The problem
An accounts payable automation product performed well on the layouts in its training set and poorly on everything else. The set had been sampled from the three largest customers.
What we did
We assembled a layout mix weighted towards the long tail rather than the volume, and annotated key-value links instead of bare field detection so the model learned the relationship rather than the position.
Every document went through a redaction verification pass before annotation, which the client's risk team had made a condition of the engagement.
Outcome
Straight-through processing improved most on the suppliers that had previously required manual review, which is where the cost actually was.
