Name the owner of every inference dependency

Hand over the deployed inventory

Provide the component list with payload scope, endpoint, identity, location settings and retention rules. Link each entry to its deployed configuration and responsible team.

Include ingestion, embedding, evaluation and support systems. These dependencies can receive sensitive material without appearing in the user-facing request. A model-only inventory leaves important operational work unowned.

Record known limitations and unresolved requirements clearly. Do not let an assumption become an inherited guarantee simply because it appears in a handover document.

Rehearse a dependency failure

Have the receiving team handle an unavailable model endpoint and verify the intended fallback or unavailable state. Then ask them to investigate a synthetic payload in a diagnostic store.

The exercise should establish who owns routing, who owns the log destination and who coordinates the incident. If every team can explain its own component but nobody owns the full request, add that coordinating responsibility.

Include credential rotation and endpoint changes in the runbook. The service can fail or route differently after those operations even when application code is unchanged.

Keep the inventory alive

Require new dependencies and significant configuration changes to update the record. Review obsolete stores, credentials and fallbacks periodically rather than retaining them indefinitely for convenience.

Assign ownership of provider-documentation and agreement reviews to the appropriate people, with engineering responsible for verifying the configured controls. A complete handover leaves a clear answer to who maintains each part of the boundary and who checks that the combined system still matches its intended data handling.