What it is

Assurance is borrowed from audit and safety engineering. Its product is not a better system but justified confidence in a statement about one: that a model meets its accuracy threshold, that a process was followed, that a management system conforms to a standard. The person who relies on the statement (a board, a regulator, a customer, an affected individual) is usually not the person who can check it, and assurance is the bridge between them.

Applied to AI, assurance has to cope with properties traditional assurance rarely meets: behaviour that is probabilistic rather than specified, systems that change when a vendor updates them, and claims (fairness, explainability, robustness) with no single agreed measure. That is why AI assurance has become a field in its own right. In September 2025 the UK Department for Science, Innovation and Technology published a roadmap to professionalise third-party AI assurance, backed by an £11 million AI Assurance Innovation Fund.

How it differs from testing, audit, certification and conformity assessment

These words are used interchangeably, and they are not the same thing.

  • Testing and evaluation produce evidence: measurements of what a system does. They are inputs to assurance, not assurance itself.
  • Assurance is the broader activity of turning evidence into a justified, communicable claim. It can be first-party (the team that built the system), second-party (an independent function inside the organisation, such as risk or internal audit) or third-party (an external body).
  • Audit is assurance against defined criteria by someone independent of what is audited, ending in a formal opinion.
  • Certification is third-party audit against a published standard by an accredited body, ending in a certificate. For AI management systems the standard is ISO/IEC 42001, and ISO/IEC 42006 sets the rules for the bodies that certify against it.
  • Conformity assessment is the EU AI Act's legal procedure for showing that a high-risk AI system meets the Act's requirements before it is placed on the market. It is a regulatory obligation, not a voluntary assurance product.

The practical consequence: a certificate assures a management system, a conformity assessment assures a product against the law, and neither assures that a particular model performs well in your use. That question needs its own evidence.

Assurance and the three lines of defence

In financial services the familiar three lines map directly. The first line, the business that owns the system, tests it and attests to it. The second line, risk and compliance, sets the standards and challenges the evidence. The third line, internal audit, gives the board independent assurance that the first two are working. External assurance providers sit outside all three. The PRA's model risk principles add a specific requirement for independent model validation, which is second-line assurance by another name; see model risk management.

What good looks like

A board or a regulator should be able to see four things.

  • A defined claim for each material system: what it is supposed to achieve and within what limits, usually drawn from the risk appetite and the control objectives beneath it.
  • Evidence tied to the claim: test and monitoring results that measure the claim directly, refreshed whenever the system, or its vendor, changes it.
  • Independence proportionate to the stakes: self-assessment for low-risk tools, second-line challenge for material ones, third-party assurance where a regulator, customer or affected group needs it.
  • An opinion someone signs: assurance ends in a statement with a name against it, not in a dashboard.

The common failure is assurance of process without assurance of outcome: a well-documented approval process for a model whose live performance nobody measures. That is the AI system lifecycle controls gap.

Frequently asked questions

What is AI assurance? The process of building and communicating justified confidence that an AI system, or the organisation operating it, meets the claims made for it.

What is the difference between AI assurance and an AI audit? An audit is one form of assurance: an independent examination against defined criteria that ends in a formal opinion. Assurance also includes internal testing, second-line review and certification.

What is third-party AI assurance? Assurance provided by a body independent of both the developer and the deploying organisation, such as an auditor or a certification body.

Does ISO/IEC 42001 certification assure an AI system? No. It assures the management system around AI. It does not show that any particular model is accurate, fair or lawful.

Where it sits in the register

AI system lifecycle controls gap is the failure assurance most often misses, inadequate AI vendor due diligence is the same gap for bought systems, and unclear AI ownership and accountability is what leaves assurance with nobody to sign it. The controls that generate the evidence are adversarial robustness testing, bias testing and fairness monitoring and AI system monitoring and logging. The AI ownership and accountability framework decides who signs.