AI for small UK accountancy practices: what actually works in 2026

It's the second week of March in a 9-staff accountancy practice somewhere in Sheffield or Cardiff. Self-Assessment season is just behind, year-end is just ahead, and the partner is staring at a list of 47 clients whose accounts need preparing for 31 March year-ends. Three of those client files are missing bank statements. Six are missing receipts for the last quarter. Two haven't responded to MLR refresh requests since November.

Her practice management system has an "AI insights" dashboard. It told her this morning that 47 jobs are open. She knew that.

Behind her, a junior is manually copying numbers from a Xero report into a working paper template. A senior is replying to client emails one at a time, two of them about the same VAT return question. Nobody in the room has touched a real automation since they bought into Karbon two years ago.

This is the practice I build for. Three to thirty staff, UK-based, running on Karbon, AccountancyManager, IRIS, CCH, or some mix, losing fee-earner hours every week to admin the software was supposed to eat. This is what I'd want a partner of one of those firms to read before they buy another "AI for accountants" pitch.

The short answer

AI is genuinely useful in UK accountancy practices in 2026, but most of the value is in narrow workflows the big vendors aren't selling: client onboarding and AML, year-end working-paper preparation, missing-records chase, email triage, capacity tracking. The work sits outside the practice management system, not inside it. For a small UK practice (3-30 staff), expect to save 4-10 hours per fee earner per week once built properly, and to pay between £15-£60 per user per month for PMS-bundled and standalone AI tools or £12,500 upwards for a custom automation layer that targets one or two high-volume workflows end to end.

Why automation is stalled in most small UK practices

The accountancy software market has been selling automation since the early 2000s. Practice management systems, time recorders, document portals, MTD-compliant bridging software. Most of it is genuinely useful. None of it solved the actual problem, which is that a 9-staff firm in Sheffield still spends most of February through April on the same admin loops it ran in 2018. Three reasons.

The PMS vendors lock most of the value inside their own product. Karbon, AccountancyManager, IRIS, CCH, Sage Practice. Each of them sells "AI" features (drafting client emails, summarising tax updates, suggesting next actions). The features are real but deliberately narrow. The vendor isn't going to build the workflow that pulls data out of their own product into a custom dashboard, or the workflow that automates a process across their PMS plus Xero plus DocuSign plus your AML provider. That's the practice's problem, and most practices don't have anyone who can solve it.

The MLR and AML regime forces caution. UK accountants are regulated under the Money Laundering Regulations and supervised by ICAEW, ACCA, AAT, CIOT, or HMRC depending on membership. Every client onboarding has to clear MLR checks. Every year, refreshes. Every high-risk client, enhanced due diligence. A partner who automates AML wrongly faces fines, loss of supervisory status, and personal liability. So the default stance on automating anything that touches AML is to do nothing. Which means the entire onboarding process stays manual, even though the parts that aren't AML risk-scoring (document collection, engagement letter drafting, ID verification chase) are completely safe to automate.

Partners who run the practice rarely have time or appetite to learn what the new tooling can do. This is the cultural piece. The same partner running 47 client files is also responsible for technology decisions. They reasonably triage: file the accounts, then think about software. Six months later, no software change has happened, and the same workflow is bleeding hours. Karbon's research on AI adoption in accounting reports that confidence in AI is growing but training and adoption lag well behind. Every small practice I've talked to confirms it: the partner reads AccountingWEB, attends a CPD webinar, knows the names of the tools, has bought none of them.

The practices that get this right stop waiting for the PMS roadmap, and stop trying to learn every tool, and instead build a narrow automation layer alongside their existing stack that targets the two or three workflows where the cost is bleeding the most.

What's actually worth automating in a 3-30 staff UK practice

Every accounting publication will tell you to automate everything. That's vendor copy. In practice some automations compound and some waste build hours. Here's what I've seen actually work in small UK practices, in rough order of ROI per workflow.

1. Client onboarding and MLR document collection

A new client agrees to engage. The current process: partner sends an MLR information request by email, chases for ID and proof of address, runs the AML check through their provider (Smartsearch, Veriphy, AMLCC, or similar), drafts the engagement letter from a Word template, sends for signature, sets up the client in the PMS, sets up the client in Xero or QuickBooks, requests authorisation codes from HMRC. The whole thing takes 90-180 minutes spread across a week or two of chasing.

A workflow that collapses this into a structured intake form, an automated AML submission to your provider's API, an LLM-drafted engagement letter with the matter-specific paragraphs filled in, a DocuSign send, automatic PMS and bookkeeping-software setup on signature, and an HMRC authorisation reminder, gets the partner out of most of that loop. Human review at the AML risk-decision step and the final letter sign-off only.

Time saved: 60-120 minutes per new client onboarding. Multiply by the 30-100 new clients a small UK practice onboards per year.

Build complexity: medium-high. The AML provider integration is the hardest. The LLM drafting is the easiest.

2. Year-end working-paper preparation

Every set of accounts has the same prep: pull the trial balance from the bookkeeping software, format into the firm's working-paper template, flag the obvious queries (overdrawn director's loan, unusual depreciation rates, missing dividends, large round-number journals), summarise the year-on-year movements. Currently a junior or semi-senior spends 60-180 minutes per client doing this manually before the senior reviews.

A workflow that pulls Xero or QuickBooks data automatically, populates the working-paper template, runs an LLM analysis pass to flag anomalies, and produces a draft set of accounts plus a queries-for-client list cuts that prep work to a 10-15 minute review.

Time saved: 60-120 minutes per client. Multiply by every year-end the practice handles.

Build complexity: medium. Xero and QuickBooks have solid APIs. The working-paper template logic is the work.

3. Missing-records and bank-reconciliation chase

The chronic admin pain in every practice. A client hasn't uploaded December receipts. Another has ten unreconciled bank items from October. Currently a bookkeeper or junior emails the client manually, chases by phone, escalates to the partner. A workflow that watches the bookkeeping software for missing data, identifies the gap (specific date range, specific category), drafts a polite client email, sends, tracks response, escalates by rule, and updates the PMS with status replaces all of that with one weekly exception report for the partner.

Time saved: 3-6 hours per week of bookkeeper or junior time across a typical small practice.

Build complexity: low to medium. Xero/QuickBooks API plus email plus a simple state machine.

4. Client email triage and query routing

Inbound email is unfiltered noise: simple one-line tax questions, document deliveries, payroll changes, client complaints, urgent HMRC enquiries, sales enquiries. Currently the practice partner or office manager triages manually. A workflow that classifies incoming email by topic and urgency, routes to the right team member, drafts a first response for simple recurring queries (where's my tax return, how do I pay HMRC, when is my return due), and escalates anything sensitive to a human, replaces 30-60 minutes a day of triage with a queue.

Time saved: 2-4 hours per week of partner or office-manager time.

Build complexity: low. Email API plus an LLM classifier plus a routing table.

5. Capacity and billable utilisation tracking

A working stack of small-practice PMS shows job status but rarely shows the partner what next week looks like in capacity terms, or whether each fee earner is hitting the target chargeable ratio. A workflow that reads from the PMS, models capacity by week and fee earner, surfaces over-allocation early, and drafts a weekly partners' meeting report makes the practice run differently. Less of a time-saver than the others; more of a decision-quality multiplier.

Time saved: indirect, but practices that run this typically recover 5-15% of fee revenue from previously-leaking time and rebalanced workloads.

Build complexity: medium. The data wrangling is straightforward; the modelling judgement is the work.

What I wouldn't automate first

Tax advice generation for client files. LLMs are improving on UK tax but still confidently wrong on specifics often enough that any client-facing advice output needs senior review. The review wipes the time savings. Use AI for research summaries and to draft internal memos, not to generate client-facing answers.

Final account sign-off. Statutory accounts, corporation tax computations, and personal tax returns need a qualified human signing off. A workflow that automates the prep is fine; a workflow that automates the sign-off is not.

MLR risk scoring. The supervisory bodies will ask. AML risk decisions should be mediated by a qualified person using an approved tool. Automate the document collection and the chase; do not automate the risk decision.

HMRC submissions without human review. Filing the wrong number is expensive and slow to undo. Automate the prep; gate the submit.

The typical stack in 2026

A working AI and automation stack in a small UK accountancy practice has four layers.

Layer 1: the practice management system. Karbon, AccountancyManager, IRIS, CCH, Sage Practice, BTCSoftware, TaxCalc Practice. The system of record for clients, jobs, deadlines, and time. Karbon and AccountancyManager have the most usable REST APIs. IRIS and CCH are larger and more closed; automation typically has to interop through scheduled exports or shared folder patterns. Sage Practice sits between.

Layer 2: the bookkeeping ledger software. Xero, QuickBooks Online, FreeAgent, Sage Business Cloud, or a self-hosted package for older clients. All have usable APIs (Xero and QuickBooks are best-in-class). Most automation pulls TB and journal data from this layer.

Layer 3: the data and compliance services. Companies House (free, indispensable), HMRC API (for authorisations and Self-Assessment status), your AML provider (Smartsearch, Veriphy, AMLCC, Credas, ID-Pal), DocuSign or SignRequest, and document processing tools (Dext, AutoEntry).

Layer 4: the LLM decision layer. Claude (Anthropic) or GPT (OpenAI) called over API from the orchestration layer. One or two LLM calls per workflow is typical. Use API tier, not consumer apps. Confirm your data is not used for training (it isn't, by default, on Anthropic and OpenAI API tiers).

The interesting question is what you use to wire layers 1-4 together. This is the orchestration layer, and it's where agency archetype matters.

Who builds this for you (and who pretends to)

I wrote a longer version of this framing in my AI automation agency buyer's guide. The short version, applied to accountancy specifically:

• The Zapier or n8n reseller will quote £800-£5,000 per workflow built in their Zapier account. Usable for the simplest jobs (intake form to PMS, missing-receipt email reminders) but breaks on anything involving AML provider integration, working-paper template logic, or fan-out across clients. Most accountancy-flavoured "automation specialists" on LinkedIn sit in this archetype.
• The AI-labelled consultancy will sell you a £12,000-£25,000 "AI in accounting" strategy engagement. The deck cites Karbon's State of AI report and ICAEW's automation guidance. The implementation is either out of scope or subcontracted.
• The engineer-led custom shop will quote £12,500-£30,000 to build a narrow automation layer that reads and writes to your PMS and bookkeeping software over their APIs, runs on your own AWS or Render account, and is documented well enough that a future engineer can maintain it. You own the code and the process.

Laires Labs sits in the third category. The build floor is £12,500 because smaller builds don't cover scoping, MLR-aware design, build, handoff, and the post-launch support a real practice automation needs. Smaller jobs are better served by the reseller tier or by PMS-bundled AI.

Real costs: SaaS+reseller vs custom build over 2 years

Same 9-staff UK practice, same desired automation (the first three workflows on the list above: onboarding, working-paper prep, missing-records chase). Here's what each path costs across two years.

SaaS-plus-reseller path

• Karbon (9 users): £69/seat/month × 9 × 24 = £14,904
• Karbon AI add-on: £20/seat/month × 9 × 24 = £4,320
• Dext Premium (3 users): £40/seat/month × 3 × 24 = £2,880
• AML provider (Smartsearch): £80/month × 24 = £1,920
• Ignition for engagement letters: £75/month × 24 = £1,800
• Zapier Professional + tasks: £150/month × 24 = £3,600
• LLM API costs (moderate): £100/month × 24 = £2,400
• Setup by Zapier-flavoured accountancy automation specialist: £6,000 one-off
• Annual tweaks: £3,000/year × 2 = £6,000

2-year total: £43,824. Of which £31,824 is recurring seats and subscriptions, £6,000 is build, £6,000 is tweaks. Net incremental automation layer cost above what you'd pay for the PMS and bookkeeping baseline: roughly £12,000.

Engineer-led custom build path

• Karbon (9 users): £69/seat/month × 9 × 24 = £14,904 (same)
• Dext Premium (3 users): £40/seat/month × 3 × 24 = £2,880 (same)
• AML provider (API direct to Smartsearch): £80/month × 24 = £1,920 (same)
• Render or AWS hosting for the automation layer: £40/month × 24 = £960
• LLM API costs: £150/month × 24 = £3,600
• Custom build by engineer-led shop: £18,000 one-off
• Retained maintenance (ad-hoc): £3,000-£6,000 across 2 years

2-year total: £45,264-£48,264. The custom path is roughly £1,500-£4,500 more expensive over 2 years on a 9-staff practice.

The numbers are close because the PMS and bookkeeping seats dominate both paths. The real argument for going custom shows in two places. First, control: a custom layer does what you decide, in the order you decide, integrating with whichever AML or document tool you choose. Second, scaling: at 18 staff a SaaS path doubles its seat fees and Zapier task volume. The custom layer's marginal cost of a tenth fee earner is roughly zero.

For practices under six staff, the SaaS+reseller path is usually right. For practices over twelve, the custom path starts to win on year two and accelerates from there.

ICAEW, ACCA, AML, and the compliance questions you'll actually face

Every automation in a UK practice has to survive three questions. A good build answers all three in writing before the partner signs off.

Where does the data live and who can see it? Client-confidential information should not be sent to a US consumer AI service. On the API tier, Anthropic and OpenAI default to not training on your data, but the contract details differ. Your data protection record needs to reflect the actual data flow: what gets sent to the LLM provider, where the response is stored, retention policy, deletion path. UK or EU regions where possible. ICAEW's technology guidance is explicit on this.

Who is accountable for the output? The supervisory rules are clear: a qualified person signs off anything reaching a client or HMRC. An automation can draft, summarise, prep, and reconcile, but it cannot send work product to a client unsupervised. Build the human review gate in. The workflow that skips the gate is the workflow that ends up in a complaint to your professional body.

Have you mapped MLR risk into the workflow? Anything client-onboarding adjacent has to respect the Money Laundering Regulations. The automation can collect documents, prompt for the right information, draft the engagement letter, and route to your AML provider. It cannot decide that a client is low-risk and skip the check. Map the human-decision points before you build.

Book a call

Twenty minutes, no pitch. Tell me your PMS, your staff size, and the two workflows your fee earners complain about most. I'll come back with a scoped build plan: what I'd build, how I'd phase it, what it would cost, and which of the workflows above I'd park for your specific practice. You keep the plan whether you hire me or not.

Don't book if you're under three staff (the maths doesn't work), if you want a "digital transformation" strategy deck, or if your managing partner isn't bought in. The cultural gap is the biggest blocker and the one I can't fix.

Book a 20-minute call or read the full AI automation agency buyer's guide if you're still deciding between archetypes.