Accounting · September 24, 2026 · guide

AI for Accounting Firms: What Works, What Is Unproven, and What the IRS Now Expects

AI for Accounting Firms: What Works, What Is Unproven, and What the IRS Now Expects
admin · Updated Sep 24, 2026

AI for accounting firms is useful today in one narrow band: tasks where every output can be checked item by item, such as reading a receipt, renaming a document, or suggesting a transaction category. Outside that band, in tax research, return preparation, and “AI coworkers,” no independent measurement covers the current generation of accounting products, and the binding constraint for US tax preparers is legal: sending identifiable client tax return information to an AI tool can be a disclosure under a criminal statute.

A small firm that adopts AI well will start with document handling, write down which tools may receive client data, put consent language in its engagement letters, and review every AI output before it reaches a client or a return.

The rules that apply to AI in a US tax practice

The IRS has said Circular 230 applies in full

On 24 June 2026, the IRS Office of Professional Responsibility issued Alert 2026-19, “Introductory Guidelines for Responsible AI Use in Federal Tax Practice.” Reporting by CNBC and practitioner publications summarizes three points: practitioners must review and verify AI-generated work, Circular 230’s due diligence, competence, and oversight duties apply to AI-assisted work in full, and billing should reflect efficiencies gained through AI.

The alert names the preparer penalty provisions directly: IRC section 6713 (civil) and section 7216(a) (criminal) for unauthorized use or disclosure of tax return information, along with Circular 230 section 10.51(a)(15). It warns that generative AI platforms may create disclosure risk, especially when data is uploaded to unsecured or public systems.

What the alert does not do is equally important. As one practitioner analysis of the text points out, it does not say whether connecting an AI tool counts as an authorized disclosure, does not address the contractor exception, and does not mention client consent. It uses the word “unauthorized” without defining it. The IRS has not answered the question most firms are asking.

CNBC also reports that the AICPA’s 2026 submission to the IRS on guidance priorities asked for additional guidance on technology, including AI, in tax preparation.

Section 7216 is the risk that matters most

Section 7216 makes it a crime for a tax return preparer to knowingly or recklessly disclose or use tax return information without authorization, with a fine of up to $1,000, imprisonment of up to one year, or both, per violation. Section 6713 adds a parallel civil penalty.

The regulations define disclosure broadly. Practitioner commentary reads that definition to cover a prompt paste, a file upload, or an API call to a third-party AI provider. Whether one of the exceptions in Treasury Regulation 301.7216-2 covers a particular AI tool depends on the vendor contract and where the data goes, and no IRS guidance settles it.

Where consent is required, Treasury Regulation 301.7216-3 sets the requirements. For Form 1040 series returns, Revenue Procedure 2013-14 specifies the required language, format, and procedure. For other taxpayers, the consent may take any format, including an engagement letter, provided it meets every requirement.

The conservative reading that most practitioner sources converge on: do not put identifiable client tax return information into a general-purpose AI tool without compliant written consent, or anonymize the input, or use an enterprise tool the firm has vetted and contracted for.

Other standards that reach AI use

The AICPA’s Statements on Standards for Tax Services, revised effective 1 January 2024, include a standard on reliance on tools (SSTS 1.4). A practitioner source reports that AI falls within the standard’s definition of a tool. Tax preparers are also subject to the FTC Safeguards Rule on protecting customer information.

In February 2026, COSO published guidance on achieving effective internal control over generative AI, which the Journal of Accountancy covered. It maps AI-specific risks to the Internal Control Integrated Framework and is relevant to firms whose clients will ask how AI is controlled.

What the evidence says about AI accuracy in accounting

Independent measurement of accounting AI is thin, and what exists predates the products firms are being sold.

The most useful peer-reviewed benchmark we found is CA-Ben, published in SN Computer Science in 2025. It tested six large language models on multiple-choice questions from India’s chartered accountancy examinations. The strongest models, Claude 3.5 Sonnet and GPT-4o, averaged about 56 percent and 54 percent accuracy. Every model showed the same two weaknesses: multi-step numerical computation and interpretation of complex legal text. Those are the two skills tax work depends on most.

The limits of that study mirror what we found in legal AI. It tested general-purpose models from 2024, on Indian exams, in multiple-choice format. It does not measure any accounting-specific product sold in 2026, and we found no independent benchmark that does. Other published benchmarks cover German tax law and Romanian tax and accounting; they are useful for researchers and do not transfer to US practice.

Adoption is moving faster than evidence. A 2024 Thomson Reuters Institute survey of 330 tax and accounting professionals found about 25 percent using public generative AI tools and 9 percent using proprietary tax-specific AI. A June 2026 survey by Blue J and CPA.com of more than 1,000 tax professionals found 60 percent using AI for tax research at least weekly, up from 33 percent in 2025. Blue J sells AI tax research, and CPA.com is the AICPA’s technology subsidiary; both have an interest in the finding.

How AI for accounting firms sorts out in the software we reviewed

We applied our three-tier framework to every AI claim in the five accounting products we reviewed. Working means the mechanism is obvious and the result is buyer-measurable. Promising means the technology works but results depend on a process the firm supplies. Marketing means the claim outruns the product, or the improvement is real but is neither new nor AI.

ClaimVendorSort
AI document renaming and taggingCanopy, TaxDome, Financial CentsWorking
Receipt and invoice data extractionDext (also Hubdoc)Working
AI File Validator (checks uploads against the request)Financial CentsPromising
AI Assist learned bookkeeping rulesDextPromising
Smart Intake, Tax Workflow Automation, Close AutomationCanopyPromising
Kai AI coworker (limited early access)KarbonPromising
“$39,191 saved per employee per year”KarbonMarketing
“56 hours and $19,200 saved per firm each year”Financial CentsMarketing
“99.9 percent accuracy”DextMarketing
“AI-driven analytics”TaxDomeMarketing, pending evidence
Coworker autonomous agent (beta)CanopyMarketing, for now

The pattern is consistent. Every working claim is a document task where a staff member can see immediately whether the AI got it right. Every promising claim depends on the firm’s own data hygiene or corrections. Every marketing claim is either a savings figure from the vendor’s own customers or an autonomy claim for a product still in beta.

Our individual reviews explain each sort: Karbon, Canopy, TaxDome, Financial Cents, and Dext.

How accounting vendors charge for AI

The vendors we reviewed use four different approaches, and the approach affects cost more than the headline price.

Included on every tier. TaxDome and Financial Cents include their document AI in every plan with no credits.

Per client, charged once. Canopy charges one Smart Intake credit per client the first time it is used, then unlimited use for that client during the term. Tax Workflow Automation starts at $34 per client credit, and Close Automation costs $10 per connected client per month after five free connections.

Per client, monthly, at a launch price. Dext charges $7 per client per month for AI Assist custom guidance “until the end of October as part of a launch offer,” and has not published the price after that.

Unpriced. Karbon has not published a price for Kai, which is in limited early access.

For a firm that grows its client count, per-client pricing scales the AI bill with the book of business. For a firm with a stable roster, Canopy’s once-per-client model is the most predictable metered approach we have seen. We track these mechanics across industries in our analysis of AI software pricing.

A practical AI policy for a small firm

This is what the rules and evidence above point to. It is a starting point for a conversation with your own counsel.

  1. List every AI tool in use, including staff use of consumer chatbots. Most firms find more than they expected.
  2. Decide which tools may receive client data, and in what form. Identifiable tax return information should go only to tools covered by a vetted contract, or only with compliant consent.
  3. Add section 7216 consent language to engagement letters where AI tools will touch return information. For Form 1040 clients, follow Revenue Procedure 2013-14’s format.
  4. Read each vendor’s data terms: whether client data trains models, how long it is retained, and where it is stored. Dext’s AI Assist, for example, processes data in EU and UK regions.
  5. Review every AI output before it goes to a client or onto a return. Alert 2026-19 makes the practitioner responsible for the result.
  6. Document the procedures. Written procedures are what an examiner or a client will ask to see.
  7. Revisit billing. The IRS alert says fees should reflect efficiencies gained through AI.
  8. Start with document handling. Renaming, extraction, and upload validation deliver measurable time savings at low risk.

Who should wait

A firm without a written consent process should hold off on any AI tool that touches identifiable return information until it has one. A firm considering an autonomous “AI coworker” should wait until the product leaves beta and publishes both a price and some evidence of what it does. And any firm tempted to use a general-purpose chatbot for tax research should treat its answers as leads to verify against primary sources, given how the best models performed on multi-step calculation and statutory interpretation.

Frequently asked questions

Can accountants legally use AI with client tax data? Nothing in IRS rules bans AI. The rules govern the data and the output. Under section 7216, disclosing identifiable tax return information to a third-party AI tool may require compliant written client consent unless an exception applies, and IRS Alert 2026-19 confirms Circular 230 duties apply in full to AI-assisted work. This is not legal advice.

What did the IRS say about AI in June 2026? The Office of Professional Responsibility’s Alert 2026-19, issued 24 June 2026, says practitioners must verify AI-generated work, applies Circular 230 duties to AI-assisted work, names sections 6713 and 7216(a), and says billing should reflect AI efficiencies. It did not address client consent.

How accurate is AI on accounting and tax questions? Independent evidence is limited. A 2025 peer-reviewed benchmark on Indian chartered accountancy exams found the best general-purpose models averaged about 56 percent, with weaknesses in multi-step calculation and legal interpretation. No independent benchmark covers current accounting-specific products.

Which AI features in accounting software are worth paying for? Document tasks where each result is checkable: renaming, extraction, and upload validation. They are included in several products at no extra cost.

The bottom line

AI is already paying off in accounting firms where it handles documents, and several practice management platforms include that capability at no extra charge. Beyond documents, buyers are being asked to pay for tools whose accuracy nobody independent has measured.

The legal picture moved in June 2026, and firms now know the IRS expects them to verify everything and protect return data. What the IRS has not said is when an AI tool is an authorized recipient of that data. Until it does, consent and careful vendor selection are the safe course.