AI Governance Sandbox
Describe an AI system. Get a risk classification mapped to the EU AI Act, the controls that follow from it, and who needs to sign off. The classification is computed from a fixed rubric, not generated by a model — the same inputs always produce the same answer.
Autonomy
How much human judgement stands between the system and the outcome.
Data
What the system is fed and where that data travels.
Affected people
Who bears the consequences, and whether they chose to be there.
Oversight
Whether anyone would notice if it started going wrong.
Prohibited practices
EU AI Act Art. 5. Any one of these is a hard stop, whatever the rest of the assessment says.
Classification
Limited risk
Transparency obligations apply; proportionate controls are enough.
Why
- Moderate exposure: meaningful consequences, but with human judgement or reversibility available.
- Art. 50 transparency obligations apply wherever people interact with the system or receive AI-generated content.
Sign-off required
System owner · Data protection lead
Review cadence
Every six months
Required controls
8 controls, each traced to the obligation it comes from.
Record the system in an AI inventory with an accountable owner named.
Write down the intended use and the uses explicitly ruled out.
Confirm a data processing agreement is in place and that inputs are excluded from vendor training.
Strip identifiers that the task does not require before sending data to the model.
Disclose AI involvement at the point of interaction, in plain language.
Insert a human review step with authority and time to actually reject the output.
Build a labelled evaluation set from real cases and record baseline accuracy before launch.
Log inputs and outputs, and sample them on a schedule for quality and drift.
Policy language
Turn the assessment into a paragraph you could put in front of a board. The classification above does not change — only the prose is generated.
Design notes
The classification is deterministic by design. A governance tool that returns a different tier on a second run is worse than no tool, so the rubric is plain code: scored dimensions, plus two hard gates that override the score entirely — Art. 5 prohibited practices, and Annex III deployment contexts that are high-risk regardless of how carefully the system is built.
A language model is used for exactly one thing: rewriting a fixed result as readable prose. It is never asked what the risk level is. If the model is unavailable, the assessment is unaffected.
Educational tool, not legal advice. Citations refer to Regulation (EU) 2024/1689 and the NIST AI Risk Management Framework 1.0.