Governance Mandate, Inventory and Decision Rights · Lesson 2
Build and classify an AI system inventory
10 min read
Lesson objectives
Analyze defining the inventory unit and linking each system to purpose, owner, data, model, interfaces and affected decisions and justify an evidence-proportionate decision.
Reading tools
Please sign in to use all the tools
Once you sign in, the reading tools are yours: larger or smaller text, typeface, line spacing, focus mode, highlighting and bold, and notes saved to your account and gathered in My study.
An AI inventory is neither a purchasing list nor a catalogue of model names. It is an operational map showing where computational capability enters a service or decision, who can explain that intervention, and who can stop it. The first design choice is the inventory unit. One unit may be a service composed of a model, a user interface, a retrieval store, and downstream rules. A shared component may instead deserve its own record when many services depend on it and its failure can propagate. Recording every component as a separate system inflates totals and obscures accountability; recording several unlike applications under one platform label conceals differences in purpose, population, and consequence. Use the decision boundary as the practical test. Keep elements together when they jointly produce one outcome with a common operator and similar operating limits. Split them when purpose, affected parties, approval authority, or failure response differs. Give every record a durable identifier that survives project renaming. Capture the actual function, influenced decision, direct user, affected parties, inputs, model source, interfaces, dependencies, deployment environment, service owner, risk owner, and lifecycle state. A binary “uses AI” field is inadequate because capability may be embedded in an ordinary product or enabled later by a supplier update. A usable record must answer three verifiable questions: where is the capability operating, what does it change, and who can intervene when it behaves badly? Store the discovery date and verification method with each answer. A product owner’s statement is useful testimony, not conclusive evidence; compare it with configuration, a data-flow diagram, a contract, or an observed transaction. Network traffic can reveal an undeclared service but does not establish purpose or authorization. The register should therefore preserve disagreement among discovery sources until a reviewer resolves it. Its managerial value lies in exposing orphaned services, duplicate deployments, purpose drift, concentrated supplier dependencies, and decision paths with no challenge route. When designed this way, the inventory supports budget choices, assurance work, and incident response instead of becoming an annual spreadsheet forgotten after sign-off.
Treat inventory maintenance as an information production process with explicit gates. Discovery gathers signals from procurement, identity administration, code repositories, data catalogues, cloud accounts, supplier registers, and structured interviews with teams. Resolution determines whether several signals describe one service or distinct operating units and assigns a persistent identifier. Enrichment asks the accountable owner to document purpose, users, data flow, decision point, and available intervention. Corroboration compares that account with a second source such as active configuration, a contract schedule, or a walkthrough of a real transaction. Classification then uses separate axes rather than a single vague label. Useful axes include consequence severity, exposure volume, reversibility, degree of autonomous action, data sensitivity, supplier dependence, and evidence maturity. A published rule may combine them into a priority tier, but the original values must remain visible so a reviewer can reconstruct why the tier was assigned. One defensible rule sends any service that performs a hard-to-reverse action or materially affects a person directly to deeper review even when volume is low. An internal drafting assistant may begin with lighter operational review, yet it still requires confidentiality and data-rights checks. Treatment is the next gate: every tier produces a named action such as completing a test, constraining scope, assigning an owner, revising a contract, or pausing operation. Reconciliation closes the loop. Compare the register with fresh discovery signals and investigate records with no activity, activity with no record, departed owners, and changed versions without review. Report decision-useful measures: the share of detected services with complete records, the share corroborated by an independent source, the age of the last verification, and the count without an accountable owner. Avoid celebrating “95 percent complete” when easy administrative fields are filled but purpose or stop authority remains blank. Inventory quality is demonstrated by the decisions the record enables, the discrepancies it surfaces, and the time required to resolve them—not by the number of columns or the color of a dashboard.
Lesson summary
The lesson connects defining the inventory unit and linking each system to purpose, owner, data, model, interfaces and affected decisions to a measure, owner and review condition, recognizing that classification prioritizes oversight; it is not a final judgment of safety or legality.