How Accountants Can Use ChatGPT Effectively: Review Notes and Memos

An accountant can use ChatGPT effectively by redacting client details first, then asking it to do the writing around the judgement: tidy review notes, structure a technical memo, or draft a client reply. Every figure and citation it produces is then checked against the source. Client names, numbers that identify a person, and raw files stay out of the chat.

Most of the value is in turning rough thinking into clean prose, not in getting answers. A reviewer's jotted comments become numbered points a junior can act on; a page of facts becomes a memo in the firm's house shape. Where firms get into trouble is trusting the tool's technical references: a paragraph number that looks right is the commonest way a fluent draft goes wrong, and only a check against the source catches it.

Follow me on Instagram@sagnikteaches

Three tiers of material, and which one may go in

Before any prompt, sort what you are about to paste. Most arguments about "is it safe" dissolve once the material is placed in a tier.

Connect on LinkedInSagnik Bhattacharya

The privacy position differs by plan. OpenAI's help centre says it won't train on a ChatGPT Business workspace's data, while on Free, Plus and Pro the setting called Improve the model for everyone, under Settings then Data controls, is on by default until the user turns it off. Check the current wording on OpenAI's pages rather than trusting a screenshot from last year. Whether your own engagement letters and professional rules allow even redacted use is a separate question for your own firm and regulator; whether it is safe for an accountant to use ChatGPT with client data goes through it in detail.

Subscribe on YouTube@sagnikteaches

Redacting before you paste: a worked pass

Redaction is not find-and-replace on one name. Identifying detail hides in amounts, dates, industries and odd phrases. Here is an illustrative case summary, before and after.

BEFORE
[Trading name] Pet Supplies Ltd (director: [initial and surname]) sold its trade to a new
company on 14 March for 212,400 and the old company owes 38,950 to
the director's loan account. Bank: account ending 4471.

AFTER
Client A, a small retailer, sold its trade to a connected company during
the year for [amount X, about 200k]. The old company has a loan from
the director of [amount Y, about 40k] outstanding. No bank details.

Notice three moves. The name and director are gone. The exact figures became approximate bands, because the technical answer rarely depends on the pennies. And the date became "during the year". Keep a private key in your own file mapping Client A to the real client, so you can put the figures back when the answer comes back.

A quick redaction checklist keeps this consistent across the team:

1. Names of people, companies, and any identifying trading name.
2. Addresses, postcodes, phone numbers, email addresses, tax and company numbers.
3. Bank, card and account fragments, even partial ones.
4. Exact amounts and dates, unless the answer depends on them. If it does, round or shift.
5. Unusual facts that identify the client on their own, such as a rare industry in a small town.
6. File names and document metadata if you are uploading, not pasting. If in doubt, paste text rather than uploading the file.

For a fuller method, see how to anonymise client data before you paste it into AI.

Pattern one: turning reviewer jottings into review notes

Review notes are the cleanest first use, because the input is your own reviewing thoughts, which usually contain no client identifiers. The usual problem is that jotted comments are cryptic to the preparer, and clarifying them costs a second conversation.

Say a reviewer scribbles these on a set of working papers for a mid-sized retail client (illustrative):

rent accrual? Q3 looks off. dep'n on new van, rate 20 pct but old vans 25.
stock count cut-off, check last 3 deliveries. director loan, interest
charged? ask. fixed asset reg doesn't agree to TB by 1.2k.

The prompt:

You are helping an accountant turn jotted review comments into review notes
for a junior preparer. Rewrite each jotting as a numbered note with:
(1) what the reviewer noticed, (2) what to check or do, (3) what evidence to
leave on file. Keep each note under 45 words. Do not add issues I did not
mention. If a jotting is ambiguous, list it under "Ask the reviewer".

A realistic reply, trimmed:

1. Rent accrual, Q3. Noticed: the accrual looks out of line with other quarters.
   Check: recompute from the lease and compare with the ledger. Evidence:
   recalculation on file.
2. Depreciation on the new van. Noticed: 20% used, older vans at 25%.
   Check: confirm the policy rate and explain any difference. Evidence: note.
3. Stock cut-off. Check the last three deliveries against count date...
Ask the reviewer: "Director loan, interest charged?" Is the question whether
the loan agreement provides for interest, or whether interest was accrued?

What you would fix: the tool correctly put the loan comment in "Ask the reviewer" rather than guessing, which is the instruction doing its job. Check that it did not add a note on the fixed asset register difference that you did mention; here it silently dropped it. Counting the jottings against the notes catches this in seconds. Say a practice reviews 40 files in a busy month and the note-writing drops from about 20 minutes to 8 per file. That is roughly eight hours back, which pays for a business-plan seat many times over. The figures are illustrative; time your own first five files before you quote any saving.

Pattern two: a technical memo in the firm's house shape

A technical memo records a question, the facts, the relevant rules, the analysis and the conclusion. ChatGPT is good at the structure and the plain-English analysis; it is unreliable at the specifics of rules. So split the job: you supply the rule wording, the tool writes around it.

An illustrative question: a client has billed a customer $12,000 upfront for a 12-month support contract starting 1 October. When should the revenue be recognised in the year ending 31 December? The skeleton prompt:

Draft a technical memo with these headings: Issue, Facts, Relevant guidance,
Analysis, Conclusion, Open points. Use only the facts and guidance text I
paste below. Where you would normally cite a paragraph, write [CITE] and
leave it for me. Do not quote any guidance that is not in my paste.

FACTS: Client A billed $12,000 upfront for a 12-month support contract
starting 1 October. Service is delivered evenly across the term.
GUIDANCE: [paste the relevant paragraphs from your own library]

The analysis section should come back along the lines of: the service is delivered over time, so revenue is recognised across the 12 months, $1,000 a month. For a year ending 31 December that is three months, $3,000, with $9,000 deferred. Do the arithmetic yourself: $12,000 divided by 12 is $1,000; three months is $3,000; $12,000 less $3,000 leaves $9,000 deferred. If the tool wrote anything different, find out why before using it.

Two habits matter. First, the [CITE] placeholder stops invented references, and you fill them from the real text. Second, the "Open points" heading gives the tool a legitimate place to list what it could not settle, so it is less tempted to fill gaps with confidence.

A second memo, where the tool went wrong

A common realistic failure: a practice asked about capitalising a batch of laptops costing $4,800 in total across six staff. The draft said the items "must" be expensed below a stated per-item threshold and named a paragraph number. There was no such threshold in the standard; it was the client's own policy, mentioned in the facts, which the tool had promoted into a rule. It showed up because the reviewer searched the memo for the word "must" and compared each instance to the pasted guidance. Make that search part of your check: find every "must", "required" and "always", and tie each to a source line.

Pattern three: client queries that touch advice

Clients ask questions like "can I claim my home office?" or "should I take the dividend now?". An AI draft can be useful for the tone and structure of the reply, but the answer is professional advice. OpenAI's usage policies bar building or offering AI tools that give others tailored advice requiring a licence without appropriate involvement by a licensed professional. A draft you review and send yourself is not that, but the principle carries over: you are the qualified professional in the loop, and the draft goes through you.

The prompt separates the parts you provide from the parts you want written:

Draft a reply to a client email. Tone: warm, plain, no jargon. Structure:
a one-sentence answer, two short reasons, what I need from them, and a
sentence saying this depends on their circumstances. The answer is: [your
conclusion]. The reasons are: [your reasons]. Do not add legal or tax
points I have not listed. Use [first name] and [your name].

That wording matters. You are handing the tool your conclusion and asking for a clear way to say it, not asking it to reach one. For more on turning technical answers into plain client wording, see how accountants explain tax to clients in plain English with AI.

A client reply, drafted and then corrected

Here is the third pattern end to end, with an illustrative sample so you can see what a correction looks like. A client asks whether they can pay their spouse a salary from the business. You decide the answer is "possibly, if the work is real and documented", and you give the tool that conclusion with two reasons.

FIRST DRAFT (from the tool)
Hi [first name], yes, you can pay your spouse a salary, and it will reduce
your tax bill. I'd recommend setting it up this month. [your name]

AFTER YOUR CORRECTION
Hi [first name], it can work, provided your spouse does real work for the
business and the pay is reasonable for it. We would need a written role,
time records and proper payroll. Whether it lowers your bill depends on your
circumstances, so let's talk it through before you commit. [your name]

The first draft turned a conditional into a promise ("it will reduce your tax bill") and added a deadline nobody asked for. Both are the sort of lines a quick read misses because the sentence flows well. The search for "will", "yes" and "recommend" catches them. Keep the corrected version in your notes as an example of the voice you want, and paste it into the Project as a style sample next time, with the client details stripped.

Finally, put a line on the file recording that the draft was tool-assisted and checked by you. It costs ten seconds, and if a client or regulator ever asks how a letter was produced, the answer is already written down.

Checking the output against source, every time

Different outputs need different checks. This table is worth printing.

If you want a second line of defence, ask the tool to critique its own draft in a new message: "List every statement in this memo that is not supported by the guidance I pasted." It will not catch everything, but it often flags sentences that smuggled in outside knowledge. Treat that list as a prompt for your own checking, not a clearance. A routine for broader source-checking is in how to check sources and citations in AI research.

Pattern four: sorting a messy client email thread into a query list

Review notes and memos are your own words. A fourth use takes the client's words: a long email thread where a client answers some of your queries, ignores others and raises new ones. Redact it first, since client emails are full of names. Then ask the tool to separate what is settled from what is open.

Below is a redacted email thread between an accountant and a client.
Produce a table with columns: Query, Status (answered / partly answered /
open), What the client said (one line), What I still need. Do not infer
answers the client did not give. Put anything the client raised that was
not one of my queries under "New points".

Suppose the thread covers a sole-trader client's year-end queries (illustrative). The tool returns a table in which "Sales to a new customer, invoice 112" is partly answered ("client says it was paid in January, no date given"), "Van purchase receipt" is open, and "New points" lists the client's question about hiring a part-timer. The last item is the valuable one. Left in the thread, it would have been missed until the next call. What to fix: check that "answered" really means answered. The tool tends to be generous when a client writes "I think so" or "should be fine". Mark those as partly answered yourself, and ask the tool, in a second message, to list every hedged phrase from the client.

The chaser that follows

Once the table exists, one more prompt writes the chase from the open rows only. Keep it to a numbered list of what you need, a date, and one sentence on why it matters, so the client can reply in a single message. The same pattern works for a quarterly query round across a dozen clients, as long as each thread is redacted and processed on its own, so facts from one client never sit in the same conversation as another's.

Writing review notes that are firm but kind

Tone is where a tool can quietly help. Review notes written at the end of a long day can read as sharp. You can ask for a rewrite that keeps the instruction but softens the delivery. Here is a before and after (illustrative):

BEFORE: Why is the accrual wrong again? Redo it and check the lease this time.
AFTER:  The Q3 rent accrual does not agree to the lease. Please recompute it
        from the lease terms and leave the calculation on file. Shout if the
        lease copy is missing and I will find it.

Notice that the second version contains the same instruction and a quicker route to help. The prompt is simple: "Rewrite this note so it is direct, specific and courteous. Keep every instruction. Do not add new ones." Check that "every instruction" survives, because softening sometimes turns "must" into "might want to". For a junior preparer, that drift can change whether the thing gets done.

What a six-person practice might actually spend and save

An illustrative sum helps decide whether to bother. Say six staff, of whom four do review or memo work. OpenAI's help centre lists a Standard ChatGPT Business seat at $25 a user a month billed monthly, or $20 a month billed annually, with at least two seats per workspace; the public pricing page shows only "starting at 2 users", so confirm before budgeting. At $25 across four seats, that is $100 a month. If each of the four saves two hours a month on notes and memo drafting, that is eight hours, and at an internal cost of $40 an hour the saving is $320 against $100 of cost. The sum is only as good as the two-hours assumption, so measure it: time the five-file trial in the five-day plan below and replace the guess with your own number.

Count the hidden cost too. Redaction takes about three minutes per case. If a memo needs it, that is three minutes you would not spend otherwise, which is why redaction suits memo-sized tasks better than a five-line email you could write faster yourself. A reasonable rule is not to use the tool at all for replies under about 80 words.

Keeping a prompt library the team trusts

Prompts drift. Someone adds "be more confident" to a memo prompt and the output begins to overstate conclusions. Keep the library in one place with a version number and the date of the last change, and assign one person to own it. Each entry should carry four fields: purpose, the prompt text, what the output must be checked for, and a known failure. For the memo prompt the known failure is invented citations; for the review-notes prompt it is dropped items; for the client reply it is overpromising. Reviewing the library quarterly, alongside the sample of chats mentioned in the questions below, keeps the habit alive without a large project.

Setting it up so the safe way is the easy way

A saved Project in ChatGPT can hold your firm's house style, memo template and tone notes, with no client facts. Staff open that Project instead of a blank chat, with the house style in the project instructions and the prompts above in a saved prompt file. OpenAI says a shared project can draw on everything in it, chats as well as files, and members can see one another's chats there, so keep the shared Project for templates only and run redacted client work in a separate personal chat. That way sharing a Project never shares client data. On a business plan, remember the two-seat minimum, so a sole practitioner may price an individual plan against a two-seat business one.

Write a one-page rule and put it where people work. A workable version:

• Redact first, using the checklist. If you cannot redact it, do not paste it.
• Use the approved plan and the firm's Project only.
• Never rely on a citation until you have opened the source.
• Number the draft's version in the file note so a reviewer knows AI helped.
• Tell the client in the engagement letter that AI tools may assist drafting, if your firm's policy requires it; AI engagement letters for accountants shows wording.

How this goes wrong in a small practice

Three realistic problems come up after the first month. A senior pastes a client's full management accounts because the redaction step "felt slow": fix it by making a redaction template with the replacement table pre-filled. A junior trusts a memo because it reads fluently: fix it by requiring a source line next to every conclusion. And the practice slowly stops reading its own review notes because the tool's version looks tidy: fix it by sampling five files a month and comparing the notes with the reviewer's original jottings. None of these is a technology fault; each is a process gap you can close in a week.

For a different slice of the work, ChatGPT prompts for bookkeepers covers client queries, chasers and notes at the bookkeeping end, while the patterns here stay with review and technical writing. Together they make a reasonable starter library.

Five days to trial the patterns on old files

1. Day one: agree the three tiers and the redaction checklist with the partners.
2. Day two: build the Project with the house memo template and the three prompts.
3. Day three: run five old files through the review-notes prompt and compare with the notes actually issued.
4. Day four: draft one real memo using the [CITE] method and record the checking time.
5. Day five: review what saved time, what was dropped, and what needed fixing, then decide whether to extend.

After that week your own files will show whether the tool earns its place in your process.

Practical questions from accountants about ChatGPT

Is a paid ChatGPT plan needed for client work?

Not for redacted, no-names material, where a Free or Plus plan with Improve the model for everyone switched off can be enough. Once staff share a workspace, or anyone is tempted to paste identifiable data, ChatGPT Business is the safer choice because OpenAI says it won't train on workspace data. It needs at least two seats; check current seat prices on OpenAI's pages.

Can ChatGPT cite the accounting standard for me?

It can point you to the right area, but it can also produce a paragraph number that does not exist. Treat any citation as a lead. Open the standard or your technical library, find the paragraph, and paste the actual wording into the memo yourself.

How do I stop staff pasting client data anyway?

Write a one-page rule with examples of what is allowed, give them a redaction checklist, and make the safe route faster than the unsafe one, for example a saved Project containing your house style and no client facts. Spot-check a sample of chats each quarter.

Does ChatGPT replace the reviewer?

No. It speeds up the writing around a review, such as turning jotted points into clear notes. The judgement about whether the numbers and treatment are right stays with the person signing off, and the tool's output should be reviewed like a junior's first draft.

Further reads

• How Accountants Use AI to Spot Errors in Client Books — Use AI to find mistakes in client books, with checks.
• 7 AI Mistakes That Put an Accounting Firm's Client Trust at Risk — Seven ways AI errors cost a firm its clients' trust.
• AI Tax Research Tools Compared: Citations, Cost and Checks — Compare research tools that show their sources.
• AI Readiness Checklist for Accountants, Solicitors, Consultants — Check whether the firm is ready to roll this out.
• Best AI Note Taker for Accountants: Which Fits a Small Practice? — Meeting notes for client calls, with retention costs.
• ChatGPT for Tax Preparers: How to Redact a Return Before You Paste — Redaction steps for a tax return specifically.
• Free ChatGPT for Accountants: When Client Work Needs a Paid Plan — When free ChatGPT stops being enough for a practice.
• ChatGPT for Sales and Marketing: B2B Battle Cards and One-Pagers — Prompts and a fact-check routine for a competitor battle card and a one-page service sheet in a small B2B firm.
• AI for Doing Tax Accounting: What to Hand Over and What to Review — A small-practice task split for document extraction, tax workpapers and client letters, with a named check before each output goes out.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI Help Centre (ChatGPT Business overview; Data controls in ChatGPT; Projects in ChatGPT), ChatGPT pricing and Business pages, and OpenAI usage policies, all read 4 November 2026. Sample figures are illustrative.