AI assurance
gobernanzaAI assurance · assurance for AI systems · independent AI verification
Independent verification that an AI system does what the organisation claims it does, and that it leaves enough evidence to defend it to a third party: auditor, regulator, customer or court.
AI assurance is the layer that checks and attests, from outside the team that built the system, that AI in production matches what the organisation says about it: what it decides, on what data, within what limits and under what human oversight. It rests on three pieces: controls defined across the lifecycle (design, deployment, operation, retirement), recorded evidence that those controls actually run, and an opinion or report from someone who is not a party to the build. It should not be confused with governance: governance is normative — it decides what is allowed, who approves and who is accountable; assurance is evidentiary — it verifies that what was decided actually happens, and documents it in a form that survives external examination. Nor is it the same as evaluations: evals measure whether the model output is good; assurance can conclude that a system with excellent metrics is indefensible because nobody recorded who approved it or what it did on the day it failed. In Europe the driver is twofold: transparency obligations already enforceable against deployers, and national supervisory authorities with inspection powers. The practical consequence is uncomfortable: the question stops being whether your AI works and becomes whether you can prove it with what you kept.
How it differs from
- AI governance
- Governance decides what is allowed, who approves and who is accountable; assurance independently verifies that this happens and keeps the proof.
- Evals
- Evals measure whether model output is correct; assurance checks whether the whole system is defensible to a third party, even when the metrics look good.
- Agent observability
- Observability serves the team that operates and debugs; assurance serves an outsider who verifies, and therefore demands evidence integrity, retention and traceability.
FAQ
- Do small companies need AI assurance?
- It depends less on headcount than on what the AI touches. If a system decides on or influences people, money, contracts or personal data, someone will eventually ask how it works: a large customer's vendor questionnaire, your insurer, a financial auditor or a supervisor. Formality scales with size: in a 50-person company assurance is not an audit report, it is an ordered dossier that exists before anyone asks for it.
- Can the same firm that implements my AI also assure it?
- It can, and often does, but then it is not independent and that has to be stated. Independence is what gives the opinion value: if the builder signs it off, what you have is a well-run self-assessment. A reasonable middle path is separating roles inside the provider and keeping evidence in a format a third party can review later without depending on them.
- Where do you start?
- With the inventory: which AI systems and agents are live today, who approved them and what they can touch. Without an inventory there is no scope, and without scope there is no assurance. Minimum controls per system and the record that evidences them come next.