New AI 'Health Check' Tells Companies How Much to Trust AI

Companies are handing real work to AI agents — software that can read files, write code and take actions without a human pressing go. There is no shared way to judge whether that is safe. So four researchers interviewed 18 senior technology leaders across telecommunications, automotive, defence, aviation, banking, energy, government and enterprise software. Their conclusion: the usual report-card style ratings, which grade a company on a single 'how advanced are you?' scale, hide the very details that actually matter.

Their replacement is the Trustworthy Autonomy Health Check — think of it as a medical checkup for AI trust rather than a score. It asks eight questions. At the system level: how much authority does the AI have, what safety nets exist, can you trust the data behind it, is it boxed in so it cannot break things, and can you trace what it did and why? At the organisation level: who governs it, how closely do humans watch it, and are staff trained to keep up? Each question is graded on a five-step scale.

The counterintuitive finding: a higher score is not automatically better. A bank cautiously using AI to draft internal memos is not 'behind' a defence contractor letting agents touch flight software. What matters is fit — matching how much power you give the AI to the safeguards built around it. When those two are mismatched, you get either dangerous risk or useless friction: an AI that can move money with nobody reviewing it, or an AI so locked down it saves no time at all. The authors call this the alignment hypothesis.

The catch: this is a diagnostic tool, not a proven recipe. The authors say it still needs real-world testing, and it is aimed at large organisations with dedicated engineering and compliance teams — not a small business using a chatbot. Still, it gives managers a shared vocabulary for a question they are already being asked by their boards and regulators: how much should we actually trust the machine?