Can ChatGPT answer bookkeeping client emails?

Published 17 October 2026. Updated 17 October 2026.

A client writes: "Why is my April mileage so low? Can I still claim it?" You open a chat assistant, paste the question, and get a polite, well-organized reply in seconds. It reads better than what you would have typed at 6 p.m. So the obvious question is whether you need anything more.

For some emails you do not. For the ones that quote a number, a date or a rule, the gap between a fluent reply and a correct one is where firms get into trouble. This guide sets out what the job actually takes, where a general chat assistant helps, where it breaks down, and how to use either approach without sending something you cannot stand behind.

What does a good reply to a client's money question need?

Strip away the polite wording and a reply to "why is this number what it is" has four parts.

1. The facts, from the ledger. The transaction, the date, the amount, the category, the period. These come from the client's books, not from general knowledge.
2. The rule, from your firm. Whether mileage counts only when a trip has a date and purpose, how you treat a payment that falls just outside the quarter, what you do when a category looks wrong. These come from your own written procedures.
3. The limits. What you cannot say: what a deduction is worth, how the tax preparer will treat it, anything about an amended return, a notice from the tax authority, an audit or a penalty.
4. A trail. Enough detail that you, or a colleague, can check the reply against the sources in a minute rather than reconstructing it.

Everything below measures an approach against those four parts.

Where does a general chat assistant help?

A general AI chat assistant such as ChatGPT is good at language. That covers a surprising amount of the work around a client email.

• Tone and clarity. Turning a blunt draft into something warmer, or a long one into something shorter.
• Explaining a concept in plain words. A client who does not know what "uncategorized" means can be told so kindly. Check the explanation against your own practice before you send it.
• Reusable templates with no numbers. A request for missing documents, a reminder that a questionnaire is overdue, a note introducing a new process.
• Rewording after you have checked the facts. If you have already pulled the ledger line and decided what to say, the assistant can polish the wording.

None of that needs the assistant to know anything about this client or this firm, which is why it works.

Where does it break down?

The trouble starts when you ask it to supply the facts or the rules instead of the wording.

It works from what you give it. A general chat assistant sees the client's books only through what you paste or connect to it. If you paste an export, it works from that text. If you paste half of it, or last month's, it works from that. Unless you ask, the reply does not say which lines it used.

It does not know your rules. Unless you give it your procedures, by pasting them or setting them up in the tool, it answers from general practice, and general practice is not your firm's practice. The reply can sound confident while contradicting what you tell every other client.

Figures can drift. Language models can misread a table, drop a row or total a column wrongly, and the result reads just as smoothly as a correct one. The error is invisible unless you recompute the figure yourself.

No citations unless you build them. You can ask for line references, but nothing checks that they match the source, so checking them is again your job.

Risky topics can still get an answer. A general assistant may add a caveat, but nothing stops it from answering a question about an IRS notice or a penalty the way it answers one about mileage. Whether it holds back depends on the tool and on your prompt.

Client data leaves your systems. Pasting a ledger export into any outside tool is a decision about your clients' information. Check the tool's data terms and your engagement letter first.

Three ways to handle the same email

The first two columns describe each approach in general. The last column states only what GroundedReply's own documentation backs.

When is a chat assistant enough?

Use the general tool, and do not feel guilty about it, when the email contains no ledger figure, no date and no rule: a scheduling note, a friendly nudge, a welcome message. The same goes for any case where you already did the research and want help with phrasing.

By hand is still the right call for a delicate email, such as a client upset about a fee or a mistake you made. There the wording is the point.

When is a cited draft the safer choice?

The cases that cost you are the ones that quote a figure and a rule together: "why was this categorized that way", "did the invoice get paid", "how much did we spend on subcontractors last quarter". For those, the value is not eloquence. It is that every number traces to a line and every rule traces to a sentence you wrote, so that checking takes a minute.

A short checklist, whichever way you work

• Pull the ledger line yourself or confirm the line the tool cites. Match date, amount and category.
• Confirm the export covers the period the client is asking about, and note the date it was pulled.
• Find the sentence in your procedures that covers the question. If there is none, answer it yourself.
• Remove any dollar amount that is not in the ledger, and any statement about what a deduction is worth.
• Hold amended returns, tax authority notices, audits and penalties for a separate, careful reply.
• Read the draft once as the client would. Would they know what to do next?
• Keep the draft and the sources together so you can show how you got there.

For a longer version of this list, see the client email reply checklist.

How GroundedReply does this step

You upload three things: the client's question, your written procedures and a ledger export from QuickBooks or Xero. GroundedReply drafts a reply that cites the ledger line and the procedure sentence behind each claim, and returns it as plain email text, with a report that carries a not-tax-advice notice. You read it and send it from your own inbox. It never sends email and is not connected to your inbox or your accounting software.

Code checks every figure against the source it cites. If a figure cannot be traced, the draft is held back and flagged for a manual answer. If no sentence of your procedures matches the question, nothing is drafted and the job needs a manual answer. Amended returns, IRS or state notices, audits, penalties, legal matters and topics your procedures escalate get only a holding reply.

A second, stronger AI model reviews every result and lists what is left for you to decide; the drafting model corrects the serious problems the review finds, and code checks every correction. Results are not verified by a person. A holding reply for an escalated question is the exception: code writes it and the second model does not read it. A reply costs 6 credits ($6 to $4.20 by pack size), with no subscription. To try it on your own question, sign up. To get more out of any tool, start with how to answer a categorization question.

Questions

Is it safe to paste client ledger data into a chat assistant?

That depends on the tool's data terms and on what your engagement letter promises your client. Read both before you paste anything with names, amounts or account details, and ask your own advisor if you are unsure.

Can I just ask the assistant to double-check its own numbers?

It can catch some slips, but it is the same kind of system checking its own work, and it only sees what you pasted. The check that holds up compares each figure to the ledger line it came from, which you or the software must do.

What if my procedures are not written down yet?

Then any tool, including a chat assistant, is guessing at your firm's rules. Write down the five answers you give most often first. The guide on written procedures for bookkeeping firms shows how to word them.