What is an AI system inventory?
An AI system inventory, also called an AI register or AI asset inventory, is a maintained record of every AI system an organisation uses or builds. Each record states the owner, purpose, vendor, models, deployment, the data that can reach it, how outputs are reviewed, its risk and its approval status. Risk assessment, approval, enforcement and audit all start from it.
Frameworks treat it as a base control. The NIST AI Risk Management Framework, under GOVERN 1.6, expects mechanisms to be in place to inventory AI systems, resourced according to organisational risk priorities. In practice the hard part is not the format. It is keeping the facts true.
Why "unknown" must be a value
A first inventory is full of gaps. A tool seen in proxy logs has no known owner. A SaaS vendor cannot say which model sits behind a feature. Nobody has checked whether personal data reaches it.
If the field is left blank, or filled with a hopeful "no", the gap disappears from view. A blank personal data field reads as "none" in a report, and a system with no recorded risk looks low risk. Make "unknown" an explicit value in every field that matters, and treat it as a reason for review, never as safe. Counting unknowns per system is also a fair measure of how mature the inventory is.
The fields that matter
| Field | Why it matters | Who supplies it |
|---|---|---|
| System name | A stable label people recognise, used in approvals and reports | Business owner |
| Business owner | Someone accountable for the use and its outcomes | Governance lead assigns, owner accepts |
| Technical owner | Someone who can answer how it is built, configured and connected | Engineering or IT |
| Purpose | Risk depends on use: the same model is low risk for drafting and higher risk for decisions about people | Business owner |
| Vendor | Links the system to vendor due diligence and contracts | Procurement |
| Models | Model changes can change behaviour, cost and data terms | Technical owner, vendor |
| Deployment | SaaS feature, API, self-hosted or agent: each has different controls | Technical owner |
| Data categories that can reach it | The main input to privacy and security risk | Business and technical owners |
| Personal data (yes, no, unknown) | Triggers privacy review and, where relevant, a DPIA decision | Business owner, DPO confirms |
| Human review of outputs | Shows whether a person checks outputs before they have effect | Business owner |
| Linked agents or MCP servers | Tools that can delete or run commands raise the risk of the whole system | Technical owner |
| Provenance (observed, declared) | Shows whether the record rests on evidence, a statement, or both | Security, owner |
| Risk level | Sets the depth of review and the review cycle | Risk or governance function |
| EU AI Act classification by the organisation | Records the organisation's own determination, made with counsel | Legal, compliance |
| Last assessment date | Shows how fresh the facts are | Governance function |
| Approval status and end date | Makes approval explicit and forces review when it expires | Approver |
| Notes | Conditions, exceptions and open questions | Anyone, reviewed by the owner |
Observed versus declared records
Records arrive in two ways. An observed record comes from evidence: proxy or DNS logs, a gateway, an endpoint inventory. It proves use but says nothing about purpose. A declared record comes from a person who states the purpose, owner and data. It explains the use but proves nothing on its own.
Keep both and say which is which. An observation should become a full record only when a person confirms it is a real AI system in use and supplies the owner and purpose. A declared system that never appears in traffic is also worth a question: it may be unused, or it may be reaching the provider by a route you do not see.
How to populate it
- Start from logs. Match a few weeks of proxy and DNS logs against a list of AI providers. Each match is an observed candidate.
- Add gateway traffic. If applications call models through a gateway, each API key, team or application is a candidate with models and volumes already known.
- Check procurement. Review contracts and renewals for AI services and for AI features added to tools you already pay for.
- Collect declarations. Send a short form to teams asking what AI they use, for what and with which data. Keep it to a few questions.
- Reconcile. Merge observed and declared candidates, assign owners and mark every gap as unknown.
- Classify and decide. Assess risk from the recorded facts, then approve with conditions and an end date, replace or block.
How to keep it current
- Feed new observations in continuously rather than in a yearly survey
- Make a model change, a new data category or a new linked agent trigger a review
- Give every approval an end date
- Tie the procurement and renewal process to the inventory
- Keep the history of each record, so you can show what was true when a decision was made
- Report unknowns and expired approvals to the governance forum
Common mistakes
| Mistake | Consequence | Better practice |
|---|---|---|
| Listing vendors, not systems | One vendor hides several uses with different risks | One record per system and purpose |
| Blank fields | Gaps read as safe in reports | Use "unknown" and review it |
| Survey only | Undeclared use never appears | Combine logs, gateway and declarations |
| Ignoring AI inside SaaS | Features switched on after approval go unrecorded | Review approved tools for new AI features |
| No agents or tools recorded | The riskiest capabilities are missing | Record linked agents and MCP servers with their tools |
| Approvals without end dates | Records go stale and nobody reviews them | Set an end date and a review trigger |
Checklist
- One owner for the inventory, and a business and technical owner per record
- "Unknown" accepted as a value and treated as a reason for review
- Provenance recorded as observed, declared or both
- Logs, gateway, procurement and declarations all feeding it
- Agents and MCP servers linked to the systems that use them
- Risk level and the organisation's own regulatory classification recorded
- Every approval with conditions and an end date
Download the template
Download the AI system inventory template (CSV). It contains the columns above and two clearly fictional example rows, one declared and one observed, so you can see how "unknown" is used. Open it in any spreadsheet tool and replace the examples with your own systems.
Sources: NIST AI RMF Playbook, GOVERN 1.6. Regulatory classification is a determination for the organisation and its counsel; this guide is not legal advice.
How MERIDVAR keeps the inventory
MERIDVAR is software you install and operate on your own network, so the inventory stays on your infrastructure and nothing is sent to us.
- Two sources. Systems are observed in gateway traffic, uploaded proxy and DNS logs and endpoint agents, or declared by a person, with owner, purpose, vendor, models, lifecycle status and provenance.
- People confirm. An observation becomes a record only when a person confirms it.
- Agents and tools. Agents have signed identities, and MCP servers are declared with tool-level risk. A tool that can delete or run commands raises the risk of every linked AI system.
- Unknown is never safe. Rules-based classification gives a reason for every factor, and approvals read "review required" when facts change.
The product is in beta. See AI governance and the product overview.