Skip to content

Discovery operations

Discovery helps identify AI that may not yet be registered. It produces signals requiring human review, not automatic declarations that a person or team is using an unauthorised system.

StageQuestion
CoverageWhich business areas, identities, cloud services, repositories or imports are observed?
DetectionWhich built-in or approved custom rules interpret the signals?
Candidate reviewIs the signal a real AI system, duplicate, permitted tool or false positive?
ReconciliationDoes it match an existing system, supplier, model or prior candidate?
Governance actionShould it be promoted, linked, dismissed, monitored or escalated?
AssuranceCan coverage, decisions and unresolved gaps be explained to management or audit?

Define the source owner, purpose, scope, cadence and lawful basis before collection. Use the smallest read scope that can produce the required signal. Store connection credentials through the approved tenant-scoped integration mechanism; do not paste credentials into candidate notes or imports.

Test a source before relying on its coverage. A healthy connection does not prove full coverage if important accounts, regions, repositories or environments are excluded.

Built-in and custom rules normalise raw indicators into reviewable candidates. Document what a rule detects, expected false positives, risk labels and the owner who can change it. Preview an import or rule change before a live scan where that option is available.

Monitor last run, next run, errors, volume and unusual changes. A sudden zero may indicate a clean environment, but it can also indicate failed collection. A sudden spike may reflect a deployment, rule change or duplicate ingestion.

For each candidate:

  1. Inspect source, detector, confidence, rationale and supporting artifacts.
  2. Search for an existing system or related candidate.
  3. Confirm owner and intended use with an accountable person.
  4. Promote a real system into the Registry or link it to the matching record.
  5. Dismiss a false positive with a reason that another reviewer can understand.
  6. Escalate material unknown use, prohibited tools or sensitive-data exposure.

Alerts should have a named disposition and owner. Bulk actions are appropriate only when the same decision basis genuinely applies to every selected item.

Reconciliation prevents duplicate inventory and preserves the chain from raw artifact to governed record. Suggested matches assist review but do not replace confirmation. Discovery playbooks can create or link follow-up risks and evidence work; confirm the target system and workspace before execution.

A defensible Discovery report explains sources, coverage, run health, rule set, candidates, decisions, unresolved artifacts, overdue alerts and limitations. It must not imply that “no candidates” means “no shadow AI” unless coverage and detector effectiveness support that conclusion.

  • Sources have owners, purpose, scope and review dates.
  • Credentials use least privilege and are not exposed in records.
  • Rules are tested and false-positive assumptions documented.
  • Failed and overdue scans are investigated.
  • Candidates have attributable decisions.
  • Promotions enter normal intake and routing.
  • Unresolved artifacts and coverage gaps are reported.