AI for accountants: 6 jobs it does well, the prompts, and the client data rules

What do accountants use AI for in 2026?

Document summaries rank third on the tax firm wish list in Thomson Reuters' 2026 survey, at 59%, just behind research and bookkeeping. Two rows of this table come straight from that list. We picked the other four because a model can do them from material you paste, and nothing gets filed on its word. Here's the tool that fits each one as of October 2026, and the check you run before anything leaves the firm:

Thomson Reuters asked tax firms which potential uses of AI interest them, and research came first at 69%, ahead of bookkeeping at 60% and document summaries at 59% (Thomson Reuters, June 2026). Interest runs ahead of use. Only a third of tax firms use generative AI in their work, and 14% use agentic AI. Elizabeth Beastrom, who heads the company's tax, audit and accounting business, put the firm-level figure at 34% in April 2026, up from 21% a year before.

Bookkeeping, the second pick, isn't a row in the table. We left it out because coding transactions is a job for the ledger software itself, and that's a separate buying decision from a chat tool. For a ledger extract you've exported, the accounting prompts pack has prompts that categorize transactions with a confidence flag and flag unusual entries for review.

The person is ahead of the firm. In Intuit's survey of 700 US accounting professionals in April 2025, 46% said they use AI every day (Intuit QuickBooks). Canada shows the same habit. Statistics Canada found that 52% of workers in professional, scientific and technical services, the group accounting firms sit in, had used generative AI at work in the previous 12 months (Statistics Canada, 17 June 2026). The figure averages survey rounds from September 2024 to July 2025.

The AICPA's own number is older and much lower. In its November 2024 Economic Outlook Survey of 273 CPA decision makers, 6% had put generative AI into a business function and 30% were experimenting. That gap, between what one accountant does in a personal chat tab and what the firm has approved, is exactly where the confidentiality rules below apply.

Can accountants put client data into ChatGPT?

Not into a consumer account that trains on your chats, and an AICPA member in public practice needs the client's specific consent before disclosing confidential client information. The AICPA rule is one sentence long: "A member in public practice shall not disclose any confidential client information without the specific consent of the client." Its Professional Ethics Division turned that into a habit in August 2026: "Before pasting client data into any AI system, stop and ask: Is this confidential client information?" Here's what each body and one US statute say, read on 11 October 2026:

The rules turn on where a pasted line goes, and the tools answer that differently. We read each vendor's own page on 11 October 2026:

The workflow that fits every row: client names become "Client A", identifiers come out, and the file goes only into a tool the firm has approved in writing. Whether a US preparer needs signed section 7216 consent before return data reaches a given AI vendor is a question for the firm's counsel, not for a chat model. The IRS's section 7216 page, reviewed 7 September 2026, still points to final regulations that took effect on 28 December 2012, a decade before ChatGPT. What never to paste into an AI chat walks through the redaction habit line by line.

Which AI tools do accountants pay for, and what do they cost?

A specialist tax research tool costs about six times a chat plan, and what the extra buys is a curated source base with citations. Blue J's individual plan is $1,498 a year, which is 6.2 times twelve months of ChatGPT Plus at $20. Prices and features below come from each vendor's US page, read 11 October 2026, in US dollars:

Read the last column first. Blue J and CCH AnswerConnect answer from an editorial database: Blue J says its sources are "Curated by tax experts", and its pricing page lists SOC 2 Type II. A general chat model searches the open web, and it can land on a firm's blog post as easily as on the statute. Intuit's AI lives inside ProConnect returns. Copilot lives inside the workbook you already have open, which makes it the fastest route to a formula and the slowest to a citation.

Whizi's Canadian CPA agent, in the Agents section of the model picker on Pro and above, sits between the two for Canadian work. It runs a web search on every turn, up to 7 searches, restricted to 42 official domains. They include canada.ca and the CRA, Justice Laws, the provincial tax and statute sites, CanLII and the federal courts, CPA Canada, FRAS Canada and the IFRS Foundation. Its instructions require a numbered Sources list at the end of every substantive answer. Where it loses: there's no ledger integration, so nothing it drafts lands in QuickBooks, Xero or a return. It's built for Canada only, and its domain list leaves out IRS.gov on purpose. And Whizi's privacy policy names no SOC 2 or ISO 27001 audit, where Blue J lists one.

Which AI prompts should accountants start with?

Start with the five prompts below, one per drafting job in the first table, each written so the model works only from what you paste and marks what it couldn't confirm. None of them asks for a tax position. That's yours to take, and a model states one with the same confidence whether it's right or wrong.

Tax research. Ask for the citation trail, not the answer:

```

I'm a CPA researching [issue] for the [2026] taxation year in [jurisdiction]. Find the governing section of the statute and any current administrative guidance from the tax authority. For each point give: the section or document number, the URL on the official site, the year the page applies to, and one sentence on what it says. Mark anything you could not confirm on an official page as UNCONFIRMED. Do not give a conclusion on any client's position.

```

The client email. It's the chase you've rewritten forty times:

```

Write a 100-word email to a client who still owes me their wage slips and two donation receipts for the 2025 return. One ask, one date (Friday 13 March), friendly, no exclamation marks. Do not promise anything about their refund or their taxes.

```

The memo. You bring the facts, the conclusion and the authorities. The model brings structure:

```

Draft a file memo from the facts, conclusion and authorities below. Sections: Facts, Issue, Authorities, Analysis, Conclusion, Open points. Use only the authorities I list and quote their section numbers exactly. If the analysis needs one I did not list, write [AUTHORITY NEEDED] instead of supplying it. The client is "Client A". Plain English, under 600 words.

Facts: [...]

My conclusion: [...]

Authorities: [...]

```

The formula. Copilot in Excel does this inside the workbook; in a chat model, describe the columns instead of pasting client rows:

```

Columns: A Date, B Vendor, C Amount, D GL code, E Month. Write one Excel formula for G2 that totals Amount for the GL code in F2 and the month in F1. Explain it one argument per line, then give me 3 test rows with the total I should see so I can check it by hand.

```

The document review. Ask for clause numbers so every point can be checked in seconds:

```

Below is a commercial lease with names and addresses removed. List the term and renewal options, base rent and escalation, who pays operating costs, any free-rent period, and termination rights. Quote the clause number for every point, and write NOT FOUND where the lease is silent. Do not suggest an accounting treatment.

```

Reconciliation help works the same way: paste the two balances and the unmatched items with account numbers removed, and ask which combination of items explains the difference. The accounting prompts pack has 20 more, from a month end close checklist to a variance note.

How do you check an AI tax answer before it goes to a client?

Open every source the answer cites and check three things: the page exists, it says what the answer claims, and it applies to your year and your jurisdiction. Tax professionals already rank wrong answers as the top risk. In Thomson Reuters' 2026 survey, 76.4% of tax respondents named inaccurate responses as a concern, and 70.5% named the privacy of what goes into the tool.

1. Open each citation. A section number with no working link is unconfirmed until you find it yourself.
2. Check the year on the source page, not the year in the answer. Rates and limits change every January.
3. Check the jurisdiction. A US rule and a Canadian one can look alike and differ on every number.
4. Write the conclusion in your own words, from the source, in the memo. The model's sentence never goes to the client.
5. Keep the source in the file, not the chat.

The AICPA's ethics staff put it plainly in August 2026: "the responsibility for the final work product rests with you." CPA Ontario said the same thing two years earlier. Models do invent sources, and AI hallucination statistics has the measured rates by model. This page describes a workflow. It takes no view on any tax position, and a model's answer to a tax question is a lead to check, never the answer you send.