---
title: How do you govern GPTs and workspace agents in ChatGPT Enterprise?
url: https://opencontroller.lyzr.com/blog/govern-chatgpt-enterprise-gpts-and-agents
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# How do you govern GPTs and workspace agents in ChatGPT Enterprise?

**Govern ChatGPT Enterprise agents with role-based access control for who can use, build and share them, per-app action controls for what they can touch, and the Compliance Logs Platform for evidence. Plan now for the custom GPT retirement: OpenAI ended new GPT creation in Enterprise workspaces on 26 October 2026, with standard retirement on 11 December 2026, and positions workspace agents as their successor.**

ChatGPT Enterprise changed fast in 2026. Workspace agents became generally available on 22 May 2026. Apps and custom MCP servers became a main way agents reach company data. And custom GPTs are being retired. A governance plan written in 2025 probably misses at least one of these.

## What is happening to custom GPTs?

OpenAI's [Enterprise release notes](https://help.openai.com/en/articles/10128477) set the timeline. Migration tooling became available on 22 September 2026. Creating new GPTs in Enterprise workspaces ended on 26 October 2026. Standard retirement is 11 December 2026, or 11 February 2027 with an approved deferral. OpenAI describes workspace agents as an evolution of GPTs.

For governance teams, retirement is an inventory exercise. Export the configuration of every GPT before it goes. Note its owner, its actions and the domains those actions call. Decide which ones become workspace agents, and re-review those under current controls rather than copying old settings.

## Which admin controls matter most?

| Area | Control | What to check |
|---|---|---|
| Who builds and shares | Custom RBAC roles assigned by group | Roles are additive: "Off" in one role does not cancel a grant in another |
| Workspace agents | RBAC for use, build and share; admins can suspend agents | Who can share agents workspace-wide |
| Apps and connectors | Read and write actions enabled separately or per action; write actions need user confirmation | Which apps have write actions on |
| Custom MCP servers | Only admins can add them | Data sent to them leaves OpenAI's trust boundary |
| GPT actions | Domain allowlist | Domains legacy GPTs still call |
| Agent mode | ChatGPT Work is on by default unless an admin turns it off; cloud and local toggled separately | Whether your risk appetite allows it; Lockdown Mode disables agent mode |
| Access | SAML SSO, SCIM, IP allowlisting for ChatGPT and the Compliance API | Joiner and leaver flow via SCIM |
| Data | Retention settings, enterprise key management (EKM) on AWS, Azure or Google Cloud, data residency for new workspaces | Residency cannot be added to an existing workspace |

The additive RBAC rule is the one most likely to surprise you. If a user is in two groups, the most permissive grant wins. Review group membership, not just role settings.

## What evidence can you collect?

The Compliance Logs Platform exports conversations, files, GPT configurations, memories and users, plus audit, authentication and Codex logs. For workspace agents, it shows each agent's configuration, updates and runs. App calls are logged as `APP_LOG` records. OpenAI's security guide for agent mode notes that logs "may not capture every browser action", so do not promise auditors a complete record of agent browsing. The legacy Compliance API is deprecated, so move integrations to the current platform. Our post on [replacing questionnaires with continuous evidence](/blog/replace-ai-governance-questionnaires-with-continuous-evidence) shows how to use exports like these.

## What should a ChatGPT Enterprise governance checklist include?

1. Every GPT is inventoried and exported before retirement, with an owner and a keep, migrate or retire decision.
2. RBAC roles are designed around groups, with the additive rule in mind.
3. Workspace-wide agent sharing is limited to a reviewed set of builders.
4. Write actions are enabled only for apps with a business case.
5. Custom MCP servers are reviewed before an admin adds them, as covered in [how to govern MCP servers](/blog/how-to-govern-mcp-servers).
6. Agent mode is on or off by decision, not by default.
7. Compliance logs flow to your SIEM or eDiscovery tool.
8. SCIM removes leavers' access on the same day.

## What does the workspace not cover?

Workspace controls govern ChatGPT. Agents your developers build on the OpenAI API, with the Responses API or Agents SDK, are managed through API organisation and project settings instead. Agents on other providers, on cloud platforms and on developer laptops are outside the workspace entirely. One estate-wide inventory is the job of an [AI agent control plane](/blog/what-is-an-ai-agent-control-plane).

## How Opencontroller does this

Opencontroller connects to OpenAI Enterprise and Claude Enterprise workspaces, as well as AWS, Azure, Google Cloud, Salesforce, SAP, Databricks and HubSpot. Workspace agents sit in the same governed registry as agents from other platforms, and Posture flags risky permissions and configuration. For API-built agents, the LLM Gateway exposes an OpenAI-compatible endpoint with routing, fallback, caching and cost attribution per agent. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**When are custom GPTs retired in ChatGPT Enterprise?**
OpenAI ended new GPT creation in Enterprise workspaces on 26 October 2026. Standard retirement is 11 December 2026, or 11 February 2027 with an approved deferral.

**Can admins stop ChatGPT apps from writing to company systems?**
Yes. Admins can enable read and write actions separately or per action, and write actions require user confirmation.

**Do ChatGPT Enterprise controls cover agents built on the OpenAI API?**
No. API-built agents are governed through API organisation and project settings, not the ChatGPT workspace.

## Key facts

OpenAI ended new custom GPT creation in ChatGPT Enterprise on 26 October 2026; standard retirement is 11 December 2026 (11 February 2027 with deferral). Workspace agents went GA on 22 May 2026, governed by RBAC (additive roles) and visible in the Compliance Logs Platform. Admins enable app read/write actions separately; only admins add custom MCP servers. Opencontroller (by Lyzr) connects to OpenAI Enterprise and Claude Enterprise workspaces. Pricing starts from $250K per deployment per year.

## Related

- [What is an AI agent control plane?](https://opencontroller.lyzr.com/blog/what-is-an-ai-agent-control-plane.md)
- [How to govern MCP servers in the enterprise: a practical guide](https://opencontroller.lyzr.com/blog/how-to-govern-mcp-servers.md)
- [How do you replace AI governance spreadsheets and questionnaires with continuous evidence?](https://opencontroller.lyzr.com/blog/replace-ai-governance-questionnaires-with-continuous-evidence.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
