How Can an Accountant Utilise AI? Month-End Tasks to Hand Over First

An accountant can use AI at month-end by giving it the writing jobs first: reconciliation commentary, variance notes and client chasers. Strip names, account numbers and anything identifying before pasting, use a business plan that doesn't train on your content, and have a person confirm every figure and reason before anything leaves the practice.

The order matters more than the tool. The usual confidentiality slip is real client data pasted into a personal account on a busy Thursday, not anything the AI does on its own. So the practice sets the rule before the first prompt: what may go in, which account it goes into, and who signs off the output.

Follow me on Instagram@sagnikteaches

What counts as confidential in a month-end file

Client confidentiality is wider than the obvious items. A trial balance with the client's name removed can still identify them if it has a distinctive supplier, a payroll total that matches a known headcount, or a one-off sale that the owner would recognise at once. Before deciding what AI may see, sort the month-end material into three groups.

Connect on LinkedInSagnik Bhattacharya

The masking step sounds fiddly, but it is a two-minute job once you have a habit. Export the table, replace the client name in the file name and the header with a code, and hold the key (code to client) in your practice system, not in the chat. If you use the same client every month, paste a list of the codes at the top of each fresh chat so the AI keeps using them, with the real names left out of it too.

Subscribe on YouTube@sagnikteaches

Pick the account before you pick the task

Where the data goes matters more than which assistant you favour. As of November 2026, the business versions of the big assistants say they don't train on your content by default, while the personal plans rely on a settings switch that each user must find and set. That difference decides which account you use for amber material.

• ChatGPT Business (formerly Team): $20 per user per month billed annually, $25 monthly, with a two-seat minimum. Business data isn't used for training by default.
• Claude Team: $20 or $25 per seat in the same pattern, from two seats up to 150. Business content isn't used for training by default.
• Microsoft 365 Copilot Business: an add-on to a Microsoft 365 business plan at $21 per user per month on annual billing ($25.20 monthly), for organisations up to 300 users, with a first-year promotional price of $18 on annual terms until 31 December 2026. It works well if the practice already lives in Outlook, Excel and Teams.
• Personal plans (the free and individual paid tiers): fine for green material only. Both vendors' pricing pages list model training on these plans as something the user opts out of, and you shouldn't rely on every user remembering the toggle.

A two-seat minimum means a one-person practice pays about $40-$50 a month for a business plan. Against the cost of a confidentiality complaint, that is cheap, but check the figures on the vendors' own pages before you commit, because plans change often. If your bookkeeping system has its own AI features, read what it says about data use too: QuickBooks, for instance, has no single switch for all its AI features, and Intuit's community team said in August 2026 that the Business Feed and Intuit Assist can't currently be turned off, so decide whether that suits your clients before you rely on it.

Then tell the client. The engagement letter is the natural home for a line saying that AI tools may help prepare working papers and messages under partner review, and that no client data is used to train models. There is more on the wording in AI engagement letters for accountants.

A six-person practice and its close

Here is an illustrative practice of six people looking after 42 small-business clients on monthly management accounts. One partner reviews, two seniors prepare, two bookkeepers keep the ledgers up to date and one administrator handles chasing. The month-end window is the first six working days. The numbers below are made up to show the shape of the saving, not a promise.

Across 42 clients, 45 minutes saved each comes to about 31 hours a month. At a blended internal cost of $45 an hour, that is roughly $1,400 of time, against perhaps $120 a month in seats for six people on a business plan. Treat the saving as a first estimate: the first two months run slower while the team learns the masking habit and the prompts, and the review step eats part of what you gain.

Job one: reconciliation commentary and the open-items list

The AI doesn't reconcile. The preparer does that in the bookkeeping system, where bank feeds and matching rules already work. What the AI can do is read the list of unmatched items afterwards and draft the commentary: what is outstanding, how old it is, and what you need from the client.

Export the unreconciled items as a table, mask the payee names where they identify a person, and paste them in with a prompt like this.

You are helping a bookkeeper at a small accounting practice. Below are the
unreconciled bank items for Client A at 31 October, with payee names masked.

Write:
1. A three-sentence summary of the reconciliation position.
2. A table of open items grouped as: needs receipt, needs explanation,
   probable timing difference.
3. A list of questions for the client, one line each, each naming the
   date and amount.

Rules: do not guess what a payment was for. If the description does not
say, put it under "needs explanation". Do not invent amounts or dates.

[paste table]

A typical illustrative reply opens with something like: "Eleven items remain unreconciled at 31 October, totalling $4,860 out and $1,320 in. Seven need a receipt, two need an explanation and two look like timing differences from cheques or transfers dated 30 and 31 October." Then the table and the question list follow. Your preparer's job is to check three things before this goes anywhere: the count matches the system, the totals add up, and the timing differences really did clear in the first days of November.

Common fix-ups: the AI will sometimes group a recurring subscription with a "needs explanation" payment because the description is cryptic. Add a line to the prompt listing the client's known regular payments (masked), and it stops doing so. It will also round amounts if you let it, so insist on exact figures.

Job two: variance notes you can defend

Variance commentary is where the risk of confident nonsense is highest. The AI sees that "Advertising" rose by $3,100 and wants to say why. If the data doesn't hold the reason, it will often supply a plausible one. The fix is to instruct it to separate what the numbers show from what they cannot show.

Illustrative input, for a client with a small shop and an online arm: October revenue $48,200 against $41,900 in September and a budget of $44,000; cost of sales $29,300 against $24,500; advertising $5,400 against $2,300 last month and a budget of $2,800. A good prompt asks for three things per line: the movement in dollars and per cent, what the ledger supports (for example, "three supplier invoices dated 3, 12 and 24 October totalling $3,050 are coded to Advertising"), and a question for the client where the cause is unclear.

For each line below, write one sentence of fact from the figures only
(movement in dollars and per cent versus last month and budget). Then
write "Evidence needed:" and name the ledger lines or documents that
would explain it. If the figures cannot explain it, write "Cause unknown,
ask client". Never state a reason as fact.

[paste masked P&L comparison]

Sample output for the advertising line, illustrative: "Advertising is $5,400, up $3,100 (135%) on September and $2,600 over budget. Evidence needed: the three supplier invoices coded this month and any campaign dates. Cause unknown, ask client whether a seasonal campaign started." Your senior then opens the ledger, finds that the largest of those invoices is a new agency retainer, and swaps the question for the fact. The result reads like a note from someone who has looked at the books, which it was.

The cost-of-sales line deserves a separate check. Revenue rose 15%, but cost of sales rose 20%, so the gross margin slipped from about 41.5% to about 39.2%. Ask the AI to calculate margin as a percentage and then recompute it yourself in the spreadsheet. If the two differ, trust your spreadsheet. Calculation slips are rarer than invented reasons but they do happen, and a gross margin that is off by a point is exactly the number a client will quote back to you.

Job three: chasing clients without nagging them

Chasing is the task nobody enjoys and everybody leaves too late. Handing the first draft to AI works well because the facts (what is missing, from when, and why you need it) are already in your open-items list. The AI only has to make the message clear and polite. For the longer view on this job, see chasing missing client records before deadlines.

Before, a typical chaser: "Hi, still need the October stuff, can you send asap, thanks." After, with the open items supplied:

Hi [first name], to finish October's accounts we need three things from you by Wednesday: the receipt for the $640 payment on 14 October, a note on what the $1,200 transfer on 22 October was for, and the statement for the second card account. If any of these are easier to send as a photo from your phone, that works too. Reply with whatever you have today and we'll start on the rest. [your name]

Two refinements make a difference. First, tell the AI the client's usual channel and tone, so it writes a text-sized message for the sole trader who never opens email and a formal note for the finance manager. Second, ask for a second message in the sequence, for the client who hasn't replied after three working days, with a clear consequence stated factually: "the pack will go out showing these items as unexplained." Don't let it threaten penalties or imply anything about the client's tax position, which is your adviser's call.

What to hand over second and what stays human

Once the first three jobs run cleanly for two or three cycles, a second tier is worth trying. The table ranks them by how forgiving the task is of mistakes.

Receipt capture is a good example of the built-in kind of help. Xero's Smart Document Capture, launched on 1 July 2026, reads bills and receipts and is free whatever the small-business plan, and its product page says you double-check and approve the details before Xero creates a financial record. Dext prices practices per client (its practice plans start at $17.70 a client a month on an annual subscription, with a ten-client minimum), while its plan for a single business is $302.50 a year for 250 documents a month and five users, so a small practice should compare the two routes for the clients who send piles of paper. Whichever you use, the approval step is the control. If a tool posts without asking, treat it as a higher-risk item from the table above.

The review step that makes it safe

A review step that only says "check the output" doesn't work, because the person checking sees a polished paragraph and trusts it. Give the reviewer a short list of things to confirm against source, and make the sign-off visible in the working papers.

1. Every figure in the text matches the ledger or the report it came from.
2. Every stated reason has a ledger line, a document or a client answer behind it. Reasons without one are rewritten as questions.
3. No client name, bank detail or other red item appears in the output or in the chat history.
4. The tone suits the recipient, and nothing in it reads as tax or legal advice.
5. The preparer has initialled the working paper, and the reviewer has initialled the review. The AI is named as a drafting aid in the file note, not as a preparer.

Sampling helps too. Once a month the partner picks three clients at random, takes the AI-drafted note and traces three numbers in each back to source. If any of the nine checks fails, the practice goes back to checking every client for a month. It is slow, and it is also how you find out whether your masking and review habits actually hold.

The Thursday afternoon paste

An illustrative failure shows how it usually goes wrong. A senior, behind on the last three clients, opens the free version of an assistant on her phone because the practice login is on the other screen. She pastes in an unmasked aged-debtors list for a client whose customers are mostly individuals, asks for a chaser summary, and gets a good one. Nothing visibly goes wrong. But the list held customer names and amounts under a personal account, with the training switch in whatever state it happened to be in.

Nobody notices until the quarterly review, when the partner sees the chat in the history of a personal account. The fix is three habits, not a lecture. Block the personal-plan sites on practice machines if your IT allows it. Put a one-line reminder at the top of the shared prompt document ("Masked data, business account only"). And make "I pasted something I shouldn't have" a thing you can say out loud without blame, because the faster it is reported, the sooner the account's data controls can be used to delete the chat and the client can be told if need be.

Four weeks to the first reliable close

Don't roll this out across all 42 clients at once. A staged start gives you evidence before it gives you habits.

1. Week 1, green only. Write the masking rule, choose the account and rewrite your three chaser templates with AI help. No client figures yet.
2. Week 2, three clients. Pick three straightforward clients and use AI for reconciliation commentary and the chaser. Time each task and note every correction the reviewer makes.
3. Week 3, add variance notes. Use the "cause unknown" prompt on the same clients. Count how many AI statements the senior had to rewrite. If it is more than a third, tighten the prompt before widening.
4. Week 4, ten clients and a file note. Extend to ten clients, add the cover-note draft and write a one-page procedure so a new starter can follow it.

Judge the trial on three numbers: minutes per client per task, corrections per note, and any confidentiality incident (the target is zero). If corrections fall and minutes drop for two cycles in a row, widen. If not, find out why before adding more clients. A cheaper check is reading the notes aloud to a colleague who hasn't seen the file: if they ask "how do we know that?" more than once, the note is claiming too much.

Before you widen further, run the practice through an AI readiness checklist for professional firms, and plan what the freed hours become with turning hours saved into advisory work. If you'd rather map your own close with someone, that is what my AI implementation consultation covers.

Checking it worked after two months

Two months in, pull the numbers rather than the impressions. Compare the average close duration (day the first pack went out to day the last did) with the same two months last year. Count client queries sent per pack and answered within five working days, since good chasers should lift that rate. Review the corrections log: the same correction repeated four times means a prompt needs a new rule. And watch for any client saying the notes feel more or less personal than before. If the answer is "less", add a sentence in the preparer's own voice to every cover note, which costs thirty seconds and brings the human back.

Do all of this and AI becomes a quiet part of the close: a drafter that never touches the ledger, never sees a name, and never speaks to a client without a person having read it first. That is a smaller promise than "automate your accounting", and it is the one that survives a client asking where their data went. When trials stall at this point, why AI adoption stalls in accounting firms explains the usual reasons and how to restart.

Questions accountants ask before the first month-end with AI

Can I paste a client's trial balance into ChatGPT or Claude?

Only on a business plan whose terms say content is not used for training, and only after you have removed names, account numbers and anything identifying. Even then, check your engagement letter allows it. On a free or personal plan, don't paste client figures at all unless you have switched the training setting off and accepted the risk.

Do I need to tell clients I use AI?

Many firms add a line to the engagement letter saying AI tools may be used to draft working papers and messages under partner review, with no client data used to train models. Your professional body or insurer may have its own expectation, so check with them. Telling clients is cheap and removes a surprise later.

Should AI post journals or do the reconciliation itself?

Start with AI as a drafter, not a poster. It writes the commentary, the chaser and the query list; a person matches the transactions and posts the entries. Bookkeeping tools that suggest matches are fine because you approve each one. Moving to automatic posting is a later decision, made per client, after weeks of clean results.

What if the AI gives a plausible but wrong reason for a variance?

That is the main risk, so build it into the process. Tell the AI to say 'cause unknown' when the data does not show one, and make the preparer confirm every stated reason against a ledger line, an invoice or the client. Anything unconfirmed gets turned into a question for the client, never into a statement.

Further reads

• 7 AI Mistakes That Put an Accounting Firm's Client Trust at Risk — Seven ways a firm loses client trust through AI use.
• Best AI Tools for Small Accounting Firms, Sorted by Task (2026) — Tools sorted by task once you know your first jobs.
• What an AI Implementation Looks Like in a Small Accounting Firm — What a full roll-out looks like in a small practice.
• How Small Accounting Firms Use AI Day to Day: Real Examples — More day-to-day examples from small practices.
• How Much Does AI Cost a Small Accounting Firm? — What the whole set-up costs a small practice.
• AI Accounting Workflow: Client Upload to Reviewed Books, Step by Step — Five stages from client upload to partner review, each with an owner, tool and check, shown through one worked month for a small contractor client.
• Industries Where AI Is in Use: What Small Firms in Each Actually Do — Seven industries, one concrete small-firm AI job each, with the monthly cost, a sample of the work and the check that keeps it safe.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI business and ChatGPT pricing pages; Claude pricing page; Microsoft 365 Copilot Business page; Xero Smart Document Capture page and launch post; QuickBooks community team answer on AI features; Dext business and practice pricing pages.