Firm AI Acceptable-Use Policy, TEMPLATE

How to use this template. This is a starting point for an AI policy your firm can adopt, not a finished legal document. Replace every [BRACKET], delete what doesn't apply, and have it reviewed by your own counsel and matched to your firm's WISP before adopting. The substance is built on, and cross-referenced to, Regulatory Foundation. This is not legal advice.

[FIRM NAME], Artificial Intelligence Acceptable-Use Policy Effective date: [DATE] · Owner: [NAME / TITLE, e.g., Qualified Individual under the WISP] · Version: [1.0]

1. Purpose & scope

This policy governs how [FIRM NAME] personnel may use artificial-intelligence tools (including generative AI such as Claude, ChatGPT, Gemini, Copilot, and any tax-research or workflow AI) in firm work. It exists to let us capture AI's benefits while meeting our legal and professional obligations, IRC §7216/§6713, the FTC Safeguards Rule and our WISP, the AICPA Code (including the Confidentiality Rule) and SSTS, Circular 230, applicable SSARS/GAAS, and our state board's rules. It applies to all partners, employees, and contractors, on any device or account used for firm work.

Guiding principle: AI drafts, organizes, and accelerates. A licensed professional reviews, decides, and signs. The license and the signature are ours, not the AI's.

2. Key definitions

• Tax Return Information (TRI), information furnished to us for return preparation, plus anything we derive or generate from it. (Broad, includes source docs, workpapers, and outputs.)

• Confidential Client Information, any non-public client information, whether or not it is TRI.

• PII, names, SSNs, EINs, account numbers, addresses, and other identifying detail.
• Approved AI Tool, a tool the firm has vetted and authorized per Section 4.
• Public/Personal AI, any consumer, free, or personal-account AI tool not under a firm-level agreement.

3. The core rule

Client TRI or Confidential Client Information may be entered into AI only through a firm-Approved AI Tool, used solely to prepare/assist the client's engagement or for permitted auxiliary services. Public or personal AI accounts may never be used with client-identifiable or return-derived information. Any use outside these limits requires §7216 review and, where applicable, valid written client consent and/or a confidentiality agreement with the vendor.

When in doubt, anonymize the information (remove identifiers; use "the client," "$X," "[STATE]") or don't use the tool.

4. Tool tiers, what's approved, restricted, and prohibited

Remember (per the Foundation doc): "enterprise" is not a magic word. Sending TRI to any third-party tool is still a disclosure under §7216, approval makes it defensible, by confirming the tool fits a permitted-use lane and is contractually controlled.

5. Permitted vs. prohibited uses

Generally permitted on an Approved Tool (low-level support to prepare this client's work):

• OCR / data extraction from source documents
• Summarizing, classifying, or organizing client-provided information
• Drafting workpapers, checklists, or client communications for our review
• Formatting, spreadsheet help, and document cleanup

Requires §7216 consent and/or is prohibited (outside the prep/auxiliary lane):

• Putting client info into public/personal AI
• Any tool/vendor that may train on, reuse, or share our inputs

• Offshore access to TRI (including foreign remote access to U.S.-stored data), special SSN limits apply; do not send SSNs offshore

• A tool making substantive tax determinations (interpreting/applying law to set the client's liability), a person reviews and owns every such conclusion

• Any non-preparation use (e.g., marketing analytics) on client-identifiable data

No client data at all is fine anywhere: generic tax research, non-client proofreading, and generic template/workflow drafting don't implicate §7216 because no TRI is used or disclosed.

6. Vendor approval, minimum diligence before a tool handles client data

[OWNER] must obtain written confirmation of all of the following before adding a tool to "Approved" (mirrors the 10-point standard in the Foundation doc and the FTC Safeguards Rule §314.4(f) service-provider duty):

1. No training / model improvement on our prompts, files, outputs, embeddings, or feedback.
2. Zero or minimal retention for the specific endpoints/features we use (not a marketing line).

3. No persistent application state unless necessary and approved (files, threads, assistants, vector stores, batch jobs, custom-GPT knowledge, project memory).

4. U.S.-only processing and access (absent valid §7216 consent for offshore).

5. No third-party/downstream tool calls with client data unless each is separately vetted.
6. Confidentiality and subprocessor terms appropriate for TRI (a DPA/BAA-equivalent).
7. Vendor human-access restrictions.
8. §7216/§6713 contractor-notice mechanics where vendor access to TRI is possible.
9. Use limitation, vendor uses our data only to provide the service to the firm.
10. Security controls consistent with our WISP (encryption in transit/at rest, access control, incident notification).

Approved tools and their approval basis are logged in [WISP / tool register location] and reassessed at least annually.

7. Consent & confidentiality

• IRC §7216: Where a use/disclosure falls outside a §301.7216-2 exception, obtain a compliant written consent (see the §7216 consent template) before the data is used/disclosed.

• AICPA Confidentiality Rule (ET §1.700.001): Independently, before Confidential Client Information reaches a third-party AI vendor, the firm must have client consent or a confidentiality agreement/DPA with that vendor. (This applies even if a §7216 exception means no tax-consent form is required.)

• State law: [If GA: Rule 20-12-.11 imposes an independent confidentiality duty; Rule 20-12-.19 incorporates AICPA standards into licensure.] [Adjust for your state.]

• Good practice: consider disclosing our general use of AI to clients in the engagement letter.

8. You are the reviewer of record

No AI output goes to a client, a taxing authority, or a workpaper/attest file without competent human review. Specifically:

• Verify every authority. Treat any Code section, regulation, case, ruling, or figure from general AI as unverified until checked against primary source. AI is not a source of law.

• Own the judgment. Under SSTS §1.4 you remain fully responsible for the work product whether or not you used AI. [Attest work: under AU-C 500, AI output is audit evidence to be evaluated, same reliability, documentation, and skepticism duties.]

9. Incident response

If client information may have been exposed through an AI tool (wrong tool used, data pasted into public AI, suspected vendor breach), report it to [OWNER] immediately. The firm will follow its WISP incident-response procedures, including breach assessment and any client/authority notification.

10. Acknowledgment

I have read and agree to comply with this AI Acceptable-Use Policy.

Name: __________________________ Signature: __________________________ Date: ____________

Template maintained with the AI Lab for Accountants library. Built on Regulatory Foundation and Guardrails. Customize and obtain legal review before adopting. Not legal advice.