THE AI INTERVIEW

She identifies a second critical gap around explainability – the ability to account for why a system reached a particular outcome.

“ Agentic AI produces outcomes through chains of micro-decisions that are not individually explainable in any meaningful sense,” says Ivana.“ What we need are governance frameworks that assess systems at the level of their behaviour over time.”

Trust by design For Ivana, compliance with regulation is merely a starting point. The more demanding standard is what she calls

“ trust by design” – building AI systems that a fully informed person would consider genuinely trustworthy.

For agentic AI, this means building systems that are legible and auditable, incorporating meaningful human oversight and designing for failure. In other words, assuming agents will make mistakes and ensuring the consequences can be recovered from.

“ The organisations that treat governance as a constraint will build systems that are compliant but not trusted,” Ivana contends.“ The organisations that treat governance as a design principle will build

aimagazine. com 25