AI Doesn't Replace Decisions. It Changes the Conditions Decisions Are Made In.
What changes before the human clicks approve
By Louize Clark
Almost every AI adoption policy contains some version of the same reassurance: a human remains in the loop. The system can draft, summarise, flag, recommend, but a person still makes the decision, still clicks approve, still signs their name to it. That's the line accountability gets hung on, and in a narrow, literal sense, it's usually true. Someone did click approve. The decision, formally, was theirs.
It's worth taking that reassurance seriously rather than dismissing it, because it isn't wrong. The decision genuinely hasn't been handed to a machine. Nobody has quietly lost the authority to say no.
What that reassurance assumes, though, is that the conditions surrounding the moment of decision are the same conditions that existed before the tool arrived, the same range of options considered, the same amount of independent checking, the same instinct to question a result that looks slightly off. That assumption is rarely tested, because the decision itself, the click, the sign-off, looks identical either way. It's everything feeding into that click that has quietly had room to move.
Picture someone new to a decision-support tool, checking anomaly flags in a finance system, say, or reviewing a CRM's suggested next action. In the first weeks, they treat its output the way anyone treats something new: with care. They check it against their own judgement, sometimes ask a colleague, occasionally overrule it. Then, gradually, because the tool keeps being right and checking takes time nobody has given them back, the checking gets lighter. Not through carelessness. Through the same ordinary, reasonable accumulation of trust that any competent person extends to something that keeps performing well. A year in, they are still the one who clicks approve. They are no longer doing what "reviewing" used to mean.
That alone would be worth noticing. But there's a second shift happening alongside it, and it isn't only about trust, it's about what the person is even being asked to choose between. A tool that summarises a document has already decided what's worth including and what isn't. A system that flags anomalies has already decided what counts as normal. A CRM that recommends a next action has already narrowed a much larger field of possibilities down to the handful it thinks are worth showing. The human still chooses. But increasingly, they're choosing from a set that something else has already shaped and they rarely see what didn't make the list, because there was never a moment where the full list was presented for them to reject.
Of course, decisions have always been shaped by the information presented to the person making them, that's what a briefing paper, an analyst or a management report has always done. What's changing is the scale, speed and consistency with which that shaping can now happen across everyday work, quietly, inside tools nobody thinks of as advisers.
Put those two things together and the decision hasn't moved. Almost everything around it has. The range of options was narrowed before the person arrived. The scrutiny they bring to what's left has quietly thinned, month by month, for entirely defensible reasons. What gets recorded - a name, a timestamp, an approval, looks exactly the same as it always did, precisely because the part that changed isn't the part anyone writes down.
There is a further shift that may be harder to see. As people spend more time working alongside a system whose suggestions continue to appear useful and credible, the suggestion can start to feel less like an input to weigh and more like the obvious answer, the one they would have reached anyway. Whether that's true in any given case is genuinely hard for the person involved to know from the inside. Has the system learned how they think? Or have they gradually adapted their own thinking to the patterns, language and options the system tends to offer? We do not yet know what years of working this way will do to independent judgement. The evidence is still developing, and it would be too early to claim that AI is simply making people less capable of thinking for themselves. But it would be equally complacent to assume that repeated reliance changes nothing.
The shift isn't limited to whether someone checks an answer, either. It can begin earlier than that. The system remembers the previous meeting, retrieves the relevant documents, identifies the unresolved issues, prepares the questions, drafts the response. By the time the person reaches the decision, much of the work that used to exercise memory, comparison and judgement has already been done somewhere upstream of them.
This is where the pattern stops being a story about one person and becomes an organisational one. An organisation may believe it has introduced one tool into one role. In practice, it has created a different working relationship for every person using it. One employee asks the system after they've already formed a view of their own. Another asks it before they have formed an initial view. One uses it to widen the possibilities in front of them. Another uses it to narrow them quickly and move on. One treats disagreeing with the system as entirely ordinary. Another increasingly experiences its first answer as the natural starting point. The licence is standardised. The decision conditions are not.
Two people can therefore reach the same decision, tick the same box, sign the same approval one having checked everything, the other almost nothing both entirely convinced they were exercising their own judgement. The record shows no difference between them at all. A decision-making policy written to govern who signs, what gets escalated and what counts as approval was never designed to detect a gap like that, because the gap doesn't live in the decision. It lives in everything that happened on the way to it.
That leaves organisations with a harder question than whether a human remained in the loop: did the person still encounter enough of the problem, enough of the evidence and enough genuine uncertainty to exercise judgement at all? Keeping a human at the end of a process is not the same thing as preserving human judgement throughout it.
AI Policies UK helps organisations see the AI infrastructure they're already standing on, chosen, inherited and embedded — before decisions like this one have to be made under pressure. Get in touch: louize@aipolicies.uk