Seoul firms test AI assistants with human rules

The pilots that survived a year have one thing in common, and it is not the model. It is that somebody wrote down what the assistant is not allowed to decide.

The scope document

Each deployment starts with a two-page document naming the decisions the assistant may make, the decisions it may draft for a human, and the decisions it may not touch. It is reviewed quarterly and signed by a named manager.

That sounds bureaucratic. It is the single strongest predictor of whether a pilot is still running twelve months later, because it converts an open-ended capability into a bounded tool that a team can be trained on and audited against.

An assistant without a scope document is a colleague nobody hired and everybody blames.

Where the failures clustered

Not in accuracy. In accountability.

The pilots that collapsed did so when something went wrong and no one could say who was responsible for the output. The assistant drafted, a human approved without reading, and the approval was treated as a formality by everyone including the person who gave it.

Firms that fixed this made approval consequential: the approver’s name attaches to the output, and a sample is reviewed. Approval rates dropped immediately, which is what a functioning control looks like.

The measurement problem

Nobody has a good productivity number. The available measures — tickets closed, drafts produced, time to first response — improve reliably and say little, because the assistant is very good at producing the artefacts those metrics count.

Two firms have stopped trying to measure output and started measuring rework: how often a drafted item comes back. It is a slower signal and a harder one to game.

What to watch

Watch approval rates after review is introduced, since a rate that does not move means the review is not real. Watch rework rather than throughput. And watch whether the scope documents are actually revised at the quarterly review or rolled forward unchanged.