ChatGPT for finance and accounting: what to use it for, and what to check

Attach last month's profit and loss and ask ChatGPT for variance commentary, and two minutes later you'll have tidy paragraphs in the house style. Ask it why marketing spend overran, and the reason it gives will be just as tidy, whether or not the ledger holds one.

That split runs through all of finance and accounting. ChatGPT is good at work done on material you supply: commentary from a report, formulas from a description, a comparison of two exports, a plain summary of a standard you attach. It is unreliable on anything you didn't supply. A balance it has to estimate, a tax rate from memory and the cause of a variance all come back sounding equally sure. Set it up for the first kind of job, and check everything in the second.

What the plans cost, from OpenAI's pricing page (learn.chatgpt.com, checked 22 October 2026):

For anything holding customer names, salaries or unpublished results, we'd use Business rather than personal accounts, mainly for that last line. If you're still deciding between assistants, Claude vs ChatGPT covers the trade-offs.

ChatGPT changed a lot in 2026, and plenty of guides to ChatGPT for accounting still describe features that have gone. Here's what we'd use for finance work now.

A note on names. Since 9 July 2026, what ChatGPT used to call connectors are apps, and apps come packaged in plugins, listed in the Plugin Directory. Older guides that send you to "Settings > Connectors" or the "app directory" mean the same place.

Two things to stop recommending:

• Agent mode is gone. OpenAI's help centre now says "ChatGPT agent is no longer available" and points to ChatGPT Work.
• Custom GPTs are being retired. Free, Go, Plus and Pro accounts can no longer create new GPTs, and OpenAI is moving GPTs to plugins. What used to be a "month-end GPT" is now a project plus a skill.

Most systems a finance team runs now have an official ChatGPT app, published by the vendor itself. Where they differ most is in what they can change, and for finance that matters more than what they can read.

Anything else, an in-house system or a data warehouse, say, can be added as a custom MCP server from Plugins. OpenAI's developer docs say the server must be reachable at a public HTTPS address or through its Secure MCP Tunnel. Its help centre says full MCP, write actions included, is for Business and Enterprise, while Pro can connect custom servers for reading only. Whichever route you take, an app sees what the account you signed in with can see. A total from it covers that account's view of the books, nothing more.

ChatGPT for Excel is OpenAI's add-in for Excel, generally available since 5 May 2026. OpenAI's help centre lists it on every plan, from Free and Go up to Business and Enterprise, with a matching add-in for Google Sheets. For most finance people it's the first place ChatGPT earns its keep, because formulas are where it saves the most time.

Formulas

Describe the result you want, name the ranges, and say what should happen at the edges. Then ask for an explanation every time. A wrong reference rarely shows in the formula, but it shows quickly in the explanation.

In the Actuals tab, column F should show the variance against Budget!C:C
for the same account code in column A, as actual minus budget.
Match on the account code with XLOOKUP, not on row position.
If the code isn't in the budget, return "NO BUDGET" rather than 0.
Explain each argument, then test the formula on rows 5, 18 and the
last row, and show me the three results.

The same pattern works for an inherited formula ("walk me through D42, one reference at a time") and for an error ("why does this return #N/A on row 31?").

What to check in a workbook ChatGPT changed

An edit can look right and still break a model. Before you rely on it:

1. Save a copy first, so you can compare the two versions.
2. Turn on Formulas > Show Formulas and look for numbers typed where a formula used to be. A hard-coded value is the most common silent break.
3. Check every range still reaches the last row. A SUM that stops at row 120 ignores everything added below it.
4. Trace precedents on the key outputs. Revenue, cash and EBITDA should still come from the input sheets, not from a cell the edit added.
5. Recheck signs and units. A cost entered as positive in one tab and negative in another doubles a variance instead of cancelling it.
6. Tie the totals back to the trial balance or the export you started from.

Be careful with workbooks from outside the company, too: a file can carry text written for the assistant rather than for you. Claude for Excel covers the same checks for Anthropic's add-in, which answers with links to the cells it used.

Each prompt is written for one feature. Swap in your own details.

Project instructions for the finance team

You help [Company]'s finance team. Read conventions.md in this project
before every answer: our basis, financial year, currency and mappings.
Use only figures from connected apps or attached files. If a figure
isn't there, write [NO DATA]. Never estimate a balance, rate or benchmark.
Do calculations in code and show it. Print the period, basis and
currency above every table.
Explain a variance only with an account, supplier or customer in the
data; otherwise write "Cause not in the data".
Don't create, send or delete anything in QuickBooks, Xero or Stripe
unless my message asks for that exact change.

The [NO DATA] line does the most work. Without it, a gap in the export gets filled with a number that looks about right.

An accounting question in deep research

How should a company reporting under [IFRS 15 / ASC 606] treat
[a three-year software licence sold with implementation services]?
Use only ifrs.org or fasb.org and the published guides of the four
largest audit firms. For every point, give the URL, the paragraph
reference and the sentence that supports it.
List any judgement the standard leaves to the company separately.
Output: the treatment in 300 words, then the judgements as a list.

Read the research plan before you approve it, and use Manage sites to hold the search to those domains. Treat the result as a brief for the conversation with your auditor, not as your accounting policy.

Journals drafted, not booked

Read the attached lease agreement. Propose the journal entries for the
first month, with the account, debit, credit and amount for each line.
Next to every amount, quote the clause it comes from.
Where an entry depends on a rate or estimate the contract doesn't give,
write [JUDGEMENT] and say what's needed. Don't post anything.

Period-end numbers get full workflows of their own below. More analysis prompts are in prompts for financial analysis.

We've written the full finance setup for Claude in Claude for finance: the conventions your numbers depend on, a shared project, the connectors and five workflows. Most of it works in ChatGPT with small changes. Rather than repeat it, here's what to take and what to change.

The three workflows below use the same prompts as the Claude guide, word for word. Only what you connect or attach changes. Where a prompt mentions "the skill", that's the finance-definitions skill above; where it says Xero or QuickBooks, use whichever app you've added.

The report most teams rebuild by hand every close. With the Xero or QuickBooks app added, ChatGPT runs the profit and loss and balance sheet, lays the month against the last one with the basis printed at the top, and checks that last month's figures still match the pack you sent. What you get is a report you could send, plus the report settings and code behind every number. For the analysis that follows, see prompts for financial analysis.

Variance commentary is where AI saves finance the most time and does the most quiet damage, because an invented reason reads exactly like a real one. This workflow reads actuals by department from the ledger, reads the approved budget from your workbook, and keeps every explanation tied to an account, supplier or customer it can point to. QuickBooks keeps budgets on its Plus and Advanced plans (QuickBooks Help); most teams keep theirs in a workbook either way. FP&A teams building the forecast itself should see AI for FP&A.

Matching Stripe payouts to the bank is slow because a payout isn't a sale: it's a batch of charges, refunds, disputes and fees, dated by when Stripe expects it to arrive. With Stripe's app added, ChatGPT reads each payout's balance transactions, matches the payouts to your bank statement export, and checks the fees against your ledger's Stripe clearing account. Anything it can't match stays on its own line.

ChatGPT's finance mistakes are rarely absurd. They're plausible, which is worse, because plausible numbers get pasted into packs.

The evidence on citations is the part to take seriously. In a 2026 study of 14 models, the best produced working links more than 94% of the time, yet only 39 to 77% of their claims were accurate (Onweller et al., Cited but Not Verified, May 2026). When Columbia's Tow Center tested eight AI search tools on finding the source of news excerpts, ChatGPT Search got 134 of 200 wrong and never once declined to answer (Tow Center, March 2025). In finance, the link that opens is the ledger report. The claim that fails is the reason given for the number in it.

The failures we'd watch for:

• A tidy reason with nothing under it. "Delayed renewals" and "seasonal slowdown" are what commentary usually says. If no transaction shows it, it's a guess.
• A variance sign read the wrong way. Spending £9,300 more than budget is a positive number and an unfavourable variance. Ask for the words, not only the sign.
• Netting differences away. A payout in transit and a missing fee can cancel to zero. That isn't reconciled.
• MRR quoted as revenue. Billing tools normalise annual plans and leave out one-off charges; recognised revenue comes from the ledger.
• Rates and thresholds from memory. VAT thresholds, filing deadlines and tax rates change. Ask for the official page and its date.
• A change nobody meant. QuickBooks' app can send and delete invoices, and Stripe's can issue refunds once someone confirms. Approve each change yourself.

A two-minute check before ChatGPT's numbers leave finance

1. Search the answer for figures. Each one should trace to a connected report or an attached file.
2. Ask: "Which figures and reasons in this answer aren't in a report you ran or a file I attached?"
3. Recompute one total from the source report yourself.
4. Ask which basis, period and currency each table uses, and compare them with the project instructions.
5. For any cited standard or tax rule, open the source and find the paragraph.

Step two is where most problems surface, and the one Count is built for: it labels every claim in an answer as a citation you can open, an assumption, model knowledge or a hallucination, with the reasoning chain behind it. It works in Claude today, not yet in ChatGPT; the buttons on this page tell us you want it there.

ChatGPT is a capable finance tool, and if your company already runs ChatGPT Business, the setup on this page is the one to use. Switching assistants costs more than any difference between them. If you're choosing from scratch, we'd set finance up in Claude, for reasons that come down to checking a number in a workbook.

The Excel row decides it for us. In finance, the question after "what's the number?" is always "where did it come from?", and Anthropic's add-in answers with links to the cells behind each answer (Claude Docs). Someone who didn't build the model can follow those links and check it. Anthropic also publishes finance agent templates, including a month-end closer and a general ledger reconciler, and Count's audit labels every claim in Claude's answers, which ChatGPT doesn't offer yet.

ChatGPT is the better pick in two cases. If most of the team works on Free or Go and lives in Excel, its add-in needs no paid plan. And if you run NetSuite, Oracle's app saves you setting up a custom connector. Either way, the checking step above doesn't change. The step-by-step Claude setup is in Claude for finance.

How do I use ChatGPT for accounting?

Can ChatGPT connect to QuickBooks or Xero?

Is ChatGPT for Excel free?

Does ChatGPT train on my financial data?

Can ChatGPT replace an accountant?

Should finance teams use ChatGPT or Claude?