Language models are remarkably good at pulling structure out of messy conversations. They are also quietly inconsistent. A prompt change, a model update or a new customer vocabulary can cut accuracy without anyone noticing — until a rep sees a wrong next step on a deal.
Golden sets per field
We maintain hand-labelled evaluation sets for every field we extract, such as next steps, stakeholders, competitors and budget signals. Each set mixes typical examples with hard cases: sarcasm, hypotheticals (“if we had budget…”), multiple speakers and switching languages. Every change to a prompt or model runs against these sets before it ships.
Measure what users feel
Overall accuracy hides the errors that matter. We track precision and recall separately per field, because a false commitment is worse than a missed one. We also weight errors by impact: getting a close date wrong matters more than misspelling a job title.
“An extraction that’s confidently wrong does more damage than one that admits it isn’t sure.”
Drift alarms in production
- Sample a slice of production extractions daily for human review.
- Alert when confidence distributions or extraction rates shift sharply.
- Treat every user correction as a labelled example and feed it back into the golden sets.
- Pin model versions and roll out upgrades behind a comparison gate.
Calibrated confidence
Every extracted claim carries a confidence score, and we calibrate those scores so 90% really means right about nine times in ten. Low-confidence claims are shown as suggestions for a person to confirm rather than facts. That one design choice does more for user trust than any accuracy improvement.