Where AI fits in tax, audit, and client advisory work

In tax, audit, and client advisory work, AI fits the reading and drafting steps on either side of a professional judgment, and the judgment and the signature stay with a licensed person. In each service line there are a handful of steps where a source document acts as the answer key and a qualified person already reviews the result, and those are the steps to hand over first.

Most published lists of accounting use cases are organized by software category: a tax research tool, an audit platform, a bookkeeping automation product. That is a useful way to shop and a poor way to decide what to change inside a practice. A partner or senior manager gets further by listing the steps of their own engagement in order and writing down, next to each one, who prepares it, who reviews it, and who signs.

Sort the work by step and by who signs

Here is the map we would draw for a regional firm. The sections below explain each row.

Service line	Steps AI can draft well	Stays with a licensed person	What the reviewer checks
Tax	Sorting client documents, tying source documents to the organizer, the open items list, the letter that explains the return	Positions, elections, estimates, the return itself, the signature	Each extracted figure against the source page, and the open items list against the prior-year file
Audit and attest	Summaries of contracts and minutes for the file, documentation completeness against the firm's checklist, footnote amounts compared to audited amounts	Risk assessment, materiality, the procedures, the conclusions, the opinion	Each summary and each flagged difference against the underlying document, with sign-off unchanged
Client advisory	Variance commentary, a first pass at the monthly package narrative, the meeting brief	The advice, the recommendations, the client conversation	Every number in the narrative against the closed books, and whether the explanation is the true one

Tax: the hours before and after preparation

In our view the time in a tax practice is lost before preparation starts and after it ends. Before, someone chases the client for documents, sorts what arrives, and works out what is still missing. After, someone writes the letter that tells the client what changed from last year and what to do next.

These steps are document reading and drafting tasks with a clear answer key:

• Intake and sorting. Identifying each uploaded document, naming it to the firm's convention, and filing it to the right place in the binder.
• Source-to-organizer tie-out. Comparing what arrived to the organizer and to last year's return, and flagging what is missing or new.
• The open items list. A draft request to the client, written in the firm's wording, that lists exactly what is outstanding.
• The client letter. A first draft, in the firm's template, that explains the result and the year-over-year changes using figures from the finished return.

A workflow built on the firm's own prior-year files is a sensible place to start, because the prior-year file shows what a complete file for this client looks like.

What stays with the preparer and reviewer

Positions, elections, estimates, and anything that requires knowing the client's facts beyond what is on the page stay with the preparer. The review and the signature stay where they are.

The professional standard is explicit on this. A Tax Adviser article from April 2026 on the AICPA's Statements on Standards for Tax Services quotes Section 1.4: members must exercise due professional care when relying on an electronic tool, and "the use of such a tool does not absolve the member of professional responsibilities." The same article quotes Section 1.3, under which a member should make reasonable efforts to safeguard taxpayer data. Which AI plan client documents may go into is a separate question from which tasks are suitable, and we cover it in our piece on the FTC Safeguards Rule, IRS Publication 4557, and Section 7216. Confirm both with your own compliance lead.

Tax research deserves a note of its own. A research answer is only useful if the reviewer can read the authority behind it. In the Thomson Reuters Future of Professionals 2026 report for tax and accounting, which surveyed professionals at tax and audit firms, 94 percent said an acceptable AI tool must ground its outputs in authoritative content, and 90 percent said it must produce reasoning that can be explained and defended. The practical rule follows from that: cite the authority after someone has read it, and never cite the tool.

Audit and attest: what the PCAOB staff has observed so far

Audit calls for the most caution of the three. In July 2024 the PCAOB staff published observations from outreach on generative AI in audits, based on conversations with auditors mainly at larger firms. The document states that it represents staff views and is not a rule. Its findings are still the best public record of what firms were doing:

• Use at the firms interviewed appeared to be focused primarily on administrative and research activities, such as initial drafts of memos and presentations and tools for researching internal accounting and auditing guidance.
• Some firms did not allow generative AI when performing audit or attest procedures, citing data privacy and concerns about the reliability of the output.
• Firms investing in these tools expected them to augment people and said human involvement remains essential to review the output.
• Some firms said their supervision and review policies had not changed: the team member who uses the tool is still responsible for the work and its documentation, and the supervisor reviews it with the same diligence as any other work.

That outreach is more than two years old and covered firms that audit public companies, so a regional firm should read it as a description of a careful starting point and check its own methodology and policy.

Summaries, completeness checks, and tie-outs that leave sign-off unchanged

The same PCAOB document lists areas the firms saw as possible future uses, including summarizing accounting policy and legal documents, evaluating the completeness of audit documentation against documentation requirements, and comparing amounts in the financial statements or notes with audited amounts. Three of those translate into work a firm can build without changing who performs or signs a procedure:

• Document summaries for the file. A structured summary of a lease, a debt agreement, or board minutes, with page references, that the auditor reads alongside the document.
• Completeness checks. A comparison of the engagement file to the firm's own checklist that lists what is missing before the manager's review.
• Footnote tie-outs. A comparison of every amount in the draft notes to the audited trial balance and workpapers, with differences listed for a person to resolve.

In each case the output is a list for a person to act on, and the preparer and reviewer sign-off stays exactly as the firm's methodology defines it.

Client advisory: preparation, so the hour goes to advice

The AICPA and CPA.com 2024 CAS Benchmark Survey, which collected 2023 data from 206 practices, reported median growth of 17 percent and described CAS as the fastest growing service area in public accounting. A practice growing at that rate has to find advisor hours somewhere.

The gain from AI here is preparation. Once the month is closed, a workflow can draft the variance commentary from the client's own books, write a first pass of the package narrative in the firm's format, and assemble a one-page brief for the meeting: what moved, what the client asked last month, and what is worth raising. The advisor edits the draft, checks the numbers, and spends the meeting advising.

Transaction coding and reconciliations are the other place firms look. CPA.com's 2025 AI in Accounting Report says vendors report time savings of 30 to 70 percent on common tasks such as bank reconciliations, transaction coding, and month-end close. That is a vendor-reported figure, so test it on your own clients before planning around it. The same report says the need for human review remains paramount in high-stakes environments like audit and tax, with accountants acting as reviewers.

A three-question test for any task

Before handing a step to AI, ask three questions about it.

1. Is there a source document to check against? Examples are a W-2, a signed lease, or a closed trial balance. If the output cannot be compared to something, the reviewer has nothing to review.
2. Does a qualified person already review this step? If the step has a reviewer today, the workflow keeps that reviewer. If it does not, add one before adding AI.
3. Would a wrong answer be caught before it reaches the client? A mislabeled document in the binder gets caught at preparation. An email that goes to a client without review is not caught by anyone.

A task that passes all three is a good candidate. A task that fails any one of them needs its process fixed first.

Why a chat window is different from a built workflow

Asking a chat window to summarize a lease works once, for the person who wrote a good prompt that day. A task the firm repeats for every client, every season, needs more than that. It needs the firm's own instructions, the firm's template, and a worked example from a real prior-year file, packaged so that every preparer who runs it gets the same result. It also needs the review point defined in advance: what the reviewer sees, what they check it against, and what they initial.

That packaging is what separates a tool the firm can stand behind from a habit that varies by person. We describe the options in AI workflow vs agent vs skill vs custom software, and the design of the checkpoint in how to design the human review step.

One first workflow to build in each service line

• Tax: the open items list. The workflow reads what the client uploaded, compares it to the organizer and last year's file, and drafts the request for what is missing. The preparer checks it against the binder before it goes out.
• Audit: the footnote tie-out. The workflow compares each amount in the draft notes to the audited amounts and lists every difference with its location. The senior resolves each one, and review proceeds as usual.
• Client advisory: the monthly package narrative. The workflow drafts variance commentary and a meeting brief from the closed books. The advisor verifies the numbers and rewrites the explanation where they know the real reason.

Each of these is small enough to build and test outside busy season, and each leaves the people who sign in exactly the same position they hold today. For the firm-level groundwork that comes first (plans, policy, and training), see our guide on how to set up a CPA firm to use AI. If you would like help mapping your own service line this way, get in touch.

Common questions

Can AI prepare a tax return?

AI can do much of the reading and drafting around a return, such as sorting client documents, tying them to the organizer, and drafting the open items list and the client letter. The positions on the return and the signature stay with the preparer and reviewer. The AICPA tax standards say that using a tool does not absolve the member of professional responsibilities.

Is it permitted to use generative AI on an audit engagement?

In a July 2024 staff publication, the PCAOB staff reported that most of the audit firms it interviewed did not view PCAOB auditing standards as impediments to using generative AI, and also that some firms did not allow it when performing audit or attest procedures. That publication is a staff view and not a rule, so the answer for your engagements depends on your firm's own policy and the standards that apply to them. Confirm with your quality management lead before a tool touches an engagement file.

Which accounting tasks should not be given to AI?

Keep any step where a licensed person exercises judgment and signs: tax positions, materiality and risk assessment, conclusions on audit evidence, and the advice given to a client. Also hold back any task with no source document to check the output against, and any task where a wrong answer would reach the client before a qualified person saw it.

How does a reviewer check AI-prepared work?

The reviewer checks it the same way they check a first-year's work, against the source. A well-built workflow makes that faster by showing the document and page each figure came from and by listing what it could not find or was unsure of. The level of diligence does not go down because a tool prepared the draft.

What is the best first AI project for a CAS practice?

We would start with the monthly package narrative: variance commentary and a short meeting brief drafted from the client's own closed books, in the firm's template. The numbers already exist and are easy to check, the advisor already reviews the package before it goes out, and the task repeats every month for every client.