Asked to sell firm data for AI training? Be very cautious

Editor’s note: The author is the communications manager for the AICPA Professional Ethics Division.

Offers to buy, receive, or otherwise use firm data for AI model training are rising. They can sound appealing, but client confidentiality obligations do not disappear because the data is redacted, labeled “anonymous,” or framed as a technology partnership. Before you say yes, you need to evaluate whether the disclosure is permitted, authorized, and consistent with your professional responsibilities.

Information you obtain to serve clients is not unrestricted inventory. The prudent default response to any request to sell, give, or share confidential client information for AI training is to decline unless you can clearly establish that the transaction is permissible.

The AICPA’s “Confidential Client Information Rule” (ET sec. 1.700.001) and related definition of “confidential client information” generally prohibit members in public practice from disclosing such information without the client’s specific consent, subject to the rule’s exceptions. A sale or other transfer intended to train an AI model does not eliminate that specific-consent requirement. Permission to process information in order to perform an engagement does not authorize you to hand the same information to a buyer (or any other party) to train an AI model.

Calling the recipient a “technology partner” does not resolve the issue. Sharing data for the recipient’s independent commercial purposes warrants different scrutiny than engaging a technology provider to help you deliver client services. Any consent you seek should clearly explain the information involved, the recipient, and the training purpose. Do not bury the transaction inside a generic technology disclosure that clients might overlook or misunderstand.

Redaction does not, by itself, turn confidential client information into nonconfidential information. Removing names and other identifying details does not change the confidential nature of the underlying information or eliminate your obligations. The AICPA’s nonauthoritative Ethics Staff Insights: AI Through an Ethics Lens addresses this point in its discussion of the “Disclosure of Client Information to Third Parties” interpretation (ET sec. 1.700.060), which applies even when client information is disclosed without the client being specifically identified. If the information remains confidential client information, that interpretation requires the client’s specific consent — preferably in writing — before you disclose or use it. Do not treat redaction, or a buyer’s “anonymous data” label, as permission to sell, give, or share the information for model training.

Tax return information requires a separate analysis. The Tax Adviser’s discussion of AICPA and IRS confidentiality requirements explains that Sec. 7216 can impose criminal consequences for knowing or reckless unauthorized disclosure or use of tax return information. Have counsel evaluate the proposed transaction, applicable exceptions, and consent requirements. Do not assume an ordinary engagement letter is sufficient.

AI training creates risks beyond the data transfer

AI model training is not simply secure document storage. As noted in the JofA article “Accounting Ethics in the Age of AI”, sensitive information entered into some AI tools could potentially resurface elsewhere. CPA.com’s security and risk guidance likewise recommends controls over data. Ask whether the recipient can meaningfully restrict downstream use and address information that has already been incorporated into a trained model.

Put a gate in front of every proposal

Before you consider selling, giving, or sharing data for AI training, require legal, ethics, and information-security review of the proposal. Inventory the proposed data. Establish the firm’s disclosure and licensing rights. Examine client agreements and privacy obligations. Evaluate consent, retention, onward-   sharing, security, incident-response, and liability terms. These steps build on AICPA-endorsed insurance program guidance that urges firms to investigate training use, access, data ownership, and contractual protections.

When you cannot establish client authorization and adequate privacy and confidentiality protections, do not proceed.

The recommendation is caution, not a claim that AICPA rules categorically prohibit every data transfer. Permission alone, however, does not answer the business question: Is a training-data payment—or a no-payment data share—worth putting client trust at risk?

Resources

AICPA Professional Ethics Division staff will be publishing a deeper dive on this topic later this quarter in the Online Ethics Library. Keep an eye out for a new Ethics Staff Insights article.

In the meantime, these are available:

• The “Confidential Client Information Rule” (ET sec. 1.700.001) and related definition.
• The “Disclosure of Client Information to Third Parties” interpretation (ET sec. 1.700.060), which refers to nonauthoritative guidance on categories of information and what members need to consider before sharing each category.
• Ethics Staff Insights: AI Through an Ethics Lens, which reinforces confidentiality considerations related to AI
• Security and Risk Considerations — CPA.com
• Should I Disclose My Use of Gen AI to Clients? — AICPA-endorsed insurance program guidance

— Kelly D. Mullins is the communications manager for the AICPA Professional Ethics Division. To comment on this article or to suggest an idea for another article, contact Jeff Drew at Jeff.Drew@aicpa-cima.com.