AI for accounting firms helps most with tax season work that is repetitive and easy to check: reading client documents, sorting files, flagging what is missing and drafting routine emails. It is risky wherever an answer needs judgment or goes to a client unreviewed. Treat AI output as a draft that a qualified person checks before anyone relies on it.
AI for accounting firms, task by task
AI for accounting firms means software that reads, sorts, summarizes or writes using models trained on large amounts of text and images. In tax season its value depends on the task. Some tasks it can draft, some it can check, and some should stay with a person. Sorting the firm's work into those three groups is the first step.
A task suits AI drafting when the output is routine and a person can check it quickly. A first-draft reminder email listing missing documents is a good example: the reviewer reads it, fixes anything off and sends it.
A task suits AI checking when the AI compares things and flags differences for a person to resolve. Comparing this year's documents with last year's, spotting a duplicate statement or noticing that a run of monthly statements skips March all fall here. The AI points; the person decides.
A task is human only when it needs judgment, signs the firm up to a position, or carries professional responsibility. Deciding how an item is treated on a return, signing a return and advising a client on what to do are in this group.
A task table for one tax season
The table sorts common tasks into the three groups. It reflects one reasonable split for a fictional firm, Marlow & Tate Tax, and a firm should draw its own lines based on its clients and its review process.
| Task | AI drafts | AI checks | Human only |
|---|---|---|---|
| Sorting uploaded files by document type | Yes | ||
| Reading figures from W-2s, 1099s and statements | Yes | Confirms key figures | |
| Comparing this year's documents with last year's | Yes | ||
| Flagging duplicate documents | Yes | ||
| Spotting gaps in a run of monthly statements | Yes | ||
| Drafting a missing-documents reminder | Yes | Reviews and sends | |
| Summarizing a long client email | Yes | ||
| Deciding how an item is reported on a return | Yes | ||
| Answering a client's tax question | Yes | ||
| Signing and filing the return | Yes | ||
| Deciding whether to extend a return | Yes |
Extension decisions are a good test of the line. AI can list which clients still have missing documents as a deadline nears, but whether to file an extension is a judgment about each client's situation. The guide to extension strategy for tax firms covers how firms make that call.
Can AI read client tax documents accurately?
AI can read many client tax documents well, especially clean PDFs of standard forms and bank statements, but no firm should assume a reading is right without checking it. Accuracy varies with image quality, layout and the document type. The practical question is not "how accurate is it?" but "how does the firm catch the errors it will make?"
Reading quality depends on the input. A PDF downloaded from an employer portal has crisp, selectable text. A phone photo of a crumpled form taken under a kitchen light does not. The same software can do well on one and poorly on the other, so a single accuracy figure from any vendor says little about a firm's own files.
Checking works best when it uses the document's own structure. For a bank statement, the opening balance plus deposits minus withdrawals should equal the closing balance. If the extracted rows add up, they are probably complete. For tax forms, compare the key figures against the page every time, and sample the rest.
Test on the firm's own files before trusting any tool. Take a handful of last season's client folders, including the worst phone photos, and run them through. Compare every extracted figure with the page by hand. The errors that turn up, and where they cluster, say more about the tool than any brochure. Repeat the test when the tool changes or when a new document type starts arriving in volume.
Good tools show their doubts. Software that marks a field it could not read clearly, or a page it could not classify, lets a reviewer go straight to the likely problems. Software that fills every field with a confident guess forces a reviewer to check everything or accept hidden errors.
Treat the reviewer as part of the system. Professional rules already expect this. Treasury Department Circular 230 (Rev. 6-2014), section 10.22, says a practitioner relying on another person's work product is presumed diligent if the practitioner "used reasonable care in engaging, supervising, training, and evaluating the person." The rule speaks of persons, not software, but the habit it describes, checking the work you rely on, applies to extraction output too.
Schema CPA reads client documents, sorts every page by type, issuer and period, and flags duplicates and statement gaps for the firm to review. It states no accuracy rate, and its output is meant to be checked before it is used.
AI for CPAs: what still needs professional judgment
AI for CPAs can gather, sort and summarize information, but the decisions that carry professional responsibility still need a CPA's judgment. These include how an item is treated on a return, whether a position is supportable, what to tell a client, and whether the file is complete enough to sign. AI can inform these decisions but cannot make them.
Treatment decisions need context AI does not have. Whether a payment is wages or a contractor fee, or whether an expense belongs to the business, depends on facts that may never appear in the documents. A CPA who knows the client asks the questions that settle it.
Completeness is a judgment, not a count. AI can report that every line on last year's list has a matching document this year. It cannot know that the client sold a rental property in August unless someone tells it. A preparer who reads the client's notes and asks about the year's changes catches what the list misses.
Representations to clients and to the IRS stay with the practitioner. Circular 230 section 10.22 requires due diligence in determining the correctness of oral or written representations made to clients and to the Treasury. A draft written by AI that the practitioner sends becomes the practitioner's representation.
Signing is the final line. Whatever tools helped prepare the return, the person who signs it is responsible for it. A firm's review process should make clear which checks happen before that signature, and who performs them.
Do client data rules limit AI in accounting firms?
Client data rules do limit AI in accounting firms, because sending a client's documents or return information to an AI service is a disclosure to a third party. Confidentiality rules, IRS consent rules for return information and the FTC Safeguards Rule all apply. None of them bans AI, but each sets conditions a firm must meet first.
Confidentiality under the AICPA Code
The AICPA Code of Professional Conduct (updated through September 2026) states in its Confidential Client Information Rule (1.700.001) that a member in public practice "shall not disclose any confidential client information without the specific consent of the client." Its interpretation on disclosing information to a third-party service provider (1.700.040) says that before disclosing, the member should either have a contract with the provider to keep the information confidential, with reasonable assurance the provider has procedures to prevent unauthorized release, or get the client's specific consent.
Tax return information under section 7216
The IRS Section 7216 information center (updated September 2026) explains that regulations under section 7216 set rules and consent requirements for the disclosure or use of tax return information by tax return preparers. A firm considering an AI tool that receives return information should read those rules before the first upload.
Safeguarding under the FTC Safeguards Rule
IRS Publication 4557 (Rev. 6-2024) explains that professional tax preparers fall under the FTC Safeguards Rule, which requires a written information security plan. Among the items that plan must cover, the publication lists the need to "select service providers that can maintain appropriate safeguards." An AI vendor that receives client data is a service provider for this purpose. The client portals and security category covers the plan and the rule in more depth.
The practical result is a short checklist before any AI tool sees client data: read the vendor's terms on how data is stored and used, decide whether a contract or client consent covers the disclosure, and add the vendor to the firm's security plan.
When AI answers should not reach a client
AI answers should not reach a client without review when they state a tax position, estimate an amount owed or refunded, interpret a notice, or tell the client what to do. These carry the firm's name and its professional responsibility. A reviewer reads every such message before it is sent, and many firms keep them out of AI drafting entirely.
Routine messages are different. A reminder that lists the documents still outstanding, with the issuer and period for each, is low risk once a person has glanced at it. That is the kind of message AI drafts well. Schema CPA drafts reminders like this from each client's missing items, and the firm reviews each draft and decides whether to send it; the client tracker demo shows the review step.
Watch for confident wrong answers. AI can produce a fluent reply that sounds right and is not. A client who receives it has no way to tell. That is reason enough to keep a person between the draft and the send button for anything beyond routine logistics.
Never let AI answer a question the firm has not seen. A chat tool that replies to client questions on the firm's behalf, without a person reading each answer, puts unreviewed advice in front of clients. Even a careful setup cannot rule out an answer that is wrong for that client's facts.
Plan the review time. AI shifts work from writing to reading, and reading still takes staff hours in the busiest weeks. The guide to busy season staffing for accounting firms covers how to plan those hours alongside the rest of the season's work.
An example of a message that needed changing
In a fictional example, an AI tool at Marlow & Tate Tax drafts a reminder to a client, Jun Park, listing a missing 1099-DIV from an example brokerage and a December statement for account XXXX-4821. The draft ends with a line saying the client "should expect a refund this year." The reviewer keeps the document list, deletes the refund line because no one has prepared the return yet, and sends the rest. In this invented example, the useful part of a draft is the list, and the risky part is anything that sounds like a conclusion.