AI Agents for Accounting Firms: What to Hand Off and What to Review.

AI agents for accounting firms can take intake, extraction, and the chasing. Here is where the CPA still signs off.

Where the line sits

AI agents for accounting firms can take on routine, document-heavy work, and a CPA still has to review what leaves the firm. The tasks sort into three buckets: hand off, review before it leaves, and the CPA keeps.

Firms use AI widely and agents rarely. Intuit commissioned a survey of 725 US accounting and bookkeeping professionals in May 2026, published as sponsored content, and 88% said they used AI for at least one client service in the previous 12 months. Thomson Reuters' 2026 AI in Professional Services Report, as summarized on its blog, puts a third of tax firms on generative AI, "with 14% using specifically agentic AI." A Financial Cents survey of 486 North American professionals found that 96% of the AI users in it mainly use general-purpose chatbots, though Financial Cents is a vendor and didn't say when or how it ran the survey.

These figures differ because each counts something different, and each has a commercial sponsor or a small sample. None represents all firms. The descriptions of what agents do come from the vendors that sell them, and the sources cited here include no independent accuracy benchmark.

Roman Kepczyk predicted in CPA Practice Advisor in November 2025 that the first agents firms actually adopt would come from accounting vendors. He also pointed to "a lot of concerns around data security, hallucinated data, and sloppy verification." That was a prediction, not a measurement. Vendors have announced agents since, though some are still in early access.

The vendor descriptions that mention review and the IRS guidance draw the same line. In a survey of dozens of firms in 8 countries from the practice-management company Uku, fielded in May and June 2026, 62% said their trust depended on a human approving before anything is sent or filed.

Bucket one: hand off the routine work

  • Chasing missing documents. In the Uku survey, 68% of respondents would trust an autonomous agent to chase clients for missing documents. A vendor announcement from May 2026 describes an agent that identifies missing client documents, drafts follow-up messages for review, and alerts staff when items arrive.
  • Onboarding steps. The same announcement says the agent triggers folder creation, questionnaire delivery, and welcome sequences when a client is created. Uku's respondents were less sure here: 21% would trust an agent to onboard new clients.
  • Extracting and matching transactions. A sponsored article from an accounting software maker describes an agent that extracts transactions from a client's PDF statements, compares them with the maker's accounting software, and flags errors with suggested fixes.
  • Flagging exceptions. The same article describes a payroll agent that notifies the person completing payroll of anomalies. The practice-management announcement describes flags on engagements where the effective hourly rate is dropping.
  • First drafts of client messages. In the Uku survey, 59% would trust an agent to prepare client emails and updates. Dext, a vendor, lists routine follow-ups and applying known rules consistently as suitable work for agents.

None of this measures how well the agents do the work.

Bucket two: review before it leaves

Where the sources describe review at all, it sits on what the agent produces: follow-up messages are drafts for review, and the payroll agent's output is a draft the user can revise and approve. Dext advises building review and approval into the process for work that touches compliance, tax treatment, or client outcomes.

Three kinds of output belong here, and for each the IRS guidance gives a reason. In June 2026 the IRS Office of Professional Responsibility (OPR) published "Introductory Guidelines for Responsible AI Use in Federal Tax Practice" (issue 2026-19), which the Journal of Accountancy covered on June 26. It applies existing Circular 230 duties to AI, and it says that "Practitioners cannot rely solely on AI; human scrutiny and editing are essential."

  • Anything sent to a client or filed, and any figure going into a return. Section 10.22 covers due diligence, and the guidance reads it as a duty to review AI-created documents before they go to a client or the IRS, and to verify facts, citations, and calculations.
  • Any tax position or written advice. Section 10.37 requires advice to rest on reasonable assumptions, and leaning on unverified AI projections may not meet that.

Bucket three: what the CPA keeps

Judgment calls, advice, and the signature stay with the CPA. Dext says not to give agents tasks that require judgment, and to keep humans in control of review and decisions. The IRS bulletin ends the same way: "final decisions must always rest with qualified professionals." It ties that line to existing Circular 230 duties, several of which stay with the firm's principals and practitioners:

  • Firm procedures. Section 10.36 makes principals take reasonable steps to give the firm adequate compliance procedures. For AI, the guidance lists staff training, secure data handling, accuracy monitoring, vetting outside tools, and documentation.
  • Fees. Section 10.27(a) bars unconscionable fees, and billing for time not actually spent because AI shortened the work may qualify, depending on the facts. The guidance recommends disclosing AI activity and passing cost savings to clients.
  • Audit responsibility. On October 13, 2026, the AICPA released a practice aid, "Artificial Intelligence (AI) Considerations for Financial Statement Auditors". The AICPA's chief auditor said the auditor remains responsible for the audit and the opinion. The practice aid covers audits, not tax or bookkeeping.

Client data: what the rules reach

Client information is the second limit, and the rules there carry legal penalties. The OPR guidance warns that one client's information can be reused for another client's matter, and that uploading data to unsecured systems risks disclosure, so practitioners must use only secure, enterprise-approved AI. It also cites IRC sections 6713 and 7216(a), which carry civil and criminal preparer penalties for unauthorized use or disclosure of tax return information, and section 10.51(a)(15), under which willful mishandling of return information can lead to discipline. In the sponsored Intuit survey, 60% said clients frequently or always asked for proof of AI data protection.

The FTC Safeguards Rule adds a written plan. IRS Publication 4557 says professional tax return preparers are "financial institutions" under the Gramm-Leach-Bliley Act and must have a written information security plan. It says to select service providers that can maintain safeguards, require that by contract, and oversee their handling of customer information. The rule itself, 16 CFR 314.4, also calls for encryption of customer information, multi-factor authentication, and telling the FTC no later than 30 days after discovering that unencrypted customer information of at least 500 consumers was acquired without authorization.

Neither the rule nor Publication 4557 mentions AI agents. Under the rule's definition of a service provider, any entity that "receives, maintains, processes, or otherwise is permitted access to customer information" through its services (16 CFR 314.2), the company that supplies or builds your agent may well qualify, so ask counsel whether your plan has to cover it.

The sources point to these questions for any vendor or builder:

  1. Is the tool secure and approved for enterprise use?
  2. Does the firm's written information security plan cover it?
  3. Does the contract require the vendor to maintain safeguards, and can the firm oversee its handling of client information?
  4. Can one client's information reach another client's work, or leave through an unsecured system?
  5. What does the agent do by itself, and what waits for a person's approval?

This is what the guidance says, not a compliance opinion. Your counsel or professional liability carrier can say how it applies to your firm.

What a build inside your tools looks like

A task the firm's software has no agent for can still get one, built inside the tools the firm already uses.

Tarsus Growth Company is an AI growth agency in West Chester, PA, and builds for accounting firms, among others. Tarsus builds systems that read and file the forms, statements, and receipts clients send, track filing and extension dates, and sort the firm's email and draft replies for review. The review steps are part of the design. In document processing, anything sensitive or uncertain waits in a queue for a person to review, and a check for missing pieces asks the sender and tells a person. In deadline tracking, a person confirms every date that carries legal or filing weight before anyone relies on it. In inbox and meeting follow-up, a reply waits for approval and nothing goes out until the firm says so. Those systems do the routine work of the first bucket and are built around the second.

If you'd like to talk through which of your firm's tasks fit which bucket, talk to Tarsus.

Keep reading.

All posts
Tattooed forearms and hands resting on a laptop on a wooden table, beside a phone and an envelope.

Foundations5 min read

Is AI Automation Worth It for a Small Business.

So is AI automation worth it for a small business? It can be, when you can name one repetitive task, count what it costs now, and measure the time saved against the full cost, checking included.