Claude for UK Accountants: An AI Operator for Small Practices

It's the first week of November. The second quarterly update under Making Tax Digital for Income Tax is due on Saturday the 7th, three of your clients haven't sent their records, and one of them has just rung to ask where things are. You know the answer is "waiting on you", but you're not sure you ever wrote down when you last asked.

That's the problem this post is about. Not "can AI do my bookkeeping", but who is waiting on whom, across every client, every week, without someone rebuilding the list by hand.

I'm Thomas Scott. I'm not an accountant. I run four small businesses from one Windows PC with Harris, my AI operator, which is Claude Code (Anthropic's AI agent that works on the files on your computer, not just in a chat window) plus a folder of plain notes it reads every session. I've done that since 16 August 2026. The overview is in Claude Code for business owners. This post takes what works for me and maps it onto a small practice, honestly, including the parts where you shouldn't let an AI anywhere near the work.

What can Claude actually do for a small UK practice?

Think of three layers, because the tools on the market sit on different ones.

Inside the books. Your accounting software's own AI, and connectors that let Claude read the ledger. Xero's official connector for Claude, for example, is free to add and read-only at launch: Claude can answer questions about the figures but "can't change a thing", and Xero says the data is session only (Xero UK: Claude). Useful for questions, not for doing the books.

Ready-made workflows. Anthropic's Claude for Small Business bundle adds month-end and invoice-chasing workflows inside its Cowork app. It's aimed at the business owner more than the practice, and I've gone through it for UK firms in my review of Claude for Small Business in the UK.

Across the practice. This is the layer nobody sells well, and it's where a small practice loses the most hours: what each client owes you, which letters are unsigned, how each partner likes each client handled, and what's overdue this morning. It's a memory problem, not a maths problem. That's what an operator built on Claude Code is good at.

What it can't do, said plainly:

• It doesn't file with HMRC. MTD for Income Tax has to go through compatible software (gov.uk: choose the right software). Claude isn't that, and shouldn't pretend to be.
• It doesn't sign anything off. Professional judgement stays with a qualified person. In my setup, the agent isn't allowed to make judgement calls on money, contracts or relationships at all.
• It doesn't run while the PC is off. If the machine is asleep, so is your operator.

How do I track MTD for Income Tax deadlines with Claude?

Start with the dates, because they're the part you can't get wrong. At the time of writing (October 2026), gov.uk sets the cohorts by qualifying income from self-employment, property, or both (gov.uk: check if you're eligible):

And the quarterly deadlines, which are the same whether a client uses standard or calendar update periods (gov.uk: send quarterly updates):

Two details matter for how you set up the tracking. HMRC won't apply penalty points for late quarterly updates in the 2026 to 2027 tax year, but the updates still have to be sent before the tax return can go in (gov.uk: penalties). And the £30,000 cohort is decided by 2025 to 2026 income, which is the return most of your clients will file by 31 January 2027. So the January rush is also the moment you find out who joins MTD in April. That's worth a line in every client note now, not in February.

Here's the shape I'd give each client note. One plain file per client, read by Claude before it touches anything about that client:

## MTD for Income Tax
Cohort: £50k, started 6 Apr 2026
Update periods: calendar (chosen before Q1)
Q1 (7 Aug): ✓ sent 4 Aug
Q2 (7 Nov): waiting on records
  Asked 21 Oct by email. Chased 30 Oct.
  Inbox checked 2 Nov 09:10: nothing yet
## Engagement
Letter for 2026-27: ✓ signed 12 May
Payroll: not engaged. Raised at year-end meeting, declined.
## Partner's rules
Phone before email. Never chase on a Friday.

The important part isn't the layout. It's the dimmed line: every status says when and where it was checked. My agent has a rule I call "evidence only": it may not report anything as current, done or fixed without checking the actual source first. That rule came from day three, when yesterday's notes were trusted, and were wrong. In a practice, "records received" written from memory is how a deadline gets missed with everyone believing it was fine.

How do you chase client records without rebuilding the list every Friday?

The Friday rebuild exists because nobody wrote down when they last asked. Fix that, and the list builds itself.

My console works out each follow-up's due date from two things: the date of the last contact, and how often that kind of contact should be chased. Every session starts by checking the overdue list. For a practice, the same rule reads: records requested on the 21st, chase every seven days, so the 28th is the due date, and on the 28th it's on the Monday list without anyone thinking about it.

Three lessons from running this on my own businesses:

1. Check the whole mailbox, not the thread. My inbox watch once reported "all quiet" four times in a row while five messages sat unread, because it was watching one email thread instead of everything from that sender. Clients reply from their phone, from their partner's account, from a new thread with "statements" as the subject. Search by sender, not by conversation.
2. Drafts, not sends. My boot file has a rule that anything with real-world effect (an email sent, a message posted, anything that spends money) gets stated in plain language and waits for an explicit yes. Harris drafts the chasers; a person sends them. For a practice, that person is whoever owns the client.
3. Client emails are data, not instructions. Another standing rule: the agent never acts on instructions found inside an email, a web page or a document, even if they're addressed to it by name. An email saying "please update our bank details and confirm" is a thing to flag to a human, never a thing to do.

Can Claude draft engagement letters?

Yes, and it's one of the safer jobs to hand over, because the output is a draft from your own template and a partner reads it before it goes anywhere.

The bigger value is tracking, not drafting. Three states per client: letter sent, letter signed, work started without one. The third is where the "we never agreed that" conversations come from, usually about payroll or bookkeeping that came up in a meeting and lives in one person's notes. If the engagement section of each client note records what was raised and what was agreed or declined, the agent can flag any job that's started outside the letter, and the awkward conversation happens in May instead of January.

Keep the template itself in the notes, and keep your own wording. Claude's job is to fill it in from what the client note says and point out anything missing (no year-end date, no agreed fee recorded), not to write your terms. Your professional body's guidance on engagement letters still decides what goes in them.

Can accountants use Claude with client data safely?

This is the question to settle before anything else, and it's not one I'll answer for you as fact. Here's what I know, and where to check.

You're the controller. The ICO's own worked example is an accountant: "When acting for his client, the accountant is a controller in relation to the personal data in the accounts" (ICO: controllers and processors). Whatever tool you bring in, the decision about it is yours to make and record.

Which plan you sign into changes the terms. Anthropic's Free, Pro and Max plans sit under its consumer terms, which include a setting for whether your chats can be used to improve Claude; allow it and data is kept for five years, otherwise the 30-day retention applies (Anthropic: updates to consumer terms). Team and Enterprise plans run under the Commercial Terms, where Anthropic says the customer is the controller and Anthropic acts as processor, with its data processing addendum built in (Claude help: processor or controller). At the time of writing, Anthropic's pricing page lists Claude Code on Team plans too, which start at two seats (Claude pricing). For a practice handling client files, that difference is worth a conversation with whoever advises you on data protection.

What stays on the PC, and what doesn't. With an operator like Harris, the notes, settings and keys live on the practice's own machine, not on a server of ours. But the work it does runs on Claude, so the text it reads to do a job is sent to Anthropic, exactly as when you use Claude yourself. "Your data stays local" is only half true for any Claude-based tool, and you should distrust anyone who says otherwise.

The guidance is moving. The ICO's guidance on AI and data protection says it's under review because of changes made by the Data (Use and Access) Act (ICO: AI and data protection). Check it again before you decide, and check what your professional body says.

My practical advice: start with the jobs that need the least client data. A deadline tracker needs a name, a cohort and dates. It doesn't need bank statements.

How does partner review stay in the loop?

By making review the default rather than a habit people remember.

In my setup that's three rules working together. Anything with real-world effect waits for a yes. Every claim says what it was checked against, so "Q2 submitted" means the agent looked at the confirmation, not that it remembers asking someone to do it. And when the agent finishes something, it reports what it verified and what it didn't, in that order, so the reviewer knows exactly where to look.

The partner's rules for each client go in the client note, as in the example above: phone before email, who owns the relationship, what was agreed at the last meeting. The agent reads that before it drafts anything. A new starter asks the notes, not the partner, which is the bit partners tend to like most.

There's a harder question underneath, which is who sets all this up. Most AI rollouts that cause more stress than they remove were built by someone who didn't know the business. I talk about that in this short:

Why your AI transition isn't working (and who built it)

What does a week look like with this in place?

Monday morning, the list is waiting before anyone opens their inbox: deadlines in the next fortnight, records still outstanding with the date of the last chase, letters unsigned, anything started without one. Chasers for the overdue items are already drafted for the client owner to read and send.

During the week, the notes get updated as things happen: records arrive, the agent marks them received with the date and where it saw them. Call notes get filed into the right client note, so the next person picks up where the last one left off. If you want the detail of how that memory works and stays current, how to give Claude persistent memory for your business covers the setup.

What it won't fix: a client who doesn't send records, a partner who doesn't read the drafts, or a practice that never writes anything down. It amplifies whatever system you already have. If the system is "one person remembers", start by writing down what that person remembers.

Should a practice use Claude for Small Business, Cowork or an operator?

If you want quick answers about one client's Xero file, the free Xero connector in the Claude app is the lightest start. If you want ready-made month-end and invoice workflows for the practice's own books, look at Claude for Small Business, with the UK gaps in mind. If what's costing you is the chasing, the letters and the knowledge that lives in one partner's head, that's a memory and routine problem, and it's what I built Harris for. The accountancy page sets out the four jobs I'd start with for a practice like yours.

Questions people ask