Purchase orders: AI that writes the extraction rules

The situation

Purchase orders arrive in many layouts. Entering them by hand is slow and costly. Sending every order through a model is faster, but the cost grows with every document, and the results are hard to inspect.

What we did

Instead of using a model to extract every item, we used AI to develop deterministic extraction patterns. The model works out the pattern for a supplier’s layout, passes hints, and learns the extraction. From then on, ordinary code applies the pattern.

What changed

A learned profile cost a twenty-fifth of an AI-extracted one, and an eightieth of manual entry.

Why it matters

The pipeline is low cost, fully inspectable and explainable: anyone can see why a value was extracted. That makes it a sound basis for future financial agentic integration. It also shows the principle at work: use AI where it helps, and keep it out of the steps where it does not.