The Organisation Has Not Changed. The Decision Has.

Imagine returning to an organisation six months after it introduced AI into an important decision process.

The organisation chart is unchanged. The same manager owns the process, the same reviewer still approves, and the escalation route remains in policy. No delegation has been rewritten. Nothing obvious suggests that authority has moved.

Yet the decision reaching that reviewer may no longer be the same kind of decision.

Before AI assistance, a case might have arrived as a collection of documents, records, notes and conflicting signals. The reviewer had to work through them, decide what mattered, connect pieces of information and gradually form a view.

Now the same case may arrive already condensed. A summary has been prepared. Anomalies are highlighted. Similar cases have been retrieved. A likely outcome may be suggested before the reviewer has examined the underlying material.

This can be genuinely useful. In many settings, it should be. There is no virtue in forcing an experienced professional to spend two hours reconstructing information that a reliable system can assemble in seconds.

Still, something has changed before the authorised person begins to decide.

Discussions about AI and authority often move too quickly towards the question of who made the decision. Was it the human, because only the human had approval rights? Was it the system, because its recommendation materially influenced the outcome?

Sometimes that is the wrong place to start.

Formal authority tells us who may commit the organisation, who may approve an exception, and who must answer for the decision afterwards. Those questions remain fundamental. But authority enters a process that is already under way.

By the time a person approves something, information has arrived through particular channels. Some details have been made prominent, others pushed to the edge. A recommendation may already have created an initial orientation. The workflow may have made one course of action easy and another cumbersome. Time pressure may have changed what still counts as a realistic alternative.

None of these conditions needs formal authority in order to affect judgement.

I find it useful here to distinguish authority from decision-shaping capacity: the practical ability to configure part of the informational or procedural environment from which judgement proceeds.

We already recognise something similar in ordinary life.

You enter a supermarket because you need bread and eggs. Nobody tells you what to buy, and whatever ends up in your basket is, in an ordinary sense, your choice. But you did not design the route through the store. You did not decide what appeared first, what sat at eye level, what was promoted or what required an extra effort to find.

If you leave with seven items instead of two, it would be wrong to say that you did not choose the other five. You did. It would also be too simple to say that those choices began only when you reached for them.

The environment had already shaped which possibilities became salient, convenient or easy to ignore.

An AI-assisted decision can develop in the same way. Suppose five reasonable courses of action exist, but the system presents three as most relevant. The reviewer still chooses among them. Yet the other two were not rejected by the person holding authority; they simply did not occupy the same field of consideration. Perhaps they remain recoverable, but only through additional effort.

That difference matters because choosing among options is not quite the same as participating in how the choice set itself was formed.

This does not make the system the decision-maker. It shows that influence can occur earlier than the final act of choice.

A model may generate a recommendation, the data determine what it was able to recognise, the interface affects what the reviewer encounters first, and policy or workflow shape the time and routes available around that recommendation. The reviewer adds experience, scepticism and contextual knowledge that may exist nowhere else in the process.

No single component has to control the outcome for the architecture of the decision to change.

The human can remain clearly visible throughout. They may still sign, still hold the same authority and still perform the role conscientiously. What changes is the work surrounding that authority.

There is a real cognitive difference between constructing an assessment and evaluating one that has already been constructed. Both require judgement, but the second begins with a candidate interpretation already present in the environment. The reviewer may reject it, revise it or ignore it. Even so, judgement begins from a field that acquired shape before the reviewer fully entered it.

Nor does AI narrow human agency by definition. A good system can surface buried evidence, expose inconsistencies, generate alternatives and remove clerical work that prevents professionals from concentrating on difficult cases.

The practical question is what remains possible after the assistance has done its work. Can the reviewer still understand the basis of the case, recover what lies outside the summary and develop another interpretation without unreasonable friction? Is there enough time to pause when uncertainty matters? If escalation is needed, can it still alter the outcome?

The same person with the same formal authority may exercise judgement under very different conditions depending on whether they have twenty minutes or four, whether primary evidence is immediately available or several layers down, and whether the recommendation appears after their own analysis or before it.

These differences become especially important later, when someone asks why a consequential decision was made.

At that point, the organisation reconstructs backwards. It finds the authorised person, the approval record, the policy and perhaps the system output. Everything may look straightforward: the reviewer approved, the policy permitted it, the process was followed.

What is harder to recover is the relationship between those facts.

Perhaps the recommendation was already present before the reviewer formed an independent view. Perhaps the evidence underneath the summary was technically available but rarely consulted. Perhaps review times had been shortened because automation increased expected throughput, while the escalation route remained designed for an older tempo.

Once the outcome is known, these details begin to look like one coherent story. That is precisely where retrospective explanation can become deceptive. The final approver is an obvious centre of gravity because their action is documented and attributable. What happened before that moment is often more distributed and much harder to narrate cleanly.

A good retrospective account therefore has to resist becoming too neat.

An approval record tells us who committed the organisation. A system log may show that a recommendation was generated. A policy may prove that escalation existed. Useful facts, all of them. They do not by themselves tell us what the reviewer saw first, what had already been interpreted, how much effort deviation required, or what remained genuinely open when the human entered the process.

This is where formal governance can remain entirely valid and still become insufficient as an explanation.

No organisation should try to record every influence surrounding every decision. That would be neither realistic nor useful. But it does need enough visibility to reconstruct the conditions that became material when authority was exercised.

The problem becomes more subtle as AI-assisted workflows mature. A team introduces summarisation. Reviewers still open source material regularly. The summaries prove reliable, so more work begins there. Recommendations are added later. The process becomes faster, expectations adjust, and eventually an interface redesign places primary evidence one level deeper because most users no longer need it in ordinary cases.

Nothing dramatic has happened. Each change has a reasonable explanation.

Months later, however, the same reviewer may be exercising the same authority inside a substantially different decision environment.

Organisations tend to recognise governance change through visible institutional events: a delegation moves, a role changes, a committee is created, a threshold is revised. Changes in decision production arrive under quieter labels - workflow optimisation, model updates, interface redesign, efficiency gains. Usually they are exactly that. Occasionally, in combination, they alter the conditions under which judgement takes place.

This is why the question “Where does effective authority actually sit?” does not always have a satisfying singular answer.

Formal authority may sit clearly with a human reviewer while the shaping of the decision is distributed across systems, interfaces, data, policies, workflows and other people. Those contributions are not equivalent, and collapsing them into one idea of who “really decided” can obscure more than it reveals.

The more useful task is to reconstruct their relationship: what shaped the decision before commitment, what remained open afterwards, and where institutional authority finally became consequential.

That shifts governance slightly away from the organisation chart and closer to the decision itself.

The chart is still useful. It tells us where authority is formally located.

It just cannot, on its own, tell us how the decision came to be.