK firms. Instead of asking a general model to "write a letter to a client", you give the system your firm's own past letters, templates and tone rules, and it drafts new correspondence from that library. The output sounds like your practice because it is built from your practice.

For an accountancy firm that means:

• Year-end letters and engagement updates drafted from last year's versions with the new figures and dates dropped in, then read once before sending.
• Query lists turned from a reviewer's terse bullet points into a polite, structured email a client can actually act on.
• Plain-English explainers of a tax position or a set of accounts, drafted for the client's level, reviewed by the person who did the work.
• Tone matching per client. The sole trader who texts you and the finance director who wants formality should not get the same letter, and a well-set-up system knows the difference.

The rule that keeps this safe is simple: AI drafts, a person sends. No client-facing message should leave the building without a human having read it. In practice that read takes seconds, because you are editing a decent draft rather than writing from a blank page.

What about chasing records and deadlines?

Deadline chasing is the least glamorous problem in practice and the one clients silently judge you on. It is also the most automatable, because it is not really a writing problem. It is a state-tracking problem: who owes what, by when, and what happens if they do not respond.

We have built this pattern for clients outside accountancy, including a veterinary booking service where bookings, payments and consent forms each had to be requested, chased and confirmed in sequence. The closely related discipline of chasing unpaid invoices with AI uses the same state-tracking shape. The structure transfers directly to practice work:

1. Every client has a record of what is outstanding: bank statements, payroll data, approval of draft accounts, a signed engagement letter.
2. Reminders go out automatically on a schedule, each one specific about what is missing rather than a vague "please send your records".
3. When the client responds, the system marks the item received and stops chasing. Nobody gets a reminder for something they sent last week, which is the failure that makes automated chasing feel robotic.
4. Non-responders escalate to a human after a set number of attempts, so the difficult conversations still happen, and happen earlier.

With quarterly MTD updates multiplying the number of deadlines per client from April 2026, this is the area where small firms will feel the squeeze first. A practice that chases manually at four times the frequency either hires for it or drops the ball. A practice that automates the chasing keeps its people for the work that needs them.

Is client data safe with AI tools?

It can be, but not by default, and this is the concern practitioners raise most often. In the Chartered Accountants Worldwide research, data security was the most-cited barrier to AI use, named by 33% of respondents.

The risks are real but manageable. The non-negotiables:

• Use business-tier tools, not free consumer accounts. Paid business tiers of the major AI providers contractually exclude your data from model training. Free tiers often do not.
• Check where data is processed. UK GDPR applies regardless of where the AI vendor sits. You need a data processing agreement, and you need to know the answer when a client asks where their data goes.
• Update engagement letters. If AI tools process client data as part of your service, say so. Most clients will not mind. All of them are entitled to know.
• Keep personal data out of general-purpose chatbots. Pasting a client's payroll into a consumer chatbot is a data breach waiting to be written up. Purpose-built tools with proper agreements exist for a reason.
• Write a short AI policy. One page: which tools are approved, what data can go into them, who reviews AI output before it goes to a client. ICAEW publishes guidance on AI for members that is a sensible reference point.

A firm that can explain its AI controls clearly has a marketing asset, not a liability. Clients are reading the same headlines you are.

Should you buy off-the-shelf tools or build something bespoke?

Both, usually, in that order.

Off-the-shelf AI features inside the software you already run, such as bookkeeping platforms and practice management suites, are the cheapest way to start and need no integration work. Their limitation is that they automate the steps the vendor chose, not the way your practice actually works.

Bespoke automation earns its keep where your process is specific to you: the way your firm chases records across multiple systems, the house style of your client letters, a workpaper format your reviewers have refined over a decade. That is connective work, joining your inbox, your practice management system and an AI model into one flow, and it is smaller and cheaper than firms expect because it builds on the tools you already have rather than replacing them.

The wrong move is the big-bang platform migration in the name of AI. The firms getting value are the ones automating one painful step at a time and keeping what already works.

Where to start

Pick one bottleneck, not a transformation programme.

1. Choose the job your team complains about most. For most practices it is records chasing or workpaper prep. One process, not five.
2. Switch on the AI features in software you already pay for. Prove the habit before spending anything new.
3. Set the review rule on day one. AI drafts, a person signs off. Write it down and apply it to everything.
4. Sort the data basics in week one. Business-tier accounts, a one-page policy, a line in your engagement letters.
5. Measure one number. Hours on the task before and after, on real jobs. If it is not saving time within a month, change the task, not the headcount.
6. Then consider bespoke. Once the easy wins are banked, the remaining friction is usually specific to your firm, and that is where tailored automation pays.

Operosus builds AI systems for UK firms that follow exactly this pattern: start with the bottleneck, build on what you already run, keep a person on every sign-off. Our accountancy industry page shows the specific systems we build for practices, and our guides to client onboarding automation and automated weekly reporting cover the two bottlenecks that usually come next. If you want a second opinion on where AI would pay back first in your practice, we are happy to give one.