Koher's tools keep the AI out of the judgement. The same split, made in a person instead of in software, is what a lot of work already does to people — and it is why those people have nothing to fall back on when a machine arrives that executes better than they do.
Koher builds tools on one rule. The AI reads and describes. Code holds the decision, so the decision can be read by anyone. A person decides. The reason is ordinary: a model asked to do both jobs starts handing out verdicts, and a verdict from a model cannot be checked, only accepted.
That rule is about software. This position is about what happens when the same split is made inside a person, which is what a great deal of work already does.
Two unlike kinds of thinking, kept in separate hands. Nothing crosses the seam.
Some work asks a person to run one channel
A registrar checks a form against a rule. An accounts clerk matches a bill to the sanction that covers it. An admissions desk applies a cut-off it did not set. In each case the person is required to perceive accurately and forbidden to conclude. Notice everything; decide nothing. The rule was written elsewhere and their job is to apply it the same way every time.
That is the same two-kinds-of-thinking problem, made in a person rather than in a system. And it does not arrive as an insult. It arrives as professional standards — consistency, reliability, no deviation — which is why it is hard to argue with, and why after enough years people believe it about themselves.
The demand is for a machine: one channel, and the other draining out of the job description.
Then a machine arrives that is better at it
An institution spends years telling somebody their competence is reliable execution. Then it hands them a tool that executes. The tool speaks in the register their training taught them to respect — confident, consistent, procedural — and they have little to set against it, because the thing they were told to be good at is the thing it is better at.
There is a measurement of this. Dratsch and colleagues, in Radiology in 2023, showed twenty-seven radiologists fifty mammograms with a purported AI suggesting a category for each, twelve of the suggestions wrong. Among the least experienced, correct assessments fell from 79.7% to 19.8%. Among the very experienced they fell from 82.3% to 45.5%. Experience did not protect them.
Dratsch et al., Radiology, 2023. The judgement was not available at the moment it was needed, because it had never been named as part of the job.
So the discipline has to hold in both places
In the software, the model never holds the decision. In the room, a Koher tool never puts a suggestion inside the procedural decision. Whatever it offers sits beside the task. The register is still balanced by the person; the form is still checked by the person. A tool that began proposing what the entry should be would have crossed to the other side of that seam, and the figures above would become a statement about it.
The obligation stays. The absence of control stays. What is supplied is brief, and sits outside the seam.
Why fun, and not something more serious
What this kind of work removes is everything about a person that is not execution. Fun is the clearest case of that: it yields nothing, it cannot be booked, and no procedure has anywhere to put it. That is exactly why it is worth supplying. It gives back a few minutes of being more than an executor, without touching the demand — and the demand is not Koher's to touch. It belongs to whoever wrote the procedure.
Two things this is not. It is not permission to do the work badly: the register still balances, the file still closes, the form is still correct. And it is not a way of making anyone happier at being single-minded, which would be the same collapse with a pleasant surface on it.
Keeping AI out of the judgement, and giving somebody back a few minutes of being more than an executor, are the same commitment. Koher makes it once in the code and once in the room.