OpenAI in Teams: what the 2026 Usage Policies require
Organizations that connect OpenAI models (a Codex or ChatGPT subscription, an OpenAI API key) and serve them to their people through Prism Teams or Enterprise must work within OpenAI’s Usage Policies, effective 2026-11-20 (usage-policies). This page is the operator summary; the connection settings are in provider connections.
What the policies require of an intermediary
Section titled “What the policies require of an intermediary”- No bypassing safeguards. When an OpenAI model refuses a request
(
content_filter, or a usage-policy rejection), the request must not be retried on another OpenAI route. Prism holds it: the routing record carriesopenai_refusal_held(see errors). - An end-user identifier. Prism sends
safety_identifierwith every request to an OpenAI model: an irreversible per-person hash derived from the installation key. If OpenAI reports a violation, the organization administrator can look the person up by SafetyID in the organization’s member card — Prism keeps the mapping locally and sends no names to OpenAI. - No training on outputs. Nothing an OpenAI model produces — answers,
scores, translations — may become training data for another model. Prism’s
own decision models are trained without OpenAI labels or translations, and
a CI gate (
laya-provenance) fails a build that adds one. - Investigate and act. The request journal shows refusals per person; the administrator can suspend a person’s requests to OpenAI models with one action, without removing them from the organization.
Subscription accounts vs API keys
Section titled “Subscription accounts vs API keys”A Codex or ChatGPT subscription is a consumer product for one person. Serving an organization’s members through it is the “intermediary” case the policies regulate: OpenAI may require prior review and approval, safeguards “consistent with our own” (the SafetyID, refusal handling and journal above), and the ability to act on violations. An API key is the standard, unambiguous path: metered, no seat semantics, no approval question.
Practical guidance:
- Prefer API keys for anything beyond light personal use of a team.
- Do not share one subscription account with several people; Prism forbids sharing any subscription connection for exactly this reason.
- If your organization serves OpenAI models to members through a subscription at scale, contact OpenAI for approval before relying on it.
Where to ask OpenAI
Section titled “Where to ask OpenAI”Approval questions go through the OpenAI API contact form (“I work with an organization that provides model access”). Have ready: what the product does, how many people it serves, which safeguards are in place (SafetyID, refusal hold, journal, suspension) and how usage is billed.