The shape of a DLP policy
A DLP policy = locations × conditions × actions:
- Locations: Exchange (mail), SharePoint/OneDrive (files), Teams (chat/channel), endpoints (Endpoint DLP — copy/paste, USB, print on managed devices), the Copilot location (excluding labelled content from AI grounding — the Copilot curriculum), and increasingly non-M365 sources via DSPM.
- Conditions: content matches SITs/labels/classifiers (the classification foundation), sensitivity thresholds, recipient/sharing context.
- Actions: block, block-with-override (user justifies), audit-only, encrypt, notify, restrict access.
The order and precedence (mirrors the other workloads)
DLP evaluates within each location's pipeline (in Exchange, it rides the transport stack — the EXO curriculum's EOP order). Multiple policies resolve by priority. The design rule: start in audit / test mode everywhere, read what WOULD have blocked for weeks, tune, THEN enforce — flipping straight to block is how you take down a legitimate business process on day one.
The false-positive economics (the real work)
A DLP rule matching a generic pattern (any 16 digits looks like a card number) fires on invoices, reference numbers, and product codes. Uncontrolled, users learn to click 'override' reflexively (defeating the control) or route around it (personal email, screenshots — the endpoint/BYO-AI gap). Controlling this is the job:
- EDM and higher-confidence SITs cut noise (match YOUR data, not any pattern).
- Context conditions (only when shared externally, only above N instances) focus the policy on real risk.
- Override-with-justification on medium risk keeps flow while capturing intent; hard-block reserved for the genuinely sensitive.
- Review override/false-positive rates monthly — flat means tuned, rising means the policy is training users to ignore it.
What to watch (proofs)
- Simulation first: run new policies in test mode and read the would-have-blocked report — the safe way to size impact before enforcing.
- Match quality: DLP alert volumes by rule and their true/false-positive split — the tuning backlog; a rule that's 95% false positive is worse than no rule.
- Coverage of locations: which locations each policy covers — endpoints and the Copilot location are the ones brownfield estates most often miss.
- The route-around signal: BYO-AI / personal-channel usage rising in populations under strict DLP (DSPM/insider risk) — the cost of over-blocking, measured.