Decision Quality in the AI Era

When AI becomes a default advisor, two things can happen at the same time: decisions get faster, and accountability gets thinner.

The failure mode no one talks about#

The most common failure mode is quiet. Teams stop thinking and start approving. AI generates a recommendation. A human reviews it. The human approves. This looks like oversight. It is often rubber-stamping.

Over time, the human loses the skill to generate the recommendation themselves. Errors become invisible because no one checks the reasoning. Accountability diffuses, and failure becomes "process" rather than ownership.

Decision quality is about keeping humans actually in the loop rather than nominally present, and none of it requires slowing down. The distinction matters because the appearance of oversight is not the same as its substance. A human who could not have produced the recommendation, and cannot explain why it is right, is not overseeing the decision. They are laundering it.

The organisations that preserve decision quality build the discipline back in: they require a human rationale for consequential calls, they keep people practising the judgement the AI is handling, and they treat the reasoning behind a decision as something to be examined rather than assumed. The goal is to ensure that when the decision matters, a capable human is still doing the deciding. Rejecting the tool achieves nothing.

The Decision Quality Protocol#

> AI can advise. It cannot carry responsibility. > > Rahim Hirji

The protocol is a short written answer to three questions, kept for each class of decision that would be expensive to get wrong. It fits on a page and it is meant to be adopted rather than admired.

One. How the decision gets made with AI in the loop. What the machine may draft, what it may decide on its own, and what a person has to do from the start. Written before the tool arrives in the workflow rather than inferred afterwards from what people ended up doing.

Two. Who owns it. One name against each class of decision. A committee in that box means nobody, and the failure mode above depends on nobody being in the box.

Three. What quality control looks like. How you would know the decision was wrong, who checks, how often, and what record is kept of the reasoning rather than the output. If the answer is that the error would show up in the result eventually, that is not quality control, it is hindsight.

The reason to write it down is that the alternative is not neutrality. AI is making everyone faster but nobody better, and speed is replacing rigour, because the pace of the tool sets the pace of the review unless something else does. A protocol is the something else.

Related SuperSkills research#

The full evidence review, including the meta-analysis showing that undesigned human-AI pairing performs worse than either party alone, is in human and AI decision making. See also Human at the Start, AI and human judgement and the verifier's discount.

Put it to work#

The operational version of this, stage by stage with a downloadable working grid, is the Delegation Boundary Map. It turns the approving-rather-than-thinking failure into an explicit set of decisions made before the work starts.

The evidence behind that failure mode, and the position it leads to, is set out in human in the loop is not a safeguard.

About this research#

Written by Rahim Hirji, author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company.

How this research works · Reviewed quarterly · Found an error? Tell me and it is corrected on the page.