Can ChatGPT do my bookkeeping? Five jobs yes, the ledger no

What can ChatGPT do for bookkeeping, and what still needs the ledger?

ChatGPT does the reading, drafting and checking half of bookkeeping and none of the posting. A chat model can read a bank export you paste, draft the rule that files "AMZN Mktp" for a landscaping client, turn a profit and loss into three sentences the owner will read, and tell you which side of a reconciliation is off. It cannot pull a bank feed, post a journal entry, or leave the audit trail a reviewer opens later. The QuickBooks app inside ChatGPT, live for US Free, Plus and Pro accounts since 5 March 2026, does not change that: Intuit's own list of what it does covers invoices, payroll questions, lending benchmarks and reports, and bank categorization and reconciliation are not on it. Where each job in a solo bookkeeper's week lands:

The survey behind the chart, 707 US accounting professionals in March 2024, found 98% had used AI to help clients in the previous 12 months, and the order of the uses is the order of the table: entry and cleanup first, checking second, explaining third. Posting is not on the list, because the people surveyed still do it themselves or their ledger app does it for them.

Can you use ChatGPT for QuickBooks?

Yes, through Intuit's QuickBooks app inside ChatGPT, which Intuit says is "live for all logged-in ChatGPT Free, Plus and Pro accounts, available on both web and mobile in the US" and which, since Intuit's 28 July 2026 update, can "Create, update, send, and delete invoices", answer payroll questions and pull profit and loss and cash flow statements. The same update put QuickBooks inside Claude. What neither app does, on Intuit's own list, is categorize the bank feed or reconcile an account. That work stays in QuickBooks, where the AI is priced by plan, and the plan step is steep.

We read the pricing pages of five ledger apps and one AI add-on on 27 September 2026, US pricing, monthly billing at the regular rate, and set them next to a chat subscription:

Read down the price column and the pattern is the fact page one misses: two of the five ledger apps charge extra for AI reconciliation. At QuickBooks, "Reconciles your books" and "Transactions auto-categorized and verified" start at Plus, $140 a month, $102 above Simple Start, and Intuit Intelligence's continuous categorization and posting is Advanced only at $340. Xero puts JAX and Smart Document Capture on its $25 Early plan and sells "Auto-reconcile transactions" there as a $5 a month add-on, included from Growing at $55. Zoho Books puts Zia on Standard at $20, Wave categorizes on Pro at $19, and FreshBooks' pricing page never uses the word AI. A $20 chat plan sits next to Xero and Zoho on price and does none of their posting, which is why the honest comparison is not chat versus ledger but chat plus ledger, at $39 a month on the cheap end ($20 chat plus Wave Pro at $19).

How can AI help me as a bookkeeper?

AI helps a bookkeeper most on the five jobs that are words about the numbers rather than the numbers themselves. Each one below is a prompt you can paste today, with the check that makes it safe to send. None of them asks the model what is deductible or which entity a client should form; that is tax advice, a chat model gives it confidently and sometimes wrongly, and it is the line this page stays behind.

The client email. A chat model writes the chase you have rewritten forty times, and writes it without the edge that creeps into the fortieth version:

```

Write a 90-word email to a client whose March bank statement and three vendor invoices I still need for the quarterly close. Friendly, one ask, one date (Friday 10 April), no exclamation marks. I am their bookkeeper, not their accountant, so do not promise anything about their taxes.

```

The categorization rules. Paste 20 to 50 bank descriptions with the amounts and the payee names left in, the client's name and account numbers taken out, and the chart of accounts:

```

Here are 40 bank descriptions from a landscaping company, with a chart of accounts below them. Draft one rule per recurring payee: the text to match, the account from the chart, and a confidence of high, medium or low. Never use an account that is not in the chart. Put every medium and low line in a separate list for me to decide, with one sentence on why it is unclear.

```

Then run the same prompt in a second model and read only the lines where the two disagree. A payee both models file under Supplies is a rule you approve in one look; a payee one files under Equipment and the other under Repairs and Maintenance is the one to open, and it is usually the one where the amount or the vendor could go either way. Whizi's compare view sends one prompt to two models in two columns with separate contexts; in any other app, open a fresh chat so the second model never sees the first answer. The disagreement list is your review queue. Where Whizi loses on this page: it has no ledger integration, so nothing you draft in it lands in QuickBooks or Xero, and there is no QuickBooks app inside it the way ChatGPT and Claude have had one since 2026.

The report explanation. Owners read four sentences and skip a page:

```

Here is a profit and loss for Q1 with the client's name removed. Write four sentences for the owner of a two-truck plumbing business: what came in, what went out, the one line that moved most against Q4, and one question I should ask them about it. Plain words, no jargon, nothing about tax.

```

Check every number in the four sentences against the report before it goes out, including the rounding ($48,212 is not "about $48,000" on a client's page) and which quarter each figure belongs to.

The spreadsheet cleanup. A bank export with mixed date formats, trailing spaces and "Last, First" names is the same job as any other messy sheet: paste it with the account column deleted, ask for the cleaned table and a change log with one line per change, and count the rows in against the rows out. How to use AI to clean up a spreadsheet has the full prompt and the row limits per tool.

The reconciliation check. The model does not reconcile; it finds where to look:

```

Bank balance $14,212.40, book balance $13,987.15. Below are the unmatched items from each side. Find the combination of items that explains the $225.25 difference. Flag any transposition (a difference divisible by 9 is the classic sign), any item that appears on one side twice, and any amount that matches an item on the other side within a few cents. Say which side each fix belongs on. Do not post anything; tell me what to open.

```

The divisible-by-9 rule is the oldest trick in the trade, and the prompt names it so the check is asked for rather than hoped for. Check the arithmetic yourself once the unmatched list runs past 30 items: ask for a table of candidate combinations with the sums shown, and add the sums yourself. The accounting prompts pack has 20 more, from a month end close checklist to a duplicate check on a vendor payment run.

What should you never paste into ChatGPT from a client's books?

Never paste a client's full account or routing numbers, card numbers, Social Security or EIN numbers, login details, or an unredacted bank statement. A model needs none of them for the five jobs above: "Chase ...4471" works as well as the full number, "Client A" as well as the name, and a bank export with the account column deleted reads exactly the same.

The reason is retention and training defaults, not hacking. On a consumer ChatGPT plan, OpenAI's pricing page lists "Content is used to train our models" for Free, Go, Plus and Pro with an opt-out, as of September 2026, and every vendor keeps a pasted line on its own schedule. What never to paste into an AI chat walks through where a pasted line goes at each vendor and the thirty-second redaction habit that makes the rest safe. The bookkeeper's version of that habit: the model gets descriptions, amounts and dates, and the ledger keeps the identifiers.

The records themselves stay in the ledger and the file, never in a chat transcript. The IRS states the rule for how long: "Generally, you must keep your records that support an item of income, deduction or credit shown on your tax return until the period of limitations for that tax return runs out." On the IRS's own page, updated 30 June 2026, that is 3 years in the ordinary case, 6 years where unreported income is more than 25% of gross income, and at least 4 years for employment tax records. A chat reply is not the supporting document; the statement, the invoice and the reconciliation report you saved from the ledger are, so the workflow that survives an audit is the one where the model reads a redacted copy and the original never moves.

One more line for anyone whose clients ask: none of this is tax advice, and neither is anything a chat model says about a deduction. A bookkeeper who wants a model's view on a tax question sends the answer to the client's CPA, not to the client.

Is AI taking away bookkeeping jobs?

The entry work, yes; the judgment, not yet. The US Bureau of Labor Statistics projects employment of bookkeeping, accounting and auditing clerks to decline 6 percent from 2025 to 2035, and still expects about 144,100 openings a year over the decade, at a 2025 median pay of $50,670 (BLS Occupational Outlook Handbook, updated 27 August 2026). The decline lands on the categorization and matching that Xero, QuickBooks, Zoho and Booke now sell as a feature on a $19 to $140 plan.

The 98% of accountants in Intuit's 2024 survey who had used AI were not replaced by it; 69% used it for data entry, which is the part of the job the ledger apps were already automating. What is left is the table at the top of this page: reading the client, drafting the rule, explaining the report, catching the $225.25. A chat subscription does the drafting half of those for $20 a month and none of the posting.

So the rule: if the task is words about the numbers, ChatGPT drafts it and you check it; if the task changes a balance, it happens in the ledger, and the model only tells you where to look.