The challenge
- AI features are easy to demonstrate and hard to make dependable on real, messy production data.
- The capability had to fit inside an existing product and its existing workflows.
- Accuracy needed to be measurable rather than asserted.
What we did
- Established what the manual process cost in time and error rate, to give the AI work a baseline to beat.
- Prototyped against real client data rather than curated samples.
- Constrained model output to a validated schema so downstream systems could rely on it.
- Routed low-confidence cases to human review instead of letting them through silently.
Where it stands
- AI capability moved from prototype into a production software product.
- Output validated against a schema before reaching downstream systems.
- A human-in-the-loop path for cases the model is not confident about.