A practice, not a product — an internal policy framework. There’s nothing to buy; it’s a document, the training around it, and the attestation that people read it.

Business objective

The “here’s what you can and can’t do with AI here” rulebook everyone signs. An acceptable-use policy (AUP) sets the organization’s baseline standards for AI: which tools are sanctioned, what data may and may not go into them, what’s flat-out prohibited (e.g. pasting MNPI or client PII into a public chatbot), who to ask when unsure, and the consequences for ignoring it.

A useful AI AUP for a fund typically covers:

• Sanctioned vs. prohibited tools — what’s approved, and that everything else routes through the ai-gateway or gets reviewed first.
• Data handling — what classes of data are forbidden in AI tools, tied to your data classification labels (MNPI, client-confidential, legal-privileged, PII).
• Disclosure & review — when AI output must be checked by a human before it’s used or sent, and when AI involvement must be disclosed.
• Acknowledgement — everyone reads and attests, so it’s enforceable and auditable.

The seed calls it “the cheapest control there is,” and that’s exactly right: it costs a document and an attestation, and it’s the precondition that makes every technical control defensible (you can’t enforce a rule nobody was told).

When you need it

Day 1 — the cheapest control there is. This is the first thing you write, before any tooling. Employees are already using AI (sanctioned or not), so the policy is what turns “we never said you could” into an enforceable standard. For a small fund it’s a one-to-two-page document plus a sign-off; it grows teeth as you add the technical controls (ai-access-governance, dlp) that detect and enforce what the policy declares.

Security role

No direct technical enforcement — a policy neither blocks nor detects anything on its own. It’s the human-layer baseline that sets expectations and reduces the easy, accidental leaks (someone pasting positions into a consumer chatbot) that no firewall should have to catch in the first place. Think of it as the floor; the technical controls are the net under the people who ignore the floor.

How it gets backed by tooling (not a shortlist)

A policy is only as good as the controls that observe and enforce it:

• ai-access-governance — discovers shadow AI and enforces inline policy on what data flows to which model; turns the AUP from words into enforcement.
• dlp — stops sanctioned-data classes from leaving via prompts, backing the data-handling clauses.
• dspm — the data classification the AUP’s “don’t put X in AI” rules reference.
• enterprise-grc — where the policy, its versions, and attestations are recorded and audited.

Adjacent categories

• risk-tiers — the AUP is the floor everyone agrees to; tiers add graduated scrutiny above it.
• ai-access-governance / dlp — the detective/preventive controls that enforce the policy.
• enterprise-grc — system of record for the policy and sign-offs.
• ai-governance-platform — maps the policy to external frameworks (NIST AI RMF, EU AI Act, SR 11-7) for regulators.

Open taxonomy questions

• The AUP overlaps every other process page (it references tiers, gates, and zones). It’s the umbrella document; the others are specific mechanisms. Keep separate so the “what you write” vs. “what you do” distinction stays clear.