AI System Inventory and Risk Classification | EU AI Act

Define the unit of inventory

Register a business use case, not merely a model name or vendor contract. The same model used for recruitment screening and internal text summarization represents different purposes, affected people, decisions and risks.

Assign a stable identifier so approvals, assessments, incidents, contracts and technical records can point to the same system.

Capture the minimum decision-grade data

The record must be useful to legal, security, procurement and the business without becoming an unmaintainable questionnaire. Use structured fields and link detailed evidence rather than copying documents into free text.

• Purpose, scope, users and affected groups
• Provider, deployer and other AI Act roles
• Models, data sources, outputs and integrations
• Autonomy, human review and consequence of error
• Lifecycle state, owner, controls and evidence links

Classify through a gated workflow

Begin with prohibited-practice screening, then assess high-risk categories and transparency obligations. Record the questions, answers, reviewer and date. Escalate uncertainty instead of forcing a convenient classification.

Classification is not permanent. Define change triggers such as a new purpose, model replacement, sensitive data, a new affected population or increased autonomy.

Connect controls to the record

Each required control needs an owner, status, evidence and next review date. This turns the inventory from a spreadsheet into a work queue and provides management with an accurate view of unresolved exposure.

Use role-based access: broad discovery, controlled editing and protected evidence for sensitive systems.

Retire systems deliberately

Retirement should revoke access, stop integrations, resolve retention duties, archive required evidence and confirm that downstream processes no longer depend on the system. Keep the historical record for auditability.