Generate an AI policy you can ship by Monday.

NIST AI Risk Management Framework (Govern / Map / Measure / Manage), EU AI Act risk classification, and ISO/IEC 42001 management system requirements. The language an auditor will recognize.

References ChatGPT, Claude, Copilot, Gemini, and Microsoft 365 Copilot by name — and gives autonomous agents an owner, a budget, and a kill switch. Gartner projects 150K+ agents per Fortune 500 by 2028, and only 13% of leaders feel adequately governed. This template assumes they’re coming.

Includes the honest-coverage language that protects you when the auditor asks “how do you know?” — Observed, Inferred, Enforced, and Not Covered, without overclaiming.

Microsoft Word .docx for legal and Markdown for engineering. Same content, both formats, no PDF lock-in.

What counts as “AI” in your environment, and who the AI actors are — the humans, apps, and agents using AI through your stack. Third-party assistants (ChatGPT, Claude, Copilot, Gemini), embedded features (Notion AI, Microsoft 365 Copilot), API usage by engineering, autonomous agents that spend and act unattended, and internal models. Without a real scope — and an owner for every actor — the rest is unenforceable.

What each AI actor can use AI for, by role and data class. Drafting and brainstorming with non-sensitive data. Code generation with non-proprietary code. Summarization within approved tools. For agents: an approved model scope and a budget, not an open-ended mandate. Clear, role-aware, not absolutist.

What’s out: sensitive customer PII into unmanaged consumer tools, regulated data (PHI, financial account data, etc.) into non-BAA providers, code containing secrets, anything covered by an NDA you can’t cite.

The list — by name — and the procurement path for adding more. Includes data-residency and retention notes from each provider’s enterprise terms. Updated when contracts change, not annually.

How the company observes and controls AI usage per actor. Which signals you collect (gateway, provider APIs, expense, SSO). What you budget, what you cap, and how you disable a misbehaving actor’s Tallin-mediated model access when spend or behavior goes wrong. The honest coverage statement that protects you in an audit: “we observe X%, infer Y%, enforce Z%, and acknowledge gaps in W% — including direct provider calls made outside Tallin and tool / MCP execution side effects.”

What happens when sensitive data goes to the wrong place. Containment steps. Provider notification. Customer notification thresholds. The same shape as your existing security incident plan — extended for AI.

Quarterly review of the provider list and coverage report. Annual policy refresh. Out-of-band review when a major new tool enters the company. Documented, with dates and owners.

Who owns the policy (typically Security or a designated AI Governance lead), who reviews it (Legal, IT, the AI Council if you have one), and who enforces it day-to-day (managers, with escalation paths). And — for agents that act without a human in the loop — a named human owner accountable for that actor’s scope, budget, and off switch.