There are two popular pictures of what AI does to work, and both are wrong in the same way.
The first is AI as a tool. A person has a task; they pick up the model the way they would pick up a calculator or a search engine; they use it; they put it down. The work is still fundamentally theirs. The model made a step faster.
The second is AI as a replacement. The model is an autonomous agent. It is handed a job and it does the job, end to end, and the person who used to do it is no longer in the picture.
Both pictures are wrong in the same way: each keeps the boundary between human and machine exactly where it has always been, and simply moves a task across it. Tool-thinking moves nothing. The human still does the work, slightly assisted. Replacement-thinking moves the whole task. The machine does it now, the human is gone. Neither is the change that matters. The change that matters is not about which side of the line a task sits on. It is about dissolving the line.
What the loop is
Call it the loop. A decision made with the model in the loop is not one a person makes with assistance, and not one the model makes alone. It is a decision designed, from the start, so that a synthetic participant is a standing presence in how it gets made.
In a loop, the model is not waiting to be picked up. It is already there. It drafts the options before they are asked for. It surfaces the consideration no one raised. It argues the opposite case. It checks the proposal against everything the company knows. It runs the version of the analysis that, under the old scarcity, was simply too expensive to commission and so never got run at all.
And a human still holds the decision. Not as a formality. As the actual locus of judgement: choosing, owning the outcome, answerable for it. The human is not reviewing the model’s homework. The human is deciding, inside an environment where most of the cognitive labour of deciding well has already been done, in parallel, by a participant that is always present and effectively free.
The question stops being whether the human or the model decides. Both are in the room the whole time. The question is how the room is designed.
What it changes
When decisions are made this way, four things move.
Speed changes, but not in the way “faster” suggests. The cost of producing one more considered option falls to almost nothing. A team that could previously afford to work up two real options can now afford to examine ten. The gain is less haste than range. Decisions get wider before they get quicker.
The cost of being wrong falls. Many decisions were taken on thin evidence, not from carelessness, but because testing the assumption properly was too slow or too dear. When the analysis is cheap, more of the company’s beliefs can be checked before they are bet on.
Decision rights widen. A great deal of seniority is, in practice, the licence to make a call because you can be trusted to have done the thinking. When the thinking is in the loop (present, thorough, available), sound decisions can safely be made further from the centre. The organisation flattens, not by decree, but because the reason for some of its height has gone.
And the work itself changes character. Less of it is gathering and preparing; more of it is judging. The human contribution concentrates on the part that was always the point.
The danger, and the discipline
A loop can fail, and it fails in a particular way.
The risk is not that the model is wrong occasionally (it is), but that its fluency makes its output easy to accept and hard to challenge. A loop drifted into rather than designed slides toward automation bias: the human stops deciding and starts ratifying, while the org chart still says a person is in charge. The model quietly becomes an alibi. When it goes wrong, no one quite owns it, because no one quite made the call.
The defence is not less model. It is more design. A real loop is built, not stumbled into. It is explicit about where the model enters and where it does not. It is explicit about which decisions the model may take alone, which it may shape but not settle, and which a human must reason through unaided. It keeps dissent alive on purpose (a participant whose job is to argue the other side), so that consensus has to be earned. And it keeps accountability unambiguous and human, so that every decision has a name attached to it.
A company that designs its loops well becomes faster, broader in what it can consider, and markedly harder to fool. A company that simply lets the model seep into its decisions gets the speed for a while, and then gets a slow, expensive failure that no one saw coming, because the role of seeing it had quietly been automated away.
The loop is not a tool the company adopts. It is a shape the company’s decisions are rebuilt into, deliberately, decision by decision. That rebuilding is the work.