Include any system where AI influences an output that people or other systems rely on, whether you built it, call it through an API, or receive it as a feature inside someone else's product. Decide edge cases in writing, because silent exclusions are where unmanaged risk collects.
The four buckets
Built by you. Models you train, fine-tune or host. These carry the most documentation: data sources, evaluation results, versions.
Called through an API. Your product or scripts send data to a hosted model. Record the provider, model name and version, what data is sent and the access mode.
Embedded in vendor software. A note-taking app with a summarizer, a CRM with a drafting assistant. NIST's generative AI profile suggests defining any inventory exemptions in policy for generative AI embedded in application software. That wording points to a deliberate choice.
Used by individuals. Coding assistants, chat tools and browser extensions that staff adopt. These are common sources of unreviewed data flow.
A simple include test
Answer three questions about each candidate:
- Does it process company, customer or personal data?
- Does anyone act on its output without checking it independently?
- Can it take actions, such as sending messages, changing records or running code?
Yes to any one means include it. Yes to the third usually means a higher tier.
Cases that cause arguments
- Spell check and autocomplete. Usually out of scope if no company data leaves the device and the output is trivial. Write the exception down.
- Search with AI summaries. In scope if staff paste internal data into queries.
- Agents and MCP servers. In scope, and worth recording the tools and permissions they hold.
- Free personal accounts used for work. In scope as a finding. The remedy may be to move users to an approved account.
What to record for an exemption
A one-line rationale, an approver, and a review date. NIST's profile also suggests listing underlying foundation models and versions in entries, so even exempted items benefit from a note about the provider.
Keeping the list usable
An inventory with every autocomplete and spell checker becomes noise, and people stop updating it. Set a minimum bar in policy, review exceptions quarterly, and mark each entry with the reason it is included. If a team wants to add a tool, a two-minute form with five fields is better than a long review. If a tool turns out to touch sensitive data later, move it from exempt to included and record the date.
Example decisions
A marketing team uses a hosted writing assistant with a company account and pastes in product plans: include, tier medium. A developer uses a local autocomplete model with no network access: exempt, with a note. An analyst connects a spreadsheet add-in that sends cells to an outside model: include, and review the data it sees. Writing three or four examples like these into your policy settles most future arguments quickly.
Quick check
Should a pilot with ten users be in the inventory?
Yes. Pilots often touch real data. Mark them as pilots and set an end date.
Who decides edge cases?
The inventory owner proposes, and the governance group or risk owner approves. Record the decision.
Sources
Last reviewed: 19 October 2026.
