AI Assurance & Validation
Evidence that an AI system does what you say it does, and keeps doing it.
In classical software, verification does most of the work: there is a specification, and you check the system against it. A trained model has no such specification. What it does is a property of its data, its objective and the conditions it runs in — so the burden shifts to validation. Does this system behave as intended, on the distribution it will actually meet, inside the boundary someone drew for it?
That is also what the regulation now asks for. Under the EU AI Act, performance is not the deliverable. Traceability, human oversight and a record that can explain a decision are.
The boundary stated explicitly: what the system is for, what it is not for, and what is supposed to happen at the edge of it. Most AI failures are boundary failures.
Training and evaluation data examined for provenance, representativeness and leakage. Performance measured on the distribution the system will meet, not the one it was built on.
Whether the person who is meant to be able to intervene can actually do so — in the time available, with the information the interface puts in front of them.
Drift detection, retraining gates and the evidence trail that keeps a conformity claim true after the model has changed.
AI assurance sits at the point where a new regulation meets the product rules that were already there.
Regulation (EU) 2024/1689. For high-risk systems: risk management, data governance, technical documentation, logging, human oversight, accuracy and robustness.
The AI Act and the product regulation applying at the same time — the MDR for medical devices, ISO/PAS 8800 and ISO 21448 for road vehicles.
IEEE 1012 for verification and validation and ISO/IEC/IEEE 29119 for testing, applied to systems whose behavior is learned rather than specified.
The model treated as part of the attack surface: data poisoning, model extraction, and prompt and tool abuse where agents can act.
Described by what the systems had to do, rather than by client name.
Systems whose output affects a person — credit, pricing, triage — where the decision has to be explainable to the person it was made about.
Agents with the ability to act, assessed for what they can reach, what they can change, and what stops them.
Pipelines built and run in environments where a regulator is part of the audience, not an afterthought.
Contributions to the international standards this discipline is now being measured against.