davisAIgovDavis & Associates Advisory, LLC
Connect with us

Build a dependency record you can use during an incident

A release-scoped inventory links a failed component to the workflows and decisions it can affect.

Richard E. Davis · · 4 min

An inventory becomes valuable when an owner can answer a specific question under pressure. If a model version, retrieval connector or identity provider changes overnight, which customer workflows can the organization still trust? I start with the release actually running and trace from that release to the controls that are supposed to contain it.

Join design evidence to observed activity

A software bill of materials gives the reviewer a component list. An AI system also depends on model and prompt versions, retrieval indexes, extensions, workload identities and the boundaries that authorize downstream action. I record an identifier at deployment and reuse it in the policy decision and the resulting event. The reviewer can then check whether the dependency map reflects actual runtime behavior.

Exercise the decision before it is urgent

Pick one dependency and exercise its loss or revocation. Identify which routes can remain available, who can suspend an affected use case and which evidence the incident team must preserve. I want the change owner to show the degraded workflow, the decision record and the retest.

What I would ask a board to review

Which dependencies have a named owner? Which events trigger reassessment? How long would it take to identify every affected use case after a supplier change? Each answer should point to a record that someone independent can inspect.

Sources

  1. NIST AI RMF 1.0
  2. ISO/IEC 42001:2023
  3. NYDFS 23 NYCRR Part 500

Related sector · Financial services and insurance · All Perspectives