An AI model inventory is a single registry listing every AI integration your organization runs — what it does, who owns it, what model and platform it uses, how risky it is, and when it was last reviewed. It is the minimum viable audit artifact: nearly every governance activity (validation, change management, incident response, decommissioning) either reads from it or writes to it. A spreadsheet is a perfectly good first implementation. What matters is the fields, the update discipline, and reconciliation against reality.
The fields that earn their place
| Field | Why it's there |
|---|---|
| System name & ID | Stable reference for reviews, incidents, and change records |
| Owner (person, not team) | Who answers questions and executes changes |
| Use case & audience | What it does; internal vs. customer-facing drives risk |
| Risk tier | Sets review depth and cadence — see risk tiering |
| Platform | Claude API, Claude Platform on AWS, Bedrock, Vertex AI, or Foundry — determines whose data-handling terms apply |
| Exact model ID | e.g. claude-sonnet-5 or Bedrock's anthropic.claude-sonnet-5; pinned version identity for validation and deprecation tracking |
| Workspace / account reference | Ties the entry to the billing and access unit it runs in |
| Data sensitivity | Highest classification of data the system touches |
| Status | Proposed / in review / live / suspended / decommissioned |
| Last review date & next review due | Makes staleness visible and schedulable |
Resist adding twenty more columns on day one. Every field is a maintenance liability; add fields when a real review needed one and couldn't find it.
Why the model ID column matters more than it looks
Model IDs are pinned snapshots — even the dateless IDs introduced with the Claude 4.6 generation are pinned, not evergreen pointers — so the inventory's model column tells you exactly which behavior was validated. It also makes vendor lifecycle events actionable: Anthropic publishes deprecation and retirement dates (for instance, the deprecated claude-opus-4-1-20250805 retires on August 5, 2026, with Opus 4.8 as the recommended migration), and a one-line query against your inventory turns that announcement into a work list of affected systems. Note that IDs vary by platform for the same model — Bedrock prefixes IDs with anthropic., and some older models use dated forms — so record the ID as your platform expects it (see model IDs across platforms).
Keeping it true: reconcile against platform data
Inventories rot because they rely on humans remembering to update them. The fix is periodic reconciliation against what the platforms report. On the Claude API, two mechanisms help. Workspaces give each project or environment a separate unit with its own scoped API keys — if your convention is "one inventory entry, one workspace," then the Admin API's workspace list is a machine-readable shadow of your inventory, and a workspace with no matching entry is shadow AI by definition. Second, the Admin Usage and Cost API breaks usage down by workspace, model, and API key: an inventory entry marked "decommissioned" that still shows tokens, or a workspace consuming a model the entry doesn't list, is a discrepancy worth a ticket. On Bedrock, Vertex AI, and Foundry, the equivalent reconciliation runs against the cloud provider's own usage and logging tooling.
Process hooks: how entries stay current without nagging
Wire the inventory into the processes that already run. The intake process creates the entry (status: proposed). Pre-launch validation flips it to live and stamps the review date. Change management updates model ID and prompt version fields. Decommissioning closes it out — on the Claude platform, archiving the workspace (which immediately revokes all its API keys) is a satisfying technical mirror of the status change. If the inventory can only be updated through these hooks, it stays honest by construction.
Where to go next
If you are starting from zero, the enumeration step in building a governance program comes first. For the broader registry concept beyond models — datasets, prompts, vendors — see the AI registry.