The AI assurance gaps nobody is talking about

AI governance has matured faster than most expected. Organisations are certifying against the likes of ISO/IEC 42001. Boards are engaged. NIST AI RMF is being referenced globally. The EU AI Act is forcing conversations that needed to happen.

That progress is real for many. But a structural gap remains, one that existing standards, tools, and schemes were never designed to close. In high-risk deployments, that gap is a material board-level exposure right now.

This talk maps the AI assurance landscape honestly across three layers.

• Layer one is high-level governance covering how organisations manage AI responsibly at the enterprise level. Standards exist. Certification infrastructure exists through accredited certification bodies. Boards have cover here. This layer is largely functional.
• Layer two is model assurance. Whether a specific AI system does what it claims, behaves consistently in your deployment context, and is being managed against its known risks. You are deploying OpenAI, Google Gemini, Anthropic Claude, AWS Bedrock. You are building RAG architectures on top of those models. Standards are forming. Tools like Credo AI, Fiddler, and Lakera are real and useful. But an independent, ongoing, accredited evaluation regime does not exist.
• Layer three is agent assurance. Evaluation of autonomous AI agents that plan, decide, and act across multiple steps with limited human oversight. AIUC-1 is the most forward-looking framework available, and it is very new. Everything else is nascent or non-existent.

Most organisations are starting to operate confidently at layer one and deploying technology at layers two and three with practically no assurance. 

The talk draws a direct parallel with cyber security. Organisations once had security policies and compliance frameworks and believed they were covered. Then the industry recognised that policy compliance said nothing about whether systems were actually secure. Red teaming and penetration testing became foundational. DevSecOps embedded security assurance into the development lifecycle. The result was a mature, independent, ongoing assurance regime that boards and regulators can rely on. AI is at the same inflection point, but will not follow the same path. The tools exist. The methodology is forming. What does not yet exist is the verified, continuous evaluation regime that connects automated assurance to real external accountability.

Attendees will leave with a clear picture of where the AI assurance landscape actually stands, what the gaps are and why they matter, and six questions every board should be asking before approving any high-risk AI deployment. 

This is an honest assessment of the gap the industry needs to close and a framework for understanding what closing it will actually require.