See training for CPA firms →

Can US CPA firms use Claude, ChatGPT and other AI tools without breaching AICPA guidance? Every managing partner we meet in 2026 asks some version of this. Associates are already pasting trial balances into consumer chatbots. Firm leadership wants a clear answer before an ethics complaint or an IRS referral forces one.

The short answer: Yes — with specific safeguards. Paid business-tier accounts with no-training terms, taxpayer anonymization before input, a §7216 consent framework, and an updated engagement letter put a firm comfortably inside AICPA Rule 1.700 and IRC §7216. Free consumer tiers with identifiable client data do not.

This piece walks through the actual rules — AICPA Code sections 1.300, 1.400 and 1.700, the IRC §7216 trap most firms overlook, and the state board nuances in California, New York, Texas and Florida — then gives you a five-step compliance checklist you can hand to a partner meeting.

AICPA Code of Professional Conduct: the three rules that matter

The AICPA Code does not mention AI by name, and it does not need to. Three existing rules do all the work. If a firm treats AI tools the way it treats any other third-party service provider — with due diligence, written agreements, and client awareness — it will usually land on the right side of all three.

Rule 1.300.001

Integrity and Objectivity

Rule 1.300 requires members to maintain objectivity and avoid misrepresentation. For AI use, that means two things. First, you cannot present AI-generated analysis as your own professional judgment without reviewing it — sign-off has to be a real review, not a signature on an unread draft. Second, if a client reasonably believes they are receiving human-prepared work, a firm that silently substitutes AI output is sailing close to misrepresentation. The fix is almost always disclosure, not prohibition.

Rule 1.400.001

Acts Discreditable — and the competency link

Rule 1.400 combines with the general standards in Rule 1.300 on professional competence. A CPA signing work they cannot evaluate — because the AI did a calculation they would not themselves know how to check — risks a competence breach. In practice this flags the same workflows that worried partners already worry about: novel international tax questions, unusual revenue recognition, and anything an associate is using AI to handle because they could not do it unaided. AI is fine as a drafting and analysis assistant. It is not fine as a substitute for the professional knowledge a licensee is certifying they hold.

Rule 1.700.001

Confidential Client Information

This is the one most firms focus on, and rightly so. Rule 1.700.001 prohibits disclosing confidential client information to a third party without specific client consent. A commercial AI provider is a third party. Whether pasting a client’s data into a chatbot counts as “disclosure” depends almost entirely on the terms of the plan you are using.

On a free consumer tier, inputs can be retained, used for training, or surfaced for human review. That is a disclosure. On a paid business or enterprise tier — Claude for Work, ChatGPT Team or Enterprise, Microsoft Copilot with the business data protection boundary — providers contractually commit not to train on inputs and to keep data within a defined processing scope. Combined with anonymization before input, this is the posture most firms land on.

IRC §7216: the hard legal constraint

Section 7216 of the Internal Revenue Code is the one AICPA guidance cannot wave away. It is a federal criminal statute. Any preparer who knowingly or recklessly discloses or uses tax return information without authorization faces fines up to $1,000 per disclosure and up to one year of imprisonment, with civil penalties in §6713 running parallel.

“Tax return information” is defined broadly: anything a client furnishes in connection with return preparation — wage figures, business income, SSNs, EINs, dependents, prior-year positions. Treas. Reg. §301.7216-3 requires taxpayer written consent in a specific format before most disclosures or uses outside return preparation itself.

The question for 2026 is whether sending tax return information to an external AI service counts as a disclosure. The conservative answer, and the one most large firms have adopted, is: assume yes unless you can point to a specific exception. The two routes firms take are:

  1. Anonymize before input. If no taxpayer-identifying information leaves the firm — names, SSNs, EINs, exact dollar amounts tied to identity are stripped or rounded — there is no tax return information in the AI prompt, and §7216 is not triggered. This is the dominant pattern.
  2. Obtain §7216 consent. Where identifiable data must be used (for example, a complex scenario modeled on a specific client’s facts), a compliant written consent naming the AI provider, the purpose and the data categories can authorize the use. Consent must be signed before the disclosure and must meet the Rev. Proc. 2013-14 format requirements.

Most firms we train pick route one for 95% of workflows and reserve route two for specific engagements where anonymization would strip the analysis of value.

State board variations

Most state boards of accountancy adopt the AICPA Code by reference, so the national rules travel. A handful have added their own commentary worth knowing.

California

The California Board of Accountancy has emphasized technology competency and the duty to understand any tool whose output a licensee signs. CA also has its own comprehensive privacy regime (CCPA/CPRA) that applies to client personal information independently of §7216.

New York

NYSED has flagged AI as part of continuing professional competence obligations. Firms serving NY clients should expect questions about AI governance in future peer review cycles, and the NY SHIELD Act adds data-security obligations on top of federal rules.

Texas

The Texas State Board of Public Accountancy has aligned closely with AICPA positioning: AI use is permitted with appropriate safeguards and client transparency, and the licensee’s professional responsibility is undiminished by the tool.

Florida

Florida’s Board has signaled that undisclosed AI use in attest engagements will be treated as a potential independence and integrity issue. Firms doing audit or review work for FL clients should put AI disclosure directly in the engagement letter.

A five-step compliance framework

This is the checklist we walk firms through. It is deliberately boring. Boring is the point.

  1. Data classification. Produce a one-page matrix listing the data your firm handles — taxpayer identifiers, financial statements, source documents, meeting notes — and mark each as “never to AI,” “anonymize first,” or “AI-acceptable.” Every associate should know which bucket they are in before they paste.
  2. Anonymization patterns. Give staff three to five concrete patterns: replace names with role tags (“Client A,” “Spouse,” “S-Corp 1”), round dollar figures within 5%, strip SSNs/EINs, remove geographic specifics tighter than state. Good training beats long policies.
  3. DPA and terms review. Confirm every AI tool in use is on a paid business tier with a Data Processing Addendum, no-training clause, and US data residency where relevant. Document the review with the date and the plan tier.
  4. Staff training. Hands-on, workflow-specific training beats a one-hour webinar every time. Two hours on the five workflows your firm actually runs changes behavior.
  5. Firm policy and engagement letter. A one-page AI policy signed by every staff member, plus a short clause in the standard engagement letter disclosing AI-assisted workflows. This is what peer reviewers and state boards will ask for first.

What the AICPA has actually said

The AICPA has published successive practice aids and guidance through 2024 and 2025 on technology-enabled auditing and professional use of generative AI, typically through its Assurance Services Executive Committee and the Center for Plain English Accounting. The through-line is consistent: AI is permitted, licensee responsibility is unchanged, confidentiality rules apply in full, and client transparency is best practice. The AICPA has not banned any specific tool or vendor, and has not published a “safe tools” list. Firms are expected to run their own due diligence.

How Claudable training addresses this

Every Cowork training we deliver to US CPA firms includes the anonymization patterns above and a short block on §7216 and Rule 1.700 in practice — what to paste, what not to, and how to redact in under 30 seconds. We are not a law firm and the training is not legal advice, but it closes the gap between “AI is allowed” and “staff actually behave safely.” Our Cowork training is delivered remotely via Google Meet to US practices anywhere in the country. A related UK-focused piece on the equivalent audit and assurance rules is here.