The prompts that save bookkeepers the most time each do one repeat job: answer a client query from facts you paste in, chase specific missing items, or turn rough call notes into a proper file note. Below are twelve templates grouped by those three jobs, each with the facts it needs, an illustrative output and the fix to make before sending.
The caveat that shapes every prompt here: ChatGPT knows nothing about your client's ledger, and it will fill gaps with confident guesses, including tax rules that belong to a different country or year. So each template carries the facts in and forbids the model from adding its own. Use a business plan (ChatGPT Business doesn't train on your content by default) and strip identifiers you don't need. One more housekeeping point: OpenAI is retiring custom GPTs, which stop running on 11 December 2026, so keep your standing instructions in a Project rather than building a GPT.
Set up one Project so you never paste your house style again
Create a Project called something like "Client correspondence" and put your standing rules in its instructions. Every chat inside it inherits them, and colleagues on the same Business workspace can share it. Here is a filled-in version for an illustrative two-person bookkeeping firm:
You draft emails and notes for a two-person bookkeeping firm.
Clients are small businesses: recruitment agencies, insurance
brokers, trades, landlords.
Rules:
- British English. Plain words. Short paragraphs.
- Never state a tax rule, rate, allowance or deadline unless
I have given it to you in the message.
- Never invent figures, dates or document names. If something
is missing, write [CHECK: what is missing] in the draft.
- Sign off as "Jo" unless told otherwise.
- Maximum 180 words for client emails unless I ask for more.
- Warm but brief. No "I hope this email finds you well".
- Account numbers: last four digits only.The [CHECK: ...] rule does more work than any other line. It turns the model's urge to guess into a visible gap you can fill, and it makes review quicker because you scan for square brackets first. If you want more on setting Projects up per client, see using ChatGPT Projects to keep client work separate.
Four prompts for client queries
1. "Why doesn't my bank balance match my profit?"
The most common question from owners who have just seen their first management accounts. The answer is always in the numbers, so paste them.
A client asks why their profit for the quarter is higher than
the increase in their bank balance. Explain it in under 150
words, using only these figures:
Profit for the quarter: $18,400
Bank balance change: +$6,900
Customer invoices issued but unpaid at quarter end: $7,200
Loan repayments (capital): $1,800
Owner drawings: $2,500
End with an offer to go through it on a call.Illustrative output: "Hi [first name], good question, and a common one. Your profit counts income when you invoice, not when you're paid. At the end of the quarter customers still owed you $7,200, so that's profit that hasn't reached the bank yet. Two other things took cash out without reducing profit: $1,800 of loan repayments and $2,500 you drew for yourself. Put together, $7,200 + $1,800 + $2,500 is $11,500, which is exactly the gap between your profit of $18,400 and the $6,900 rise in your balance. Happy to walk through it on a quick call if that helps. Jo"
What to fix: check the arithmetic yourself every time. Here it holds ($18,400 minus $6,900 is $11,500), but models do slip on sums, and one wrong total in a reply about money undoes the whole explanation. If your figures don't reconcile exactly, the model may invent a balancing item; the [CHECK] rule should catch it, but look.
2. The long, worried email
Some client emails bundle six questions and a lot of anxiety into one paragraph. Ask the model to separate them before it drafts anything.
Below is a client email. First, list each separate question or
request as a numbered line. Second, draft a calm reply that
answers only items I mark as ANSWER and says we'll come back on
items I mark LATER. Don't answer anything yourself yet.
Email: {paste}Before: "Hi, I got a letter about my return and I'm panicking, also my accountant before you said I didn't need to register for anything but now my friend says I do, and I haven't done the payroll for the new starter yet, is that a problem, and can you also look at why my card payments look wrong?" After the first step, the model returns four numbered items: the letter, a registration question, payroll for a new starter, and card payment reconciliation. You mark 3 and 4 as ANSWER, 1 and 2 as LATER (they need the letter and proper advice), and the draft comes back with a clear structure instead of a wall of reassurance. The split step takes ten seconds and prevents the most common failure: a draft that confidently answers the registration question with rules the model made up.
3. "Can I claim this?"
Whether an expense is allowable depends on the tax system and the facts, so the useful prompt drafts a holding reply that gathers facts rather than answering.
A client asks whether they can claim {item} as a business
expense. Do NOT say yes or no. Draft a short reply that:
- thanks them for checking before spending,
- asks the 3-4 facts we need to decide (business vs personal
use, who paid, whether they have an invoice, what it's for),
- says we'll confirm once we have those.
Under 120 words.Illustrative output for "a new laptop": "Thanks for checking first, that makes it much easier to get right. A few quick questions: will the laptop be used only for the business, or partly for personal use too? Was it bought on the business card or personally? Do you have the invoice showing the business name? And roughly what will you use it for day to day? Once we have those, we'll confirm how it's treated in your accounts. Jo"
Edits needed: almost none, which is the point. The only edit is adding the portal link for the invoice. Where models go wrong with this prompt is when you remove the "do NOT say yes or no" line; they then answer with rules from whichever tax system dominated their training data.
4. The request that's outside your engagement
A client has asked us to {request}. This is outside our
engagement, which covers {scope}. Draft a friendly reply that
says we can help, that it's additional work, and offers two
options: a fixed fee of {fee} or a referral to {referral}.
No apologising more than once.What this looks like for a mortgage adviser client who asks for "a quick letter confirming my income for the lender": the draft explains that an income confirmation is separate from the monthly bookkeeping, quotes your fixed fee, and gives a turnaround. The fix is usually tone: models apologise three times by default, which reads as if you're embarrassed to charge. The "no apologising more than once" line mostly cures it. Also check the scope wording the model uses to describe your engagement: it should match your engagement letter's words, not a paraphrase, because a client who later disputes a fee will quote your email back to you.
Four prompts for chasers
5. The specific missing-items chaser
Write a chaser to {first_name} for these outstanding items,
each with its reason in brackets:
{items}
Deadline: {date}. Upload link: {link}.
Friendly, under 120 words. List items as bullets, word for word.The phrase "word for word" matters. Without it, "Loan statement, finance agreement ending 2291" becomes "your finance paperwork", and the client can't tell which document you mean. If you're chasing dozens of clients against one deadline, turn this into a schedule rather than one-off prompts; chasing missing client records before deadlines sets out the full ladder.
6. Turning unexplained bank lines into a client query list
Every month-end leaves a handful of transactions you can't code. Pasting them as raw bank text and asking the client "what are these?" gets slow, partial answers. Let the model build a friendlier list.
Turn these unexplained bank transactions into a short table
for the client with columns: Date | Amount | Description as
shown | Our question. Make each question answerable in a few
words. Keep amounts and dates exactly as given.
{paste lines}Illustrative output (first three rows): "12 Jun | $240.00 | CARD 4471 AMZN MKTP | Was this for the business? If so, what was it? / 14 Jun | $1,150.00 | TFR J SMITH | Is this a payment to a subcontractor, a loan, or something else? / 19 Jun | $36.99 | DD CLOUDSTORE | Is this a business subscription, and do you have the invoice?"
The check here: make sure the model didn't "tidy" the descriptions. In one realistic slip, "TFR J SMITH" came back as "Transfer to John Smith", a name the model invented from an initial. The client then denies knowing any John, and you've lost a day. Keep the description column exactly as the bank shows it.
7. Chasing your own unpaid fees
Draft three versions of a reminder for our unpaid invoice
{number} for {amount}, due {date}, now {days} days overdue:
A) friendly nudge, B) firm second reminder, C) final notice
stating we'll pause work on {date} if unpaid. Each under 90 words.Illustrative version B: "Hi [first name], a reminder that invoice 1043 for $420 was due on 2 May and is now 21 days overdue. If it's already on its way, thank you and please ignore this. If there's a problem with the invoice, reply and we'll sort it out. Otherwise, could you arrange payment this week? Thanks, Jo"
Asking for all three at once gives you a consistent sequence you can save and reuse. The "if there's a problem with the invoice" line is worth keeping in every version; it gives a client who is quietly unhappy about the bill a route to say so, which is better than silence.
A second check: read version C against your engagement letter: if your terms don't allow you to pause work, the model has just threatened something you can't do. That is a realistic mistake worth catching before it goes to a financial planner client who reads contracts for a living.
8. The final written reminder before a filing deadline
Write a final written reminder to {first_name}. The deadline is
{date}. Still outstanding: {items}. Include this sentence exactly:
"{approved_consequence_line}". Offer a 10-minute call on
{two_slots}. Under 130 words. No other warnings.Facts to paste in: first name, the deadline, the outstanding items as they appear on your list, your approved consequence line, and two real call slots from your calendar. If any of those is missing, stop and find it rather than letting the model improvise; this is the one email where an invented detail can cause a complaint.
The approved consequence line is written once by a partner and reused; for example, "If we don't receive these by 8 March, we may not be able to file on time, which can lead to penalties." The model's job is the framing around it, not the warning itself. Offering two specific call slots gets more responses than "let us know if you'd like a call".
Four prompts for file notes
9. Call scribbles into a file note
Turn these rough call notes into a file note with headings:
Date and attendees | What the client told us | What we advised
or agreed | Actions (who, what, by when) | Open questions.
Don't add anything I didn't write. Mark unclear items [CHECK].
Notes: {paste}Before: "call w/ [client] 10 mins - new van ?k on finance, starts Jul, wants to know re monthly cost. also recruiter contract ending Sept, might go employed?? told him we'd look at it. send loan agmt." (Your own notes will be this messy; that's fine.) After: a tidy note with the van finance under "What the client told us", "We said we'd review options once we have the loan agreement" under "What we advised", an action for the client to send the agreement, and two [CHECK] flags: the van's cost and what "might go employed" meant. The flags are the valuable part; they're the questions you would otherwise forget to ask. Notice the model left "?k" alone rather than guessing a price; that is the "don't add anything" line doing its job.
10. Month-end handover note for a colleague
I'm handing this client's month-end to a colleague. From the
notes below, write a handover under 200 words: status, what's
done, what's outstanding, quirks of this client, and the one
thing most likely to go wrong. Notes: {paste}"The one thing most likely to go wrong" is the line colleagues thank you for. For an insurance broker client, the model might surface from your notes that commission statements arrive two weeks after month end and the colleague shouldn't chase them early. Check the quirks section against what you actually know; models sometimes promote a one-off comment into a "pattern".
11. A long email thread into a client history
Summarise this email thread for the client file in 6-8 bullet
points, oldest first, each starting with the date. Include
decisions, figures agreed and anything the client promised to
send. Leave out pleasantries. Thread: {paste}Useful when a client relationship has lived in one person's inbox for two years. The check is simple: every date and figure in the summary should appear in the thread. If a bullet says "agreed fee of $150 a month" and you can't find "150" in the original, the model got it from somewhere else.
12. A recurring question into an internal FAQ entry
We've answered this client question several times. Write an
internal FAQ entry: the question as clients phrase it, a
3-sentence answer in our house style, the facts to check before
using it, and when NOT to use it. Past answers: {paste 2-3}Feed it two or three replies you've actually sent, and it will produce a consistent entry your team can reuse. The "when not to use it" field is the safeguard. For "Why is my profit different from my bank balance?", the entry might say not to use the standard answer for clients who account on a cash basis, where the explanation is different. Over a few months these entries become the core of a shared prompt library for your team.
A Monday inbox run through the twelve prompts
Individual prompts are easy to admire and hard to picture in a working week, so here is an illustrative Monday for the two-person firm from the Project example, with 60 clients on monthly bookkeeping. The inbox holds 26 client emails from the weekend and Friday afternoon.
- Sorting (10 minutes). Jo reads subject lines and first sentences and tags each email by prompt number. Nine are uploads or thank-yous that need no reply. Seventeen need something.
- Queries (35 minutes for eight emails). Three "why is my bank balance lower?" questions go through prompt 1 with each client's figures pasted in. Two long emails go through prompt 2; one of them, from a recruitment agency owner, turns out to contain a question about a disputed supplier bill that Jo marks LATER and books a call for. Two "can I claim this?" questions get prompt 3 holding replies. One request for a cash-flow forecast gets prompt 4 with the firm's fixed fee.
- Chasers (20 minutes for six clients). Four clients have unexplained transactions from last month, so prompt 6 builds their query tables. Two clients are late paying the firm's own invoices; prompt 7 produces version A for one and version B for the other.
- Notes (15 minutes). Friday's three phone calls become file notes through prompt 9, and one thread about a client's new premises goes into the file via prompt 11.
- Review (20 minutes). Jo scans every draft for square brackets first, then checks every figure against the ledger. Two drafts need a sentence rewritten; one prompt 1 reply had the loan repayment figure in the wrong place and is corrected.
That is about an hour and forty minutes for work that, written from scratch, typically fills most of a Monday morning and spills into the afternoon. The figures are illustrative, but the shape holds for most practices: the time saved is in the drafting, and the time that remains is in reading figures carefully. If the review step starts shrinking below about a fifth of the total, that's a warning sign, not an efficiency.
What this looks like for a firm with a different client mix: a bookkeeper whose clients are mostly law firms will lean more heavily on prompts 10 and 11, because client-account reconciliations generate long email threads and handovers matter when a colleague covers month end. A bookkeeper for trades gets far more prompt 6 work, because card payments at builders' merchants arrive with cryptic descriptions.
What never goes into these prompts
- Full bank statements or ledgers when a few lines will do. Paste the transactions you're asking about, not the whole month.
- Personal identifiers the draft doesn't need. A first name is enough for a reply. Tax reference numbers, dates of birth and full account numbers add risk and no value; anonymising client data before you paste it into AI shows a quick routine.
- Anything on a free or personal account. If the team uses personal ChatGPT logins, switch off the model-training setting at minimum, and move client work to a business workspace.
- Bank details for payments. Never let a draft include payment instructions; those should come from your practice software or a fixed template, which also protects clients from payment-diversion fraud.
- Questions that need advice. Prompts 2 and 3 exist precisely so the model gathers facts instead of advising. Keep it that way.
Keeping the prompts sharp after the first month
Prompts drift. Someone adds a line, someone else deletes one, and in three months nobody knows which version produces the good chasers. Keep the twelve templates in one shared document with a version date, and change them deliberately: when a draft needs the same fix twice, add a line to the template that prevents it. The fixes noted above ("word for word", "don't tidy descriptions", "no apologising more than once") all started as repeated edits.
Once a month, pick five sent emails and ask two questions: did the client understand it without a follow-up, and did we have to rewrite more than a sentence? If the answer to the second is often yes, the problem is usually the facts you're pasting, not the wording of the prompt. For whether these tools are safe with client figures at all, is it safe for an accountant to use ChatGPT with client data? covers the settings and plans in detail.
Further reads
- ChatGPT Prompts for Small Business Owners: 50 Tested Examples — Broader prompt set for the owner-side tasks your clients ask about.
- AI Email Triage for Professional Firms: Sort, Summarise, Draft — Sort the inbox before these prompts draft the replies.
- A Five-Minute Fact-Check Routine for AI Output Before It Goes Out — A five-minute check to run on every client-facing draft.
- Is AI Worth It for a One-Person Bookkeeping Business? — Whether a paid plan pays back for a sole bookkeeper.
- How to Write Reply Templates That Keep AI Replies On-Script — Turn your best replies into templates the AI follows.
- How Accountants Use AI to Explain Tax to Clients in Plain English — The same approach applied to explaining tax to clients.
- Why AI Gives Different Answers Each Time, and How to Fix It — Why the same prompt gets a different answer, why 'set temperature to zero' no longer works, and the fixes and consistency test that do.
- Rolling Out AI in a Bookkeeping Practice Without Losing Control — A staged AI rollout for bookkeeping practices: control points, an inventory of AI already in your software, a pilot, sampling rates and client wording.
- 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 Small Tax Practices Use AI Through the Busy Season — A season-long plan for tax preparers: what to build eight weeks out, how AI handles intake, chasers and the inbox, and where preparer review stays in charge.
- How Accountants Use AI to Spot Errors in Client Books — The error checks worth running on every client file, the tools that run them, and how to turn thirty flags into the six corrections that matter.
- How to Chase Late Payments With AI Reminders That Sound Human — A reminder sequence, AI-personalised from each client's history, with tone rules, sample wording for different payers, reply sorting and a hand-off to a person.
- How to Categorise Transactions With AI and Check Its Work — A category guide, bank rules for the routine 60%, AI for the rest, and a monthly sampling check that tells you when its coding has drifted.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI help pages (Custom GPT retirement and migration FAQ; Projects in ChatGPT); OpenAI ChatGPT Business pricing and data-use pages.