OpenAI · Chat assistant
ChatGPT Enterprise
OpenAI's managed assistant for organisations — the surface with the widest non-technical adoption and the largest connector governance surface.
What you are actually governing
ChatGPT Enterprise is usually the AI surface with the most seats in a company and the least technical user base, which shapes the governance problem. The controls that matter are administrative rather than architectural: who has access, which integrations are permitted, what is retained, and what an auditor can retrieve later.
OpenAI provides a genuinely useful set of primitives here — SSO and directory provisioning, workspace administration, retention configuration, and a compliance API for programmatic access to conversation and audit data. That last item is more valuable than it sounds, because it makes eDiscovery and incident review a query rather than a support request.
Two artifact classes need explicit ownership. Connectors attach the assistant to real systems, granted through user-facing consent, which makes them privilege grants issued by people who do not think of themselves as issuing privilege grants. Custom GPTs package instructions, knowledge files and actions into something shareable — which is to say a distributable agent artifact, created inside the organisation, typically with no review process at all.
The gap that persists after every admin control is configured is the same one as elsewhere: the connector and custom-GPT inventory is a list of permissions, not an assessment. Knowing that a connector is enabled tells you nothing about what its tools can do, whether its definitions changed last week, or whether the same underlying server is also connected to a coding agent by a different team.
Artifact classes this surface loads
Each of these changes what the agent does, and each arrives through a channel with no dependency review. This is the inventory a governance programme has to cover — not a feature list.
- Connectors
- Custom GPTs
- Custom instructions
- Projects
- MCP servers
What OpenAI genuinely controls
Documented controls worth configuring fully before buying anything. Every one is cited to OpenAI's own documentation below.
- Workspace administration
Central control over membership, roles, and which features and integrations are available to users.
- SSO and SCIM provisioning
Identity and lifecycle managed through the organisation's directory rather than individual signups.
- Compliance API and audit logs
Programmatic access to conversation and administrative event data for eDiscovery, review and SIEM ingestion.
- Retention configuration
Organisation-controlled retention, with enterprise terms that differ materially from consumer plans on training use.
What is still open once those are configured
Not criticism of the product — the honest boundary of a per-machine configuration model. These are fleet questions, and local settings were never designed to answer them.
- Connectors are permitted, not assessed
Admin controls decide availability. They do not evaluate what a connector's tools can do or detect when its definitions change.
- Custom GPTs are ungoverned internal artifacts
Instructions, knowledge files and actions bundled and shared internally, usually with no review, no versioning and no owner.
- Personal-account usage is out of scope
Everything here governs the managed workspace. Consumer accounts on personal devices remain invisible.
- No shared artifact registry
The same MCP server or knowledge source is often used by other AI surfaces in the company, each governed separately and none with a complete view.
Frequently asked
- Does the compliance API give us everything we need for an AI incident review?
- It covers conversations and administrative events within this workspace, which is a strong start. It does not cover what other AI surfaces did, or which shared artifact versions were active — cross-surface attribution has to be assembled elsewhere.
- Are custom GPTs a supply-chain risk?
- Yes, an internal one. A custom GPT is a bundle of instructions, knowledge and actions distributed to colleagues — the same artifact shape as a skill, usually with even less review.
Sources
Compiled from OpenAI's own documentation on 2026-07-25. Products in this category change quickly — verify against the source before a decision depends on a specific mechanism.