The AI model inventory your auditor will ask for
Under the EU AI Act and most enterprise security reviews, "we use some LLMs" is not an answer. Here is the minimum viable inventory and how to keep it current automatically.
The question you will be asked
Sooner or later — an enterprise security review, an ISO audit, an EU AI Act readiness assessment, or a customer's procurement questionnaire — someone will ask: *which AI models does your product use, for what purpose, on what data?*
Most scaleups cannot answer. Not because they are careless, but because the answer is distributed across a dozen repositories, three teams and a handful of API keys nobody has audited since they were created.
What a minimum viable inventory contains
One row per **use case**, not per model. A single model used for three different purposes carries three different risk profiles.
For each row:
- **Use case name and owner.** A named person, not a team mailbox.
- **Model and provider,** including the exact version or snapshot identifier.
- **Purpose,** in one plain sentence a non-engineer can read.
- **Input data categories.** Does it see personal data, payment data, health data, or none of those?
- **Human oversight.** Is there a human in the loop, on the loop, or none — and where does the output land?
- **Risk classification.** Under the EU AI Act framing: minimal, limited, high, or prohibited.
- **Deployment status.** Prototype, internal, or customer-facing.
- **Last reviewed date.**
That is eight columns. A spreadsheet is a perfectly legitimate starting point. What is not legitimate is a spreadsheet that was accurate in March.
Keeping it current without a compliance headcount
Manual inventories decay within one quarter. The fix is to derive as much of it as possible from systems that are already authoritative.
**Route model calls through a gateway.** If every LLM call passes through one control point, the list of models actually in use is a query, not an interview. Anything calling a provider directly is by definition off-inventory — and that discrepancy is itself the most valuable signal the inventory produces.
**Tag calls with a use-case identifier.** One header, set at the call site. It costs a developer thirty seconds and turns your gateway logs into an automatically maintained usage register.
**Reconcile monthly.** Compare gateway-observed use-case IDs against the registered inventory. New ID with no row: someone shipped something unregistered. Registered row with no traffic in 90 days: a candidate for decommissioning, and one fewer thing to defend in an audit.
**Attach the inventory to change management.** Adding a new use case becomes part of the same review that ships the feature, rather than a separate compliance task that gets deferred.
Why this pays for itself before any audit
The inventory is usually justified as a compliance artefact, but its first real return is commercial. Enterprise deals stall on security questionnaires. A team that can answer the AI section in an hour instead of three weeks closes faster. The same register also answers the cost question — which use case is generating which share of your model spend — because it is built on the same gateway data.
Compliance and cost control turn out to need the same underlying visibility. Build it once.