The KYC agent was the easy part

Automating a KYC check is a solved problem. Getting a compliance officer to accept the output is not, and that is where the project actually lives.

The pipeline we built does what you would expect: intake, extraction, risk flags, and a human-on-the-loop interface over the top. It ran on sandbox data before it ran on anything real, and it went to production with evals and audit trails attached. Any competent team can assemble those parts now.

What determines whether it ships is a different question. When the agent is wrong, how quickly does a person find out, and how much does it cost them to overrule it?

Two constraints did most of the work. The first is that the officer can see the evidence without reconstructing it. A decision arrives with what produced it, and the trail behind it is a record someone outside the team could read, not a debug log. An officer who has to rebuild the reasoning will go back to the documents, and then you have paid for the system and the manual work both.

The second is that overruling has to be cheap. If correcting the agent costs more than doing the task by hand, nobody corrects it. They stop using it, quietly, and you find out months later.

None of that is model work. All of it is what makes the model work usable, and it is consistently underestimated in scoping, including by us.

If you are scoping something similar and want to compare notes, write to hello@windmill.digital.