Build an AI inventory by defining scope, collecting candidates from several sources, recording a small set of fields per system, assigning an owner, and setting a refresh date. NIST lists an inventory mechanism under GOVERN 1.6, and the Playbook notes partial inventories offer less value than full ones.
Day 1: decide scope and owner
Name one person or team to maintain the inventory. NIST's Playbook suggests exactly that, plus a policy defining what gets inventoried. Aim for full coverage, with high-risk or high-stakes systems as the floor.
Decide what counts. Include models you train, models you call through an API, AI features inside vendor products, coding assistants, and agents with tool access.
Days 2 and 3: collect candidates
No single source is complete. Combine several:
- Procurement and expense records for AI subscriptions.
- Single sign-on and OAuth app grants.
- Network or proxy logs for known AI service domains.
- Repository searches for AI SDK imports and API key variables.
- Cloud accounts for managed model services.
- A short survey to each team lead.
Expect surprises. People use tools that were never procured.
Day 4: record the fields
Keep it small enough to finish. The Playbook mentions documentation, source code links, incident response plans, data dictionaries and contacts. The generative AI profile adds items worth considering: data provenance, known issues from internal bug tracking or public incident sources, human oversight roles, special handling for intellectual property or sensitive data, and the underlying foundation models, versions and access modes.
A workable row: name, owner, purpose, vendor or in-house, model and version, data it touches, who can use it, risk tier, links to documentation, last review date.
Day 5: tier and publish
Assign a rough risk tier to each entry (see how to tier AI systems by risk). Share the list with security, legal and the business owners and ask them to correct it.
Keep it current
An inventory decays quickly. Set a quarterly review, make adding an entry part of the procurement and launch checklists, and compare against fresh discovery data on a schedule.
Where Safeguard fits
Safeguard includes asset discovery that can help find software and services in your environment as one input to this list. It does not replace asking teams, so treat it as one source among several.
Common gaps after the first pass
First inventories usually miss three things: AI features that arrived in a software update without anyone buying them, personal accounts used for work, and agents or automations built by individual teams with API keys. Ask directly about each in your survey. Also record the status of every entry (pilot, live, paused, retired), so the list shows lifecycle and not just existence.
Quick check
Do embedded AI features in vendor software count?
The generative AI profile suggests defining any inventory exemptions in policy for such embedded features, which implies you should decide deliberately rather than leave them out by default.
Is a spreadsheet enough?
For a first version, yes. Move to tooling when the number of systems or the pace of change makes manual upkeep unreliable.
Sources
- NIST, AI Risk Management Framework 1.0 (NIST AI 100-1)
- NIST, AI RMF Playbook
- NIST, Playbook: Govern
- NIST, Generative AI Profile (NIST AI 600-1)
Last reviewed: 19 October 2026.
