Compliance & audit
Workloads & AI inventory
A workload is a registered AI use: the app behind one or more keys. Regulators expect an inventory of every AI use with a named owner; workloads are that inventory.
What a workload records
- Owner, team and purpose. Apps without an owner show under Needs attention.
- Data types it may handle, and whether every request should be treated as containing them.
- Task (chat, code, retrieval, generation) and autonomy (assistive or autonomous).
- EU AI Act tier: minimal, limited, high or prohibited.
- Monthly budget. When it's used up, requests are refused with
budget_exceeded.
Lifecycle
| Status | Meaning |
|---|---|
draft | Registered, not yet started. |
pending_approval | High-risk AI waiting for someone who manages workloads to approve it. Requests are refused until then. |
running | Can call models. Connect an app starts workloads directly when the pre-launch checks pass. |
stopped | Paused, for example by a kill switch. |
retired | No longer in use; kept for the record. |
Kill switches
From an incident or the risk pages, stop a workload, a key, a provider, or all workloads at once. A reason is required and recorded in the audit log and the incident timeline.
Approvals
Requests held for review (risky requests, or autonomous high-risk use) wait under Approvals (opens in a new tab) until a person releases or rejects them.