How do we inventory and classify AI systems by risk level?

Begin with the AI systems already in use. Include personal accounts, AI features inside vendor software, and departmental subscriptions that bypassed procurement. These tools can remain absent from an IT-approved software list.

Review vendor contracts for AI capabilities in existing software. Ask employees which AI tools they use in their workflows. Have IT audit network traffic for unauthorized calls to language model providers. Run these checks together; each covers gaps in the others.

Apply a consistent risk classification to the inventory. A common summary of the EU AI Act distinguishes unacceptable, high, limited, and minimal risk. Prohibited uses require exclusion. High-risk uses face strict requirements, while certain other uses carry transparency duties. Minimal-risk uses generally have no specific AI Act obligations. Teams outside the Act’s scope can use this distinction as an organizing reference. Check statutory criteria before assigning a legal classification.

Assess data sensitivity first. Does the system process personal data, health information, financial records, or another regulated category? Then assess decision impact. Does it affect hiring, credit, benefits, or access to services? Finally, check regulatory exposure. Financial services, healthcare, and employment may have specific requirements.

Use these dimensions to assign an internal risk tier. High scores across all three call for rigorous oversight. Review classifications annually and whenever a system or its use changes. An internal score does not replace the legal classification required by an applicable regulation.

Assign an owner, record the risk tier, and schedule reviews for every system. Classify new deployments before launch. Add an AI capability check to procurement and vendor management so new tools enter the inventory.

Choose an inventory tool that filters by risk tier, business unit, data type, and regulatory framework. Keep records ready to explain which systems you operate and how you oversee them.

NIST's draft guidance on public-facing AI documentation specifies that each system should be documented across several dimensions: its purpose, intended uses, performance characteristics, and known limitations. Using structured templates for these fields makes records consistent across your inventory and easier to compare during governance reviews. Documentation should be completed before deployment, not after.

The guidance also calls for audit-ready records that can substantiate claims during third-party or regulatory review, and for external-facing disclosures written in terms accessible to affected parties. Aligning your inventory records with NIST AI RMF categories, which the draft explicitly encourages, lets a single record serve both internal classification and external transparency purposes. Public comments on the draft are open until September 16, 2026, so practitioners should monitor the final version for changes to template requirements.