Blog /Accounting

ChatGPT and Copilot for Accounting: What They Can and Can't Do

Anirudh Badam

Anirudh Badam

Co-Founder and CAIO, Adopt AI

11 November 2026

A smartphone showing the ChatGPT app resting on a laptop keyboard

Your controller already uses ChatGPT. So does the staff accountant two desks over, and someone on your team has Copilot open inside Excel right now. They are not wrong to. These tools are good at a real slice of accounting work, and a finance leader who bans them loses that slice for nothing.

The better question is where the slice ends. This post lays out what ChatGPT for accounting does well, what Copilot adds inside Microsoft 365, and the jobs neither one can run for a finance team.

What ChatGPT does well in accounting

A general-purpose model earns its place on tasks that start and finish inside one conversation.

  • Explaining a standard. Ask how ASC 842 treats a lease modification and you get a readable summary in seconds. Verify it against the source before it reaches a memo.
  • Drafting prose. Variance commentary, an email to a vendor about a duplicate invoice, a first pass at a policy document.
  • Writing formulas and macros. A nested lookup across two tabs, a pivot layout, a script that reformats a bank export.
  • Working through a messy problem. Paste in a confusing intercompany balance and ask what could produce it. The list of hypotheses is often useful.

For one person doing one task, this saves real time. A senior accountant who writes a flux explanation in ten minutes instead of thirty has gained something.

What Copilot adds

Microsoft Copilot sits inside the tools your team already opens. In Excel it can suggest formulas, summarize a sheet, and build a chart from a range. In Outlook and Word it drafts and summarizes. For a company already on Microsoft 365, the licensing question is simple and the learning curve is short.

Copilot works on the files and mailboxes the user can already reach. It does not change what is in those files, and it does not connect to your general ledger, your bank portals, or your AP queue on its own.

Where both stop

The limits show up when the work has to repeat, span systems, or survive an audit.

1. They work one chat at a time. A reconciliation for 12 bank accounts across four entities is 48 jobs, every month. A chat window handles one at a time, and a person has to start each one, paste the data, and check the result. That is still manual work with a faster typist.

2. They do not hold the source. A bank rec needs the statement from the bank, the ledger detail from the ERP, and the list of reconciling items from last month. ChatGPT and Copilot see what a person pastes or attaches. They do not pull the statement themselves or confirm that the file is complete.

3. They are not consistent run to run. Ask the same question twice with different wording and the answer can change. Finance teams that use these tools on close tasks report the same complaint: results are not consistent each time. For drafting, that is tolerable. For a tie-out that an auditor will test, it is not.

4. They hold no review gate. Accounting depends on a preparer and a reviewer, with the work recorded for both. A chat transcript does not show who approved what, and it does not stop an unreviewed number from reaching the books.

5. They leave no trace your auditors accept. Ask how a figure was produced and the answer is a paste buffer and a conversation that lives in one person's account. An audit trail needs source documents, computations, and a named approver.

6. Data handling needs a policy. Pasting a trial balance or a vendor list into a consumer chat tool raises questions about what leaves your environment. Our post on AI accounting data privacy covers what to ask.

A simple test for any task

Before you hand a job to a chat tool, ask four questions:

  1. Does the job start and finish in one sitting?
  2. Is one person responsible for it?
  3. Is the output a draft that a human will rewrite?
  4. Would it matter if the answer came out slightly different tomorrow?

Four yeses mean the task fits ChatGPT or Copilot. A no on any of them means you are describing a process, and a process needs more than a conversation.

What the process layer looks like

The close, the reconciliations, AP, and AR all run on a schedule, across entities, with data from systems of record. A tool for that work has to do what a chat window cannot:

  • Pull data from your ERP, banks, and documents without a person pasting it
  • Run the same steps on every entity, every period
  • Log every step with its source
  • Stop at a review gate wherever you want one
  • Leave your ERP as the system of record

Adopt is built for that layer. Agents prepare the reconciliations, the tie-outs, and the variance write-ups. A named reviewer approves before anything posts. Your team can keep using ChatGPT and Copilot for drafting and formulas, and run the repeatable work through agents that keep a record. For a deeper look at the buying questions, see our guide to evaluating AI finance tools.

How to decide this quarter

Take the three tasks that consume the most hours on your close calendar. Run each through the four-question test above. The ones that pass can go to ChatGPT or Copilot today, with a policy on what data goes in. The ones that fail are your candidates for a workflow with agents and review gates.

If you want to see one of those on your own data, bring us your messiest reconciliation.