The four interactions
- Grounding-time exclusion (with DLP-for-Copilot): labeled content can be barred from answers — the previous concept, the enforcement rail.
- Encryption/usage rights: label-encrypted content requires the asker to hold EXTRACT rights for Copilot to use it. No extract right = the file opens for reading but never grounds an answer. This works with NO DLP policy at all — encryption labels are self-enforcing against Copilot.
- Label inheritance on output: content Copilot creates FROM labeled sources takes the strictest source's label in supported flows (a summary of a Confidential doc lands Confidential) — the anti-laundering property: you cannot wash a label off by asking for a summary. Verify per surface; inheritance coverage has grown release by release.
- Answer-time visibility: citations and referenced items surface their labels in the UI — users SEE they are drawing on Confidential material, which is half the awareness battle for free.
Design moves that exploit the interactions
- Put encryption with narrow EXTRACT on the truly explosive content (M&A, board) — protection that doesn't depend on any Copilot-specific policy existing.
- Use DLP-for-Copilot on the broad sensitive tier (Confidential) where encryption everywhere would break workflows.
- Test inheritance paths your users actually use (summarise-in-Word, chat-across-files, Pages) and document which carry labels — the gap list feeds awareness training.
- Auto-labeling remains the coverage engine (previous concept) — all four interactions are label-powered, so label coverage is the master metric.
What to watch (proofs)
- The extract-right demo: two test users (with/without EXTRACT) on one encrypted doc, same prompt — the self-enforcing claim, proven.
- Inheritance audit: generate from labeled sources across your main flows; record which outputs carried labels — YOUR tenant's inheritance map, dated (it improves over releases; retest quarterly).
- Label analytics: coverage and label-change events (Purview) — including the downgrade-after-generation pattern that suggests users laundering manually.
- The visibility check: citations showing labels in the pilot users' real answers — screenshot for the awareness deck.