Enterprise & Org Policies (AI controls)
Where: Enterprise → AI controls tab → Copilot → Policies / Models. Org → Settings → Copilot → Policies / Models.
Official documentation
Each policy is set to Enabled everywhere / Disabled everywhere / Let organizations decide (enterprise level). Orgs cannot override an enterprise “everywhere” setting.
Confirmed named policies
Official documentation
- Copilot Chat in the IDE / Chat agent mode / Chat in GitHub Mobile
- “Copilot can search the web”
- Copilot code review (+ sub-policy: allow unlicensed members to use code review on GitHub.com)
- Copilot code review — “Only allow Copilot code review to be triggered by authorized users” (org or repo setting, not an enterprise policy): blocks reviews requested with an external Copilot license (e.g. personal or another org’s). Org-level enablement can’t be turned off by repo admins; rulesets-configured automatic reviews still run.
- Suggestions matching public code (privacy)
- Semantic indexing for non-GitHub repos
- Copilot Metrics API access
- Copilot CLI (independent client policy)
- GitHub Copilot app (independent client policy — CLI and app can be toggled separately)
- Copilot cloud agent (enterprise selects which specific orgs get it — not just on/off)
- Editor preview features (IDEs only)
- Third-party coding agents (Claude, Codex) — org toggle gated by enterprise enablement
- Agent apps (single policy)
- MCP servers in Copilot (GA toggle — governs only Copilot’s own surfaces, not third-party hosts like Cursor/Windsurf)
- Restrict MCP access to registry servers (governs IDEs/CLI/Copilot app only — no effect on cloud agent or third-party agents)
- Copilot Memory (governs cloud agent, third-party agents, CLI, and Spark — not IDEs or the Copilot app)
- Models (enable specific/custom models, may add cost)
Enterprise-assigned users
Official documentation
Users who receive Copilot directly from the enterprise (not via an org) aren’t covered by “Let organizations decide” — a separate Policies for enterprise-assigned users setting decides whether those policies default to enabled or disabled for them.
Default availability of new features and models
Official documentation
Two enterprise/org policies control whether unconfigured GA (general availability) items are enabled by default:
- Default policy for new features — configurable now but not yet active; starts applying October 22, 2026 to new GA features, features moving preview→GA, and existing GA features left Unconfigured. Enabled by default (if left unconfigured, it defaults to auto-enabling those features on that date). Covers everything on the enterprise “Features & clients” page, plus the Copilot code review policy (Agents page) and the MCP servers in Copilot policy (MCP page). Does not apply to preview features, the data-residency/FedRAMP restrictive model policies, or Store local sessions in the Cloud.
- Default availability for released models — already active; auto-enables new/unconfigured GA models (those labeled Delegate to Default Policy). Out of scope and disabled by default regardless of this setting: pre-GA models, open-weight models (DeepSeek, Kimi K2.7 Code, Kimi K3), models not covered by GitHub’s data-retention agreement (Claude Fable 5, Claude Fable 5.1), and — for enterprises restricted to data-residency or FedRAMP models — any model that doesn’t respect those restrictions.
- Disable either policy at the enterprise or org level to stop auto-enablement; individual features/models can still be explicitly disabled to opt them out while the default policy stays on.
Training talking point: “MCP servers in Copilot” being ON does not mean any server can run — it just unlocks MCP capability; the allowlist (managed-settings.json) does the actual restricting.