An AI agent can take most of the writing out of an accounting practice: chase emails, variance commentary, missing-information lists, plain-English replies about HMRC letters. You paste in the figures or the client's message, it drafts, and you check and send. It doesn't post journals, file returns or make tax decisions.
What makes it more than a chat window is memory and routine. It learns that one client sends photographs of receipts and another a tidy export, it knows how your firm writes, and it can message you on the first working day of the month with the deadlines nobody has started.
What does an agent do that your accounts software doesn't?
Your ledger holds the records, runs the reports and files the returns. The agent works on text you choose to give it: a pasted bank extract, a trial balance, an email thread, the wording of an HMRC letter. It isn't connected to Xero, your practice management system or your inbox, and for a profession with confidentiality duties that's a sensible default. It only ever sees what you deliberately paste.
The flip side is that quality depends on the extract. Give it the whole set of statement lines, the prior-year comparatives and the client's sector, and the draft comes back close. Give it a vague summary and you'll get vague commentary.
The tasks below are grouped by the practice calendar, starting where the work starts.
Taking on a new client
The first email from a prospective client usually holds the scope, the deadline and three things they haven't thought about. Paste it in and ask for an onboarding checklist: what you know, what to ask before quoting, what to send them.
This is an email from a prospective client, a limited company with two cafés. Turn it into an onboarding checklist and list the questions I need answered before I quote.
Clients moving from another firm bring a chart of accounts you'd never have built. Paste theirs and yours as two lists and ask for a mapping, with anything lacking an obvious home flagged and the uncertain matches marked as such. You still make the calls, but you start from a table rather than a blank sheet.
The scope section of the engagement letter is where misunderstandings begin, so ask for that paragraph in plain English with the exclusions spelled out and the specifics left as placeholders. Pair it with a document request list that fits how this client keeps records (the spreadsheet-and-receipt-photos client needs a different list from the one with a bookkeeper).
Month-end: reconciliations and management accounts
Categorising a statement. Paste forty bank lines as text and ask for a suggested nominal on each, with a confidence column and anything that looks personal rather than business picked out. Nothing gets posted. You get a starting point and a list of guesses to check.
Finding duplicates. A second pair of eyes on a supplier ledger extract catches what you'd miss at six on a Friday.
Here is a quarter's supplier ledger extract. Find possible duplicate payments: same supplier, similar amount, within 14 days. Show each pair and your reasoning.
Explaining a difference. Finding the cause yourself is the easy part; explaining it to a client is the work.
Our bank reconciliation is out by 412.30. Below are the statement lines and the ledger lines. List the likely causes in order, and what to check first for each.
You get ranked causes, each paired with the check that would confirm or rule it out.
Variance commentary. For most practices this is the single most valuable job on the page. The figures for a café client's management accounts might be done by half ten, and the commentary can still eat the rest of the morning, because explaining usefully why turnover rose while gross margin fell takes longer than the numbers did.
Here are this month's figures and the same month last year, by line. Write client commentary: what moved, by how much, the likely reason. Six paragraphs. Mark anything the figures alone can't explain.
The movements come back ranked by size, with the ones the numbers can't account for flagged for you to ask about. Check every figure it quotes against your source, and treat its stated reasons as hypotheses. A fluent paragraph about margin makes you more inclined to believe it, which is exactly why you test it.
Chasing receipts. Fourteen expenses with no receipt become one short message listing them by date and amount, explaining why you need them and giving a clear date, plus a table for the file.
Year end and tax season
January is when this helps most. A self assessment client with rental income, dividends and a share disposal has sent you some of what you need across four messages and a phone call. Paste it all in and ask what a competent preparer would still need before starting the return, ordered by what blocks progress first.
HMRC letters arrive as panicked forwards. Paste the text of the letter and ask what HMRC wants, by when, and what happens if the date is missed. Ask for the dates pulled out separately, then a short reply to the client in plain words. What used to take forty minutes of reading twice and rewriting takes five.
Where you've already decided a position, the agent can explain it. Say you're treating a roof job as a repair rather than an improvement:
I'm treating this as a repair, not an improvement, for the reasons below. Draft a paragraph for the client explaining the position. Don't add any technical argument I haven't given you.
That last instruction matters. It keeps the reasoning yours and stops the draft inventing support you'd then have to defend.
If someone else signs off, ask for a short note to the reviewer: your working, the two points you're unsure about, and exactly what you need confirmed. Reviews go faster when the question is stated up front.
Chasing clients without souring the relationship
Records that haven't arrived three weeks before a deadline call for a sequence: a nudge now, a firmer message in ten days, a final one a week out. Writing all three in one go, calmly, means you're not drafting the third one while irritated.
The covering note for a set of final accounts is the part the client reads. Ask for the three things they should notice, the one decision they need to make and what happens next, in under 200 words.
The hardest of these is the fee that went up. Paste both years' fees and what was different about the work, and ask for a reply that sets the difference out factually without sounding defensive. Then make it sound like you before it goes.
Reminders that follow the practice calendar
Scheduled messages are where the agent starts coming to you. You ask in plain language and adjust the wording until the output is something you read. Scheduled check-ins explains the mechanics.
Every Monday at 8am, ask me to paste the week's deadlines and job list, then tell me what must finish this week, what's waiting on a client and what I should start now.
Every Wednesday at 4pm, remind me to paste in what clients still owe us, and draft chase messages. Keep them short and vary the wording.
On the first working day of each month, ask for the client deadline list and tell me which deadlines in the next 60 days have no work started.
Every quarter, show me what you remember about our clients and house style, and ask what's changed.
The quarterly one is easy to skip and shouldn't be. An agent working from a stale note about a client writes with the same confidence as one working from a fresh note.
Turning good drafts into house procedures
Once a task has gone well twice, ask the agent to write down the method and save it, so next time it starts from your standard. Agent skills explained covers how that works.
The variance commentary format is the obvious first one: which movements get a sentence, which get a table, how much the client wants explained. A house style for client letters comes next. Build it from three letters you're happy with, ask the agent to state the rules it can see, and correct them. The year-end file structure (order of work, standard queries, checks before review) helps a new junior get productive sooner.
Look back at saved procedures every so often. An agent will repeat a mistake you saved as faithfully as a good method.
Which judgements stay with the accountant?
Plenty of hard technical work is fine to hand over, because you'll check it anyway. The line is accountability.
- Tax positions. What to claim, how to treat a transaction, whether an arrangement works. The agent can explain a decision you've made; it shouldn't be where the decision comes from.
- What goes to HMRC as the firm's view. Correspondence can be drafted, but the position stated needs a qualified person behind it.
- Sign-off and materiality. What matters in a set of accounts is a judgement about this client in this year, and once a draft becomes the accounts, someone has reviewed it.
- Anything you couldn't explain. If a reviewer asks how a figure was reached, "the agent produced it" won't do.
The seven bodies behind Professional Conduct in Relation to Taxation, including ICAEW, published AI guidance in January 2026. It adds no new rules. It reads the existing five fundamental principles (integrity, objectivity, professional competence and due care, confidentiality, professional behaviour) through an AI lens, and its position is that AI can support the work but doesn't replace professional judgement, with responsibility staying with you. Read ICAEW's PCRT pages in full rather than relying on a summary like this one.
Is it safe to paste client data into an agent?
Only once your firm has decided it is, and only the minimum the task needs.
A variance commentary needs figures and a sector. It doesn't need the client's name, the director's circumstances or the address. Use "Client A" and keep the mapping in your own records. Watch for combinations, too: a café in a named town with two sites and a stated turnover identifies a client as well as a name does, so drop the geography and round figures where the exact number doesn't matter.
Keep personal data out of memory. An agent that remembers how a client sends records is useful; one that remembers a client's health or family situation is a liability you created. Ask it to forget anything of that kind.
Any hosted service processes text on a server somewhere. Under the UK GDPR, GDPR or your local data-protection law, putting client personal data into a third-party tool is a processing decision the firm should make deliberately, with your own policy applied as written: if client data doesn't go into outside tools, that covers an agent. What happens to your data lists the questions to put to any provider. Your professional body and compliance lead are the right people to settle what your obligations require.
Getting started: what we set up, then your first week
We set Able° up for you, connect it to Telegram or LINE with you, and go through what it should know at the start: what the practice does, the kind of clients you serve, how you write and what never belongs in a draft. You read back its summary and correct it. That conversation does more for the quality of later drafts than anything else.
Then your week, about twenty minutes a day:
- Draft one routine email twice. A chase or a covering note. Edit the first, tell it what you changed and why, and compare the second.
- Run one analysis task. Paste real figures and ask for variance commentary, then check every number against the source. You'll learn where it's reliable and where it's fluent and wrong.
- Set two schedules. The Monday plan and the month-start deadline check. Live with them before adding more.
- Save your first procedure. Take whichever task worked best, ask for the method to be saved with your corrections, and try it on a different client.
Practices that won't get much from it
If most of your day is posting transactions and preparing returns inside software that already does the heavy lifting, the drafting share is small. The same goes when the repetitive work is locked in a system and extracting the text takes longer than writing the email. A sole practitioner with four clients holds everything in their head, so the memory advantage disappears. And a firm that would treat a draft as finished work is worse off with an agent than without one.
Common questions
Can an AI agent do my bookkeeping?
No. It can't connect to a bank feed, post a journal or reconcile an account. It can read a pasted statement and suggest codes, spot likely duplicates and talk through a difference, but posting stays in your software and under your control.
Will it get the tax treatment right?
Sometimes, and you can't tell which times without checking. Rules change, and it never has a client's full circumstances. Use it to explain positions you've already reached, not to reach them.
Does it work with Xero, QuickBooks or Sage?
It doesn't connect to any of them. You export or copy the report you need (a trial balance, an aged debtors list, a ledger extract) and paste it into the chat. That's slower than an integration but means the agent sees only what you chose to share.
How do I stop it inventing figures?
Tell it to mark anything it's guessing and to add nothing you haven't supplied, then check the flagged items first, because that's where invention tends to cluster. Ask it to quote figures exactly as they appear in what you pasted, so a mismatch is easy to spot.