Data governance and access control

Govern AI like you govern your data

Rolling out ChatGPT across an organization opens a new channel to your data. AI governance isn’t a separate project; it extends the governance you already apply to your business tools. Three pillars hold it together—classification, access control, and traceability—and none works without the other two.

Classify before you configure

Before you touch the admin console, know what data your teams handle and how sensitive it is. In practice, four levels are enough.

Level 1, Public, covers information already available outside the company: sales brochures, press releases, product documentation. No restrictions. Level 2, Internal, covers material intended for employees: meeting notes, procedures, training materials. You can use this content on a professional plan—Business, Enterprise, or Edu—but never from a personal account.

Level 3, Confidential, includes customer data, contracts, and non-public financial information. You can still use it on a professional plan, preferably Enterprise for its adjustable retention and logs, but with precautions: go through connectors, which respect source permissions, rather than copying documents into prompts.

Level 4, Restricted, gathers trade secrets, health data, and sensitive personal data under the GDPR (General Data Protection Regulation). Do not send these to ChatGPT without a dedicated legal and security review; when in doubt, handle them outside the tool.

Where access control really happens

ChatGPT Enterprise inherits permissions from your existing systems but adds a layer you must manage directly. Model access is controlled in workspace settings and by role: an admin can disable a model, and the Pro option, powered by GPT-6 Astra, has its own access control. Reserving the most expensive options for teams that need them controls both costs and usage.

Creation and sharing of GPTs are also role-based, and groups determine who can access a shared GPT. However, GPTs will be retired on December 11, 2026, in favor of plugins, or on February 11, 2027, for workspaces with an exemption: your governance must already cover plugins and the skills that replace them.

Document access via Google Drive or SharePoint connectors follows existing permissions: no one reaches a file in ChatGPT that they cannot access at the source. This guarantee holds only if your sharing settings are clean to begin with.

Finally, a shared project opens with either conversation access or edit access, granted to individuals or groups. ChatGPT exposes all its content—conversations, files, instructions—so anyone added later sees the full history: add people deliberately.

Audit and trace

On Enterprise and Edu, the Compliance Platform provides access to workspace logs and metadata—audit, authentication, applications, and conversation messages if the owner allows—for ingestion into your eDiscovery, data loss prevention (DLP), or SIEM tools. Compliance logs are immutable, and messages from locked conversations appear there too: the code that locks a conversation protects access inside ChatGPT, not access granted by the Compliance Platform.

Alerts then land in the console your security team already uses, not in a dashboard nobody checks. Usage statistics in the Admin Console complement this monitoring without exposing conversation content.

Formalize the policy

An AI governance document fits on a few pages and covers five points: scope (which AI tools are allowed); classification (which data can be used in each); responsibilities (split among IT, the DPO, managers, and users); incident response (what to do if an employee enters restricted data); and review frequency (at least quarterly).

This split is not theoretical. The DPO validates GDPR compliance, IT manages infrastructure—SSO, SCIM, connectors—workspace admins and group managers oversee daily use, and AI champions, whom we’ll revisit in Lesson 9, spread best practices across each team.

Shared assistants—GPTs, plugins, skills—need special handling because they encapsulate instructions and sometimes durable reference data. Before opening them widely, run them through four steps.

1. The creator documents the objective, instructions, and reference data.
2. A workspace admin checks compliance with the governance policy.
3. IT steps in if the assistant accesses sensitive data sources.
4. Publication happens with an identified owner responsible for maintenance.

What this can look like

Here’s a fictional example for a 500-person company organized into groups.

This setup gives each team the tools it needs without exposing other teams’ data. The two most exposed groups, Executive and Finance, forgo document connectors: convenience yields to risk here. None of this slows adoption; governance secures it.

Your turn

Classify ten data types from your daily work. List ten data types your teams handle weekly, assign each one of the four levels, then note the required setting for each level: plan, role, connector, memory. The disagreements that surface between IT and business units are exactly what your policy must resolve.

Official sources: Managing workspace settings in ChatGPT Enterprise · ChatGPT Enterprise and Edu models and limits · OpenAI Compliance Platform for Enterprise and Edu customers · Projects in ChatGPT · SharePoint app and setup in ChatGPT (OpenAI Help Center)

Key takeaways

• Classify your data into four levels: Public, Internal, Confidential (professional plan, with precautions), Restricted (not in ChatGPT without a dedicated review).
• Access control applies to models by role, shared assistants, permissions inherited from connectors, and the scope of shared projects.
• On Enterprise and Edu, the Compliance Platform feeds logs and metadata into your eDiscovery, DLP, or SIEM tools.
• GPTs will be retired on December 11, 2026: extend your governance to plugins and skills now.
• Any widely shared assistant goes through a four-step validation and keeps an identified owner.