Day to day, small accounting firms use AI in five places: summarising and drafting client emails, pulling data from receipts and bank statements, chasing missing records, turning meeting recordings into file notes, and drafting management-account commentary and plain-English explanations. A qualified person checks each output, and advice, sign-off and client decisions stay human.
Most of it doesn't happen in ChatGPT. It runs inside software practices already use for other reasons: the practice-management system, the ledger, the receipt-capture app. A chat assistant fills the gaps. The week below follows one composite practice, built from the way small firms commonly work rather than from any single client, using real products and their real features. The prompts, outputs and numbers are illustrative.
The practice in this example, before AI
Five people: the owner (a qualified accountant), a manager, two bookkeepers and a practice coordinator. Around 150 clients, mostly sole traders and small companies, with a cluster in hospitality: a boutique hotel, a tour operator, a letting agency, a campsite and about twenty holiday-let owners. Most clients are on Xero, a handful on QuickBooks. The practice runs on Microsoft 365, uses Karbon for workflow and email, and Dext for receipts.
Before switching anything on, the owner timed a normal week. Email ran to about 14 hours across the team, data entry and reconciliation about 16, chasing clients 5, meeting notes 3, and writing commentary and explanations about 6. That baseline is what the numbers at the end are measured against.
Monday: the inbox after the weekend
The coordinator starts with Karbon's triage inbox. Karbon's AI features, which are included in its paid plans at no extra cost, can summarise long email threads, suggest quick replies and draft or refine an email from a short prompt. The useful one on a Monday is the thread summary. The letting agency's director has sent five emails over the weekend about a change of year-end, copying in their office manager, who replied twice.
Thread summary (illustrative):
- Director wants to move year-end from 31 March to 31 December.
- Office manager asks whether this affects the October payroll run.
- Director asks for a call this week and a fee for the extra work.
- No decision made yet; director unsure about timing.That takes ten seconds to read instead of three minutes. The coordinator books the call and assigns the thread to the manager. What she doesn't do is let the suggested quick reply go out, because it said "Yes, we can certainly change your year-end", which is a decision for the owner, not a pleasantry.
Across a morning, the pattern is the same: AI summaries for anything longer than three messages, AI drafts for routine replies (portal logins, copy invoices, appointment times), and a person writing anything that contains advice, a figure or a fee. More on setting up this kind of sorting is in AI email triage for professional firms.
Tuesday: receipts, bank statements and a date that flipped
Bookkeeping day. Dext reads receipts and supplier invoices and extracts the supplier, date, amounts and tax, then publishes them to Xero or QuickBooks. For the holiday-let owners, most of whom photograph receipts on their phones, this replaced a lot of typing. Xero's own document capture now pulls transactions out of uploaded bank statements, which helps where a client's bank feed is missing.
The bookkeepers' job has moved from typing to checking. Their rule: anything the software flags as uncertain, anything over a set value, and a sample of the rest.
Written down for the holiday-let clients, the rule reads: check every item Dext marks as low confidence, every item over $250, and five random items per client per month. The sum shows why that's manageable. Twenty holiday-let owners send about 35 documents each a month, so 700 in total. Roughly 60 get flagged or exceed the threshold, and the random sample adds 100, so the bookkeepers look closely at about 160 documents instead of typing 700. At a minute or so each, that's under three hours a month for the whole group. If the sample starts finding more than one error in twenty, the sample size for that client doubles until the cause is found.
This Tuesday the sample catches a realistic mistake. A cleaning-supplies invoice for a holiday-let owner, issued by a supplier abroad, shows the date 03/04/2026. The extraction read it as 4 March, not 3 April, which put it in the wrong quarter. Nothing in the software looked wrong; the bookkeeper noticed only because the owner hadn't bought anything in March. The fix was a supplier-level date rule and a note on the client file. The wider set-up, including how categorisation rules learn, is in AI receipt capture and bank categorisation.
On the ledger side, bank-feed suggestions do the matching and the bookkeeper accepts or corrects. For the tour operator, where one customer payment often covers three bookings, suggestions are wrong often enough that the bookkeeper matches those by hand. Xero has announced better handling of split payments as part of its JAX roadmap, but a practice should plan around what the software does today.
Wednesday: chasing what's missing
The coordinator exports the list of unreconciled transactions older than 30 days that have no document attached, anonymised to client codes, and asks ChatGPT Business (the practice has two seats) to group them into a chaser per client.
Prompt:
For each client code below, write a short, friendly email asking
for the missing documents. List each transaction with date, amount
and payee. Maximum 80 words of text plus the list. Don't mention
deadlines or penalties. Sign off as "The team at [practice]".
[pasted table]
Sample output for one client (illustrative):
"Hi [first name], we're just missing a few receipts to finish your
August and September books. Could you upload these when you get a
moment?
12/08 Fuel station 64.20
19/08 Hardware store 212.00
02/09 Online marketplace 38.99
Thanks so much, and hope the season has gone well!
The team at [practice]"Before anything is sent, the coordinator does three checks: the client isn't marked "do not chase" in Karbon, the documents haven't landed in Dext since the export, and the placeholders are filled. This week two of fourteen drafts are deleted: one client uploaded everything on Tuesday night, and one is on the do-not-chase list after a family bereavement. The ending "hope the season has gone well" is fine for the campsite owner and odd for the letting agency, so she edits it out of theirs.
Replies bring the next job. The campsite owner answers: "The 212.00 at the hardware store was cash for fence posts, there's no receipt, sorry." The coordinator doesn't ask the assistant what to do with that. She passes it to the bookkeeper, who records it with the client's explanation as the supporting note, and adds the reply to the client file so the same transaction isn't chased again next month. Where a client says "I'll look" and nothing arrives, the second chaser is drafted the same way a fortnight later, with one change the coordinator makes by hand: it names the one or two items that actually matter for the return, rather than the whole list again.
Xero announced in August 2026 that JAX will be able to spot missing documentation, email the client, send reminders and ask follow-up questions, rolling out over time. When that arrives, the same three checks still apply. The full routine is in how to chase missing client records with AI.
Thursday: a planning meeting and a file note that nearly misled
The owner meets the boutique hotel's two directors on Google Meet to talk about whether to put the property into a company. The practice's Microsoft 365 set-up doesn't cover Meet, but the directors use Google Workspace and have "Take notes for me" switched on, which is available on Business Standard and above. The owner asks at the start whether everyone is happy for notes to be taken, and says the practice will write its own file note from them. (Workspace admins can now require explicit consent from each participant before note-taking starts; the setting is off by default.)
The AI notes land in the directors' Drive and they share them afterwards. One line reads:
"Decision: the directors agreed to incorporate from 1 April and
asked the accountant to prepare the paperwork."What was actually said: the directors liked the idea, wanted figures for both options, and would decide after speaking to their bank. The owner's file note corrects it plainly, and the email to the directors says what was agreed in their own words. If the practice had filed the AI note as its record, it would have been evidence of a decision nobody made. Every meeting summary gets read against memory the same day, while the conversation is still fresh, and only the corrected version is allowed to generate tasks.
Friday: commentary and plain-English explanations
The manager prepares the quarterly management pack for the tour operator. The numbers come from Xero; the commentary used to take her an hour per client. Now she writes three lines of notes from her conversation with the owner, pastes the summary figures and the notes into ChatGPT Business, and asks for commentary that uses only the reasons in the notes.
Her notes were: deposits for next spring up sharply after an early-booking offer; coach hire costs up after a supplier changed; one guide left in July and wasn't replaced. The draft comes back well structured. She fixes two things: it described the deposit increase as "revenue growth", when deposits are money held for trips not yet run, and it rounded a cost increase of 8.6% to "almost 10%". Both are the kind of slip a reviewer catches in thirty seconds and a client might not catch at all.
The owner uses the same approach for a sole trader who has asked why he owes tax when his bank account is nearly empty. She gives the assistant the figures and the reason (drawings and a van purchase), asks for 150 words at a reading age of about twelve, and edits the result. An extract from the illustrative draft:
"Your tax is worked out on your profit, not on what's left in your
bank account. This year your business made a profit of [figure]. You
took money out for yourself during the year, and you bought a van,
which is why the account is low. The van will reduce your tax, but
only part of its cost counts this year..."The first two sentences are exactly right and she keeps them. The van line she rewrites, because how much of a vehicle's cost counts in the first year depends on the rules that apply to him and the choices made in his accounts; the draft stated a general rule as if it were his. She also adds what the draft couldn't know: the date the tax is due, and an offer to work out a monthly amount he can put aside so next year isn't a surprise. The method, including prompts for common questions, is in how accountants use AI to explain tax in plain English.
What the week added up to
After three months of running these routines, the owner timed another normal week. Illustrative figures:
| Work | Hours before | Hours after | Where the change came from |
|---|---|---|---|
| 14 | 10 | Thread summaries, drafted routine replies | |
| Data entry and reconciliation | 16 | 10 | Receipt extraction, bank-statement extraction, feed suggestions |
| Chasing clients | 5 | 2 | Drafted chasers, grouped by client |
| Meeting notes | 3 | 2 | AI notes as a first draft for the file note |
| Commentary and explanations | 6 | 4 | Drafts from notes and figures |
| Total | 44 | 28 | About 16 hours a week, across five people |
The review work is inside the "after" column: checking a receipt sample, reading every summary, fixing drafts. Practices that leave checking out of their sums overstate the saving, then feel misled.
New spending was modest: two ChatGPT Business seats at $20 each a month on annual billing, so $40 a month. The Karbon AI features, receipt extraction and ledger AI came inside subscriptions the practice held anyway, and Microsoft 365 business plans include Copilot Chat at no extra cost. The larger cost was time: roughly 30 hours of the owner's and manager's time over three months to set up, test and write the review rules. For a costed view of the options, see the best AI tools for small accounting firms by task.
What stayed human, on purpose
- Signing off accounts and returns. AI prepares; a qualified person reviews and signs.
- Advice. Incorporation, pricing, tax planning and anything a client might act on is written or at least rewritten by the owner or manager.
- Fees and scope. The Monday quick reply that agreed to a year-end change is the reason this is on the list.
- Client identity and money-laundering checks. Identity documents don't go into a chat assistant, and the judgement on a client's risk isn't delegated.
- Bad news. A large tax bill, a cash shortfall or a missed deadline is delivered by a person, usually by phone.
Clients do ask. The hotel's directors, having seen their own meeting tool take notes, asked whether the practice "puts our accounts into ChatGPT". The owner now has two sentences she uses in the engagement letter and repeats on the phone:
We use AI features inside our accounting and practice software to read receipts, summarise emails and draft routine messages, and a member of our team checks everything before it is used or sent. Your identity documents never go into these tools, and advice about your affairs is always written by a qualified member of staff.
It's short because it has to be true in every detail. If a practice can't write that paragraph honestly about its own set-up, the gap is usually a personal chat account someone is using on the side.
What the first month taught the practice
Four lessons that most practices learn in roughly the same order:
- Samples beat trust. The date flip on Tuesday wasn't a one-off. Sampling twenty extracted documents per client each month caught a steady trickle of small errors that would otherwise have reached year-end.
- Summaries overstate agreement. AI meeting notes tend to turn "we'll think about it" into "agreed". Reading them against memory, the same day, is non-negotiable.
- The easy route wins. In the first fortnight one bookkeeper kept using her personal chat app because the practice workspace wasn't set up on her laptop. Once it was, the habit moved. Make the approved tool the quickest one to reach.
- Built-in features are uneven. Some worked well from day one, some needed settings changed, and some advertised features weren't available yet. Test each against a normal week before relying on it, and keep the practice's own checks regardless of what the software promises. The two ledgers compare differently on this; Xero vs QuickBooks AI sets out where each saves bookkeeping time.
The practice kept a plain error log from week one, one row per mistake the checks caught, and reviewed it every Friday. Four of the first month's rows, illustratively:
| Week | Routine | What went wrong | What changed |
|---|---|---|---|
| 1 | Receipts | The hotel's linen hire was categorised as cleaning, so a cost line in its management pack looked wrong | Supplier rule set to the correct account; the hotel's first month of linen invoices re-checked |
| 2 | Email summary | A thread summary for the campsite left out an attached letter about a payment plan | Summaries are never used for any thread with an attachment until the attachment is opened |
| 3 | Chasers | A letting-agency chaser listed a transaction the director had explained in a client note | Client notes checked before export, and explained items tagged so they drop off the list |
| 4 | Commentary | A draft called a one-off insurance refund "improved margins" | Prompt now says: flag one-off items separately and don't describe them as trends |
By the end of the second month the log was down to one or two rows a week, and every row had changed a rule, a setting or a prompt. That's the point of it: an error that's only corrected, and not written down, turns up again in a different client's file.
None of this made the practice a different kind of firm. It made the same firm about 16 hours a week lighter, with a checking habit that catches the errors AI brings with it.
Questions practices ask after seeing a week like this
Do we need a chat assistant at all if our practice software has AI built in?
Often not at first. Practice-management, ledger and receipt-capture tools now cover email summaries, data extraction and some chasing. A chat assistant earns its place for the gaps: drafting explanations, reworking commentary, screening exports and one-off analysis. Start with what your existing software offers, list the jobs it can't do, then decide whether two business seats are worth it.
How long did it take the practice in the example to get to this week?
In a composite like this, around two to three months: a few weeks switching on built-in features and fixing their settings, a month running each new routine alongside the old one, and a further few weeks writing prompts and review rules. Busy season is the wrong time to start; the quarter after a deadline peak is the right one.
Which of these uses carries the most risk?
Meeting notes and anything client-facing that contains a figure or a rule. Extraction errors get caught at reconciliation, but a summary that records the wrong decision or an email with a wrong deadline goes straight to the client. Those two get a named reviewer every time, while lower-risk tasks such as internal summaries can be spot-checked.
Further reads
- What an AI Implementation Looks Like in a Small Accounting Firm — How a practice gets from zero to a week like this one.
- How Much Does AI Cost a Small Accounting Firm? — What the tools in this week cost a small practice per year.
- How Accountants Use AI to Spot Errors in Client Books — Another daily use: AI as a second pair of eyes on the ledger.
- How to Turn Meeting Notes Into Tasks Automatically With AI — Turn the Thursday meeting notes into tasks automatically.
- Turning Hours Saved by AI Into Advisory Work Clients Pay For — What to do with the hours a week like this frees up.
- How Small Tax Practices Use AI Through the Busy Season — How the same tools change during the busiest months.
- How to Automate Client Onboarding in a Small Accounting Practice — Six stages to take a small accounting practice from signed proposal to first job with fewer manual touches, and where AI helps or must not decide.
- How to Choose an Accountant Who Uses AI Well — What good AI use inside an accountancy practice looks like, eight questions for the first meeting, and how a café compared three firms.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Karbon AI feature page (features and availability); Dext product pages; Accounting Today report on Xero JAX announcements (24 August 2026); QuickBooks Intuit AI overview; Google Workspace Updates on Meet consent controls (April 2026); OpenAI ChatGPT Business pricing.