Skip to main content
TaxCTO.
← Back to Dispatches

Drafting an Acceptable Use Policy for Generative AI

The barrier to entry for Artificial Intelligence is now a web browser. If your firm does not have a legally binding Acceptable Use Policy (AUP) for Generative AI, your staff are likely violating IRC Section 7216 right now.

The Shadow IT Threat Model

Tax preparation involves synthesizing massive amounts of unstructured data (brokerage statements, K-1s, prior year returns). Large Language Models (LLMs) excel at this. Recognizing this efficiency, junior staff are increasingly creating free accounts on OpenAI (ChatGPT) or Anthropic (Claude) and uploading client documents to have the AI "summarize the K-1 footnotes" or "draft a response to this IRS notice."

By default, public LLMs use user inputs to train their models. When a staff member uploads a client document containing Personally Identifiable Information (PII) to a public consumer tier AI, they have executed an unauthorized disclosure of taxpayer information.

Core Components of the AI Policy

A defensive AUP must explicitly state what is forbidden and provide a sanctioned alternative.

  1. Absolute Prohibition of Public LLMs for Client Data: Explicitly ban the use of free or consumer-tier AI tools (ChatGPT, Claude, Google Gemini) for any task involving client data, firm financial data, or proprietary methodologies.
  2. Sanctioned Alternatives: You cannot simply ban a highly efficient tool; staff will find workarounds. The firm must deploy an enterprise-grade solution (e.g., Microsoft Copilot for Microsoft 365, or a private Azure OpenAI instance) where the vendor contractually guarantees data is not used for model training and remains within the firm's compliance boundary.
  3. Output Verification Mandate: Establish that AI outputs (hallucinations, incorrect tax code citations) do not absolve the preparer of liability. All AI-generated work must be reviewed and verified by a human expert before being incorporated into workpapers or client communications.

Technical Enforcement

Policy without enforcement is merely a suggestion. A CTO will enforce the AUP technically by utilizing the firm's firewall or Endpoint Detection and Response (EDR) agent to block access to unauthorized consumer AI domains on all firm-issued devices, while simultaneously provisioning access to the secure, enterprise-tier equivalent.

Common Mistakes in Infrastructure

A critical failure mode we observe is assuming that paying an MSP a monthly fee absolves the firm's partners of legal liability under FTC guidelines. (Reported by Ponemon Institute, Q3 2023).

Empirical Data & Metrics

Metric Industry Average Bureau Standard
MFA Adoption 42% 100% (Zero Exceptions)
Bare-Metal Restore Time 5+ Days < 12 Hours

Frequently Asked Questions

How does this impact our cyber insurance?

Insurance carriers now require attested proof of EDR and MFA. Failing to maintain these technically, even if stated on the policy application, can result in a denied claim.

Can we handle this internally?

Unless your firm employs a dedicated, full-time cloud security architect, attempting to self-manage enterprise compliance usually results in critical gaps.

Related Internal Resources