An assistant can read a supplier file.
What else is it allowed to do?

OpenAI's current Company Knowledge help describes a plugin replacing the previous company-knowledge option in ChatGPT. Its instructions cover the web experience for eligible Business, Enterprise and Edu workspaces. Installation alone does not connect individual accounts or grant source-data access: supported sources, workspace availability and any required provider authorization still matter. The page says Company Knowledge respects connected-source permissions. It also says apps can include search, interactive interfaces and writes. That is not a statement that every source supports changes. Ask which app, account and capabilities the proposed supplier-document task actually uses.

Keep the separate API example separate. OpenAI's MCP guide covers the Responses API, using remote MCP servers or a tunnel to a private server. MCP, the Model Context Protocol, provides a way for an assistant to use external tools; it does not by itself establish business permission. The guide also distinguishes legacy connectors: connector_id is deprecated for models released after September 1, 2026, while existing models retain support. That is a compatibility cutoff, not the page's publication date. Neither this API guidance nor the ChatGPT help proves that a particular document store and supplier-record system are connected.

Suppose the proposed task sends a certificate excerpt to an external search service to clarify a reference. The service might make no change to a record, yet receive supplier information in the request. This is a review scenario, not an observed deployment. Ask what text leaves, which service receives it and who can authorize that disclosure. For the Responses API, OpenAI documents approval by default before data is shared with a connector or remote MCP server. Each approval applies to one tool call. Developers can configure selected tools, or all tools on a server, to skip approvals. The guide recommends approval flows for sensitive actions and warns that remote tool behavior can change. Request evidence of the proposed outgoing arguments and effective approval configuration. A visible prompt is useful evidence about a call, not proof that its recipient or contents are appropriate.

ChatGPT app permissions govern use after access is available. OpenAI says they do not grant provider permissions, change source-system access or override workspace policies. Available controls depend on the account, app, connection and workspace. The documented options include asking before reads and changes, allowing reads while asking before changes, and automatic approvals under specified conditions. Some requests may instead be denied. Do not promise that every write receives a confirmation: inspect the actual action controls and approval behavior for this connection. The personal-account and managed-workspace experiences are not interchangeable. There is another boundary: restrictions on direct app actions do not necessarily restrict retrieval from synced data, and sync content-selection settings do not necessarily restrict separate live actions. Ask the technical owner to identify which path supplies the certificate. Separately, have the business owner establish who may authorize disclosure and who may approve a supplier-record change. The cited controls do not establish those organizational assignments.

We propose one review record with separate entries for source access, outgoing context, available external actions and approval conditions. Name the exact connection and task, attach authorized configuration evidence and identify the owner of each decision. Do not include confidential payloads merely to make the review packet convincing. For a proposed evaluation, use authorized, non-confidential material to examine retrieval, a proposed outbound request and a proposed record change separately. Agree on expected allowed, denied and confirmation behavior before evaluating it. These are our suggested checks, not tests performed here or vendor-mandated acceptance procedures. Changing a ChatGPT app permission does not disconnect the account or revoke provider access already granted. The MCP guide also says third-party servers apply their own retention and residency policies. Treat disconnection, revocation and existing copies as separate follow-up questions; none was verified here. Revisit the record when accounts, tools or configurations change.