Enterprise AI assistant glossary
Plain-English definitions for IT identity, data-retention, and licensing reviews.
- Inherited permissions vs shared-credential scope
- Inherited permissions restrict answers to records the signed-in user can view in the connected app; custom connectors that use a shared credential reach whatever that credential can access (rather than restricting access to each individual user’s source permissions). Compare connector ACLs · Claude connector source.
- Connector visibility modes & full-crawl lag (Copilot connectors)
- Microsoft Learn states connectors apply either “Only people with access to this data source” (source ACLs mapped to Microsoft Entra ID) or “Visible to everyone” (tenant-wide access), and changing visibility requires deleting and recreating the connection. Incremental crawls do not update ACL permission changes or deletions; periodic full crawls are required and ACL changes take up to 24 hours. Microsoft connector source.
- No model training by default
- Contractual commitment that customer prompts, retrieved files, and outputs are not used to train foundation LLMs (stated for Microsoft 365 Copilot, ChatGPT Enterprise unless explicitly opted in, Claude Enterprise, Gemini Enterprise, and Glean). A no-training rule does not by itself define how long prompts or conversations are stored. Microsoft · OpenAI · Anthropic · Google.
- Zero LLM data retention
- Commitment that prompts and outputs are not stored by the LLM provider after generating a response (explicitly documented for Glean across every deployment). Contrast this with workspace admin-controlled retention windows (OpenAI deletes removed conversations within 30 days unless legally required; Anthropic Enterprise documents custom data retention controls). Glean source · OpenAI retention source.
Discovery versus access
A discovery control limits what appears in search or an assistant. It can leave a person’s permission to open the original file unchanged. See the documented example →