• Workforce & AI Skills

We are training people to use AI. Too often, we are not training them to disagree with it.

That gap worries me more than any model limitation, and it is what I find myself talking about most since being named in the Foundever AI 100 UK this year. I explored the governance side of it in my AI Week 2026 session this October. The conversations that followed convinced me it is just as much a skills problem.

Most organisations are investing in two things: AI tools, and policies to control them. Between the two sits a moment that rarely appears on any roadmap. A person looks at an AI output and decides what to do next. I call this the AI decision gap: the space between an AI recommendation and an accountable human decision.

Here is what it looks like in practice. Consider this illustrative scenario from a service management setting, where AI now categorises incidents, suggests priorities and recommends next actions. Picture a tool flagging a ticket as low priority. A busy analyst accepts it, because the tool is usually right and the queue is long. Days later, that ticket turns out to be linked to a payroll failure affecting hundreds of staff.

Nobody broke a rule. The policy said a human should review AI recommendations, and a human did. But reviewing had quietly become clicking. Nobody had been equipped to ask the right question at the right moment.

We tell people to keep a human in the loop, then give them no guidance on what being in the loop means. When things go wrong, we ask whether the model was flawed. The more useful question is how the decision was made.

Three lessons have shaped how I think about closing that gap.

Make ownership explicit. Not ownership of the system or the policy, but of acting on this particular output. When accountability is vague, people defer to the machine, because agreeing with it feels safer than taking a position.

Teach interrogation, not review. Reviewing asks, “Does this look right?” Interrogating asks, “What would have to be true for this to be wrong?” What context can the tool not see? Is this case unusual? What does a mistake cost here? These are thinking skills, not technical ones, and they can be taught. A handful of well-chosen questions changes behaviour more than another page of policy.

Make reasoning visible. When someone overrides an AI recommendation, or accepts one where the stakes are high, one sentence explaining why is enormously valuable. It makes the basis for the decision clearer, gives governance teams an audit trail they can genuinely use, and over time reveals where the tool helps and where it quietly leads people astray.

None of this need’s new software. It needs us to treat judgement as an AI skill in its own right, alongside data literacy and prompt writing. When I ask teams about their AI training, the answer is almost always about how to use the tool. Very little is about how to challenge it.

That, for me, is where the workforce conversation needs to go next. The organisations that get lasting value from AI will not simply have the best models. They will have people confident enough to question an output, clear enough about their accountability to act on that question, and leaders who reward good judgement, not just speed.

So here is a simple test for any leader reading this. Ask your team what would trigger them to challenge an AI recommendation, who could authorise an override, and how they would record the reason. If those answers are unclear, the gap is already open.” Never overriding AI does not, by itself, prove poor judgement.

People also need the authority and support to pause or escalate a recommendation when the consequences are unclear.” Skills matter, but confidence alone cannot compensate for a workplace that penalises challenge.

Responsible AI depends on responsible human decisions. The technology will keep improving. Our job is to make sure the people using it improve alongside it.

---

Read More Workforce & AI Skills

Comments are closed.