AI for Accountants: Eight Workflows, Copyable Prompts and the Guardrails Behind Them

Most of an accountant’s week is not spent on judgment calls. It goes on the work around them: chasing documents, writing up variances, explaining a notice to a nervous client, turning a messy trial balance into something a partner can review. That surrounding work is where general-purpose AI tools earn their place, and it is also where the confidentiality rules bite hardest, because the raw material is client data.

This guide sets out eight workflows that suit AI well, each with a prompt that can be copied and adapted. It then covers what the IRS, the FTC and the AICPA say about putting client information into these tools, and how a small firm can turn that into a one-page policy. Rules and vendor terms were checked on November 9, 2026.

Where AI fits in an accounting week

AI tools are good at language and structure: summarizing, drafting, reformatting, spotting gaps in a list. They are unreliable at arithmetic on long tables and at citing law, because they can produce confident but wrong output. NIST’s generative AI risk profile (July 26, 2024) calls this “confabulation” and lists it among the core risks of the technology. The practical split looks like this:

Every prompt below is a starting point. Output varies by tool and model version, and the accountant who signs off on the work is responsible for checking it.

The eight workflows

1. Turn an engagement scope into a client document request list

Document chasing is repetitive and easy to template. The AI only needs the engagement type and the client’s general situation, not their figures.

You are helping an accounting firm prepare a document request list.

Engagement: [ENGAGEMENT TYPE, e.g. 2026 individual return with Schedule C and one rental property]
Client situation: [GENERAL FACTS WITH NO NAMES OR ID NUMBERS, e.g. sole proprietor, moved states mid-year]
Prior-year items to carry forward: [LIST]

Produce a checklist grouped by category (income, deductions, property, estimates paid, other).
For each item give a one-line plain-English explanation the client will understand.
Flag any item that depends on facts I haven't given you, and list the questions I should ask.

2. Draft month-end variance commentary

Flux analysis needs numbers, but it rarely needs the client’s identity. Many firms export account names and balances with the entity name removed.

Below is a month-over-month comparison for a [INDUSTRY] business. Entity names are removed.

[PASTE: account name | prior month | current month | change | change %]

Materiality threshold: changes over [AMOUNT] or [PERCENT].
For each account over the threshold:
1. Write a neutral one-sentence description of the change.
2. List two or three possible causes that someone should confirm with the client.
3. Do not state a cause as fact.
Return a table, then a short list of questions for the client.

The second instruction matters. The tool has no idea why rent went up; it can only list possibilities for someone to confirm.

3. Write reconciliation exception notes

Once a reconciliation is done in the ledger, the leftover items need a clear note for the reviewer.

I have finished a bank reconciliation. These items remain unreconciled (no account numbers included):

[PASTE: date | description | amount | side (book or bank)]

Group them by likely type (timing, missing entry, possible duplicate, unknown).
Write one short reviewer note per group saying what I should check next.
Keep the language factual. Do not invent explanations for individual items.

4. Explain an IRS or state notice to a client in plain English

Clients panic at notices. A plain-language first draft saves time, as long as the identifying details stay out.

Rewrite the following explanation for a client with no tax background.

Notice type: [e.g. CP2000 proposed adjustment]
What the notice says, in my words: [SUMMARY WITHOUT NAMES, SSNs OR AMOUNTS THAT IDENTIFY THE CLIENT]
What we plan to do: [NEXT STEPS AND DEADLINE]

Keep it under 200 words. Use short sentences. End with what the client needs to send us and by when.
Do not promise an outcome.

5. Build and explain spreadsheet checks

AI tools handle formula syntax well, and the risk is low when the data stays in the spreadsheet and only the structure goes into the prompt.

I have an Excel sheet with these columns: [COLUMN LETTERS AND HEADERS].
I need a formula that [GOAL, e.g. flags any invoice where the tax amount isn't 8.25% of the net amount, within 1 cent].
Give me the formula for row 2, explain each part in one line, and suggest two test cases with expected results so I can confirm it works.

6. Get a research starting map, then go to the source

AI is useful for working out which questions to ask and which code sections or standards might apply. It is not a citation source.

I'm researching this question: [QUESTION, e.g. treatment of a client's home office when they also rent part of the home].

Give me:
- the issues I should consider, as a numbered list;
- the code sections, regulations or standards that may apply, labeled "to verify";
- the facts I need from the client to answer it.

Do not quote any rule text. I will read the primary sources myself.

The “to verify” label is the point. Anything the tool names has to be found and read in the actual code, regulation or standard before it goes into a memo.

7. Draft a close checklist or procedure

Procedures are low-risk and high-value: no client data, lots of structure.

Draft a month-end close checklist for a [SIZE] firm that closes books for [NUMBER] small-business clients on [ACCOUNTING SOFTWARE].
Include: bank and card reconciliations, AP and AR review, accruals, payroll tie-out, sales tax, review sign-off.
For each step give: owner role, due day (business day 1 to 10), and the evidence the reviewer should see.
Format as a table.

8. Prepare a client meeting agenda from typed notes

When the accountant’s own typed notes from earlier meetings are stripped of identifiers, the tool can turn them into an agenda and a follow-up list.

Here are my notes from previous conversations with a client (identifiers removed):
[NOTES]

Build a 30-minute meeting agenda with time blocks.
Then list open items, who owns each, and what decision is needed.
Point out anything in my notes that looks unresolved or contradictory.

A longer set of prompts, sorted by task, is in ChatGPT prompts for accountants.

What the rules say about client data

Don’t paste this in

Tax return information. Section 7216 makes it a crime for return preparers to disclose or use it without an exception or the taxpayer’s signed consent. The IRS Section 7216 FAQ gives the penalties: a fine of up to $1,000, up to a year in prison, or both for knowing or reckless violations, and a $250 civil penalty under Section 6713 that needs no intent. SSNs, EINs, bank and card numbers, payroll records, and anything else covered by the firm’s written information security plan should also stay out. Neither should any client file that the firm hasn’t approved the tool to receive.

The rules that govern this come from three directions.

Section 7216. The IRS FAQ (which the IRS now labels as historical, though the regulations it describes are still in force) says tax return information is read broadly. It covers information preparers get from taxpayers or other sources in connection with a return. Consent must be signed and dated, it must follow the format set out in Rev. Proc. 2013-14, and it lasts one year unless a different period is stated. Disclosing information to a preparer outside the United States always needs consent. Whether sending return data to an AI vendor counts as a permitted disclosure to a service provider is not settled in the IRS materials. Practitioner commentary leans toward treating it as a disclosure that needs consent unless an exception clearly applies.

The FTC Safeguards Rule. IRS Publication 5708 states that under the Gramm-Leach-Bliley Act and the Safeguards Rule, tax and accounting professionals count as financial institutions whatever their size. They must keep a written information security plan (WISP). One requirement is to select service providers that can maintain appropriate safeguards and to oversee how they handle customer information. An AI vendor that receives client data is a service provider, so it belongs in the WISP.

The AICPA Code. Members in public practice are bound by the Confidential Client Information Rule (ET sec. 1.700.001). A Journal of Accountancy article from August 2026, written by AICPA ethics staff, suggests three questions before pasting anything in. Is this confidential client information? Is consent required, and has the client specifically consented? Could this use lead to disclosure to third parties? A follow-up piece dated November 2, 2026, adds a point that matters for every workflow above: redaction does not, on its own, turn confidential client information into nonconfidential information. Removing names lowers the risk. It doesn’t remove the duty.

IRS Office of Professional Responsibility. On June 24, 2026, OPR issued introductory guidelines on AI in federal tax practice (Issue 2026-19). The guidelines create no new rules. They map existing Circular 230 duties onto AI use: due diligence in reviewing output (section 10.22), competence (10.35), firm procedures and training (10.36), verified facts in written advice (10.37), and confidentiality of return information (10.51(a)(15) with sections 6713 and 7216). On fees (10.27), the bulletin warns against billing for time not actually spent.

Picking the right tier of tool

The confidentiality analysis changes depending on which version of a tool the firm uses, so this is a settings question as much as a software question.

A business tier with a signed data processing agreement is a better fit for client work than a personal login. It doesn’t answer the Section 7216 consent question by itself, though.

A one-page firm AI policy

Firms don’t need a long document. They need clear answers that staff can apply in the moment. The AICPA published a small firm generative AI policy template on April 29, 2026 (members only). Whether or not a firm uses it, the policy should cover:

1. Approved tools. Name them, with the account tier (for example, the firm’s business workspace, not personal accounts).
2. Data classes. List what may go in (procedures, anonymized structures, public information), what needs client consent first, and what never goes in.
3. Consent. Say when a Section 7216 consent is required and who drafts it. Include the reminder that, under Treas. Reg. 301.7216-3, making services conditional on consent generally makes the consent involuntary.
4. Review. Assign who checks AI-assisted work, and record that the check happened.
5. Billing. Bill for time actually spent, in line with Circular 230 section 10.27 and the OPR bulletin.
6. WISP update. List the AI vendor as a service provider and note how its safeguards were assessed.
7. Training. Cover how often staff are trained and who answers questions.

A review checklist for AI-assisted work

Before anything drafted with AI leaves the firm, the reviewer can run through:

• Every figure traced to the ledger, return or source document, not to the AI output.
• Every code section, regulation or standard read in the original.
• No client identifiers in the prompt history of a tool that isn’t approved for them.
• Tone checked for anything that reads as a promise or guarantee.
• The file notes that AI was used, where firm policy requires it.

Common questions

Is it safe to paste a trial balance into a chatbot? It depends on the account tier, the vendor’s terms and whether the data identifies the client. A trial balance with the entity name removed is still confidential client information under the AICPA Code. The safer pattern is a firm-approved business tool, covered in the WISP, with client consent where the firm’s policy requires it.

Does every tax client need to sign an AI consent? The IRS hasn’t issued AI-specific Section 7216 guidance. Consent is needed for disclosures and uses the regulations don’t specifically permit, and for any disclosure outside the United States. Firms that send return information to an AI vendor should have their position reviewed by a tax attorney and, where consent is needed, use a form that follows Rev. Proc. 2013-14.

Can AI prepare a return? Tax software already automates calculation. General chatbots should not be treated as preparers, and OPR’s guidelines say practitioners must not rely solely on AI output.

Do clients need to be told? No federal rule requires a general disclosure of AI use. Consent rules apply where client data is disclosed, though, and some engagement terms may require more. Many firms add a short description of their AI practices to engagement letters.