What counts, and what the record holds
Lesson 2 of 5 in AI Inventory, Use-Case Intake, and Risk Triage.
Discovery immediately forces a definitional question: what earns a registry entry? Draw the line too wide and the registry drowns in spreadsheets and macros, burying the systems that matter; too narrow and the first regulator who applies the EU AI Act’s Article 3(1) definition finds systems you excluded.
Anchor on the legal definition — a machine-based system that infers from inputs how to generate outputs that influence environments (the anatomy of the AI system definition is unpacked in the foundations module What Is AI?, and we will not re-teach it here) — then resolve the edge cases with a consistent internal rule. The classic edges: a hand-written rule engine is out of scope of AI law but may still deserve tracking as an automated decision system (Robodebt was "just" arithmetic — and destroyed lives at scale); a statistical spreadsheet model used for consequential decisions is a governance question even when the AI definition is debatable; a general-purpose model and the systems built on it get separate treatment — the model appears once as an asset, each use case appears as its own entry, because the chatbot for HR and the chatbot for customer refunds carry entirely different risks.
Practical rule from working programs: registry scope should be slightly wider than the statutory definition. It costs little to track a borderline system at minimal tier; it costs an enforcement action to have excluded one wrongly.
Interactive sorting exercise: Your registry scope rule: AI systems per Art 3(1) in — tracked individually per use case; consequential rule-based automation in — flagged as ADS, not AI; pure conventional software out. Classify each discovered item.
With scope settled, design the record. The discipline is a minimum viable record — the fields you can realistically populate for every system on day one — expanding to a full lifecycle record only for the higher tiers. Ten empty columns of aspirational metadata help no one; ten filled ones answer a regulator.
| Field | What it captures | The question it answers under pressure |
|---|---|---|
System name & ID | Unique, stable identifier | Are we all talking about the same system? |
Owner | A named individual (never a team), plus business sponsor | Who is accountable right now? The single most diagnostic field — stale owners mean a dead registry. |
Business purpose | What decision or output, affecting whom | The input every risk-tiering and legal-mapping judgment starts from. |
Data used | Categories in and out; personal, sensitive, customer data flags | What GDPR/privacy exposure rides on this system? |
Model type & provenance | Built / bought / embedded; underlying GPAI model if any | Whose duties are whose across the value chain? |
Vendor & contract link | Provider identity, contract reference, update cadence | Whom do we call — and what did they promise — when it misbehaves? |
Deployment status | Idea / pilot / production / retired, with dates | Is this live on real people, and since when? |
Risk tier | Output of triage (next lessons), with date and rationale link | How much scrutiny does everything else owe this system? |
Applicable regimes | EU AI Act category, Colorado, NYC LL144, sectoral rules | Which statutory clocks are ticking on it? |
Linked artifacts | Assessments, model card, monitoring dashboard, incidents | Where is the evidence? A registry that links nothing is a phone book. |
Tooling follows the same maturity arc as the last module’s program: every registry starts as a spreadsheet, and the spreadsheet is fine — until synchronisation breaks. The migration path runs to either a GRC platform module (strong on workflow, attestations, and audit trails; weak on technical truth) or an MLOps model registry (knows exactly what is deployed, automatically; knows nothing about vendors, business purpose, or law). Mature setups wire the two together: the model registry feeds technical facts upward, the GRC layer holds governance facts, and neither is typed twice.
Whatever the tool, the registry decays without a heartbeat. The standing fix is the attestation cycle: every owner confirms, on a fixed cadence (quarterly for high tiers, annually for the rest), that their record is accurate — status, data, purpose, tier still true. Attestation catches what automation cannot: the model quietly repurposed, the owner who left in March, the pilot that became production without telling anyone.
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.