AI Acceptable Use Policy

An AI acceptable use policy is a written company rule that tells staff and vendors how to use AI tools at work, which inputs are banned, and which outputs need review. It sets guardrails that protect client data, compliance, and brand safety.

Outsourcing teams sit closest to the risk. A Manila agent handling US health records, or a Bangalore developer piping code through a chatbot, can expose regulated data in one paste. A clear policy converts AI risk into rules a shift lead can enforce.

Good policies split what’s allowed, restricted, and forbidden. Public generative AI tools might be fine for internal brainstorming, blocked for client PII, and forbidden for source code.

Vendor tools sit in their own tier — their data-handling promises differ from consumer chatbots.

Regulators have not waited for consensus. The NIST AI Risk Management Framework from January 2023, and the EU AI Act in force since August 2024, both assume an organisation already has an internal AUP. The policy is now table stakes for enterprise deals.

Key takeaways

• Names AI use rules across staff, contractors, and vendors, with named consequences for breach.
• Splits tools by risk tier — approved, restricted, prohibited — with the data class each tier can touch.
• Requires human review of any AI output that reaches a client, a filing, or production code.
• References external frameworks like NIST AI RMF and the EU AI Act so audits map cleanly to regulator expectations.
• Is reviewed at least every six months because model capabilities shift faster than annual cycles.

How it works

An AI acceptable use policy works like an access-control matrix. It ranks tools by risk, ranks data by sensitivity, and defines which combinations are allowed. Governance leads publish it, security enforces via DLP, and managers approve exceptions.

Tools map to tiers. Data maps to classes: public, internal, confidential, restricted. The policy answers a single question per cell: can this class of data pass through this tier of tool, and if yes, under what review?

Most policies map three axes — user role, tool category, and data class. A payroll analyst using ChatGPT with sample vendor names is one combination. That same analyst pasting a live pay stub is a policy violation the DLP layer should block.

Enforcement runs on three surfaces. Endpoint DLP watches for regulated strings heading to unapproved AI domains. SSO logs surface which AI tools staff are actually using each week. Quarterly manager attestations catch the shadow-AI cases neither logs nor DLP can spot.

Examples

Real acceptable use policies are already published. They vary in specificity, but the strongest ones name banned tools, banned data types, and named consequences instead of vague ethical language. Here are four with dates.

Samsung (2023): Samsung banned staff use of ChatGPT in May 2023 after engineers pasted proprietary chip source code into the tool. The internal policy that replaced the ban limits AI prompts to 1024 bytes and blocks upload of any Samsung code base.

JPMorgan Chase (2023): JPMorgan restricted staff use of ChatGPT in February 2023 under existing controls on non-approved software. The bank has since built its own internal LLM assistant so employees have a compliant alternative that keeps prompts inside the tenant.

US Federal government (2024): The Office of Management and Budget issued Memorandum M-24-10 in March 2024, requiring every federal agency to publish a public AI use case inventory and designate a Chief AI Officer.

European Union (2024): The EU AI Act, in force since August 2024, bans subliminal manipulation and social scoring outright, while classifying most business AI as either high-risk or limited-risk with specific transparency duties.

Related terms

AI acceptable use policies sit inside a wider governance stack. They inherit from data policies, feed vendor risk assessments, and get audited alongside compliance controls. The related concepts below map the neighbouring terrain.

• Artificial Intelligence: the umbrella technology this policy governs, spanning generative, predictive, and agentic systems.
• Compliance: the operational discipline that turns written policy into evidence an auditor will accept.
• Risk Management: the frame that classifies AI use cases by likelihood and impact, then matches each to a control tier.
• Generative AI: the tool category that triggered most current AUPs because it accepts free-text prompts and returns free-text answers over any data pasted in.
• Quality Assurance: the review layer that catches AI hallucinations and misuses before either reaches a client deliverable.
• Business Process Outsourcing (BPO): the delivery model where AI acceptable use rules must cross corporate boundaries into vendor teams.

FAQ

What must every AI acceptable use policy define?

Every policy must name approved tools, banned data types, required review steps, and named consequences for violation. Anything vaguer than that fails the first audit or breach investigation. Regulators expect specifics, not aspirational language about ethics.

Who owns the AI acceptable use policy inside a company?

Ownership usually sits with the Chief Information Security Officer or a dedicated Chief AI Officer, with input from legal, HR, and the head of engineering. Governance is shared, but accountability names one person on the executive team.

How is an AI acceptable use policy different from a general IT policy?

An AI policy addresses prompt handling, model risk, and hallucination review — controls a standard IT policy never anticipated. It supplements rather than replaces the general acceptable use rules.

How often should the policy be reviewed?

Every six months at a minimum, and immediately when a new class of AI tool enters the stack. Model capabilities shift faster than annual policy cycles can track, and older policies miss features like agent tool-use or persistent memory.

Do outsourced teams need their own version?

Outsourcing partners work under the client’s policy plus their own local law, so a signed addendum in the master services agreement handles the overlap. One integrated document is cleaner than two competing sets of rules.

What happens when the policy is broken?

Consequences range from retraining to termination and, in regulated sectors, mandatory breach disclosure to authorities.

Compare vetted BPO partners with documented AI governance policies in the OA Directory.