What the AI Listens For
As the transcript flows in, the AI pulls structured facts out of it and drops them onto the fact canvas. Some kinds it nails on day one. Some take a few weeks of corrections before they're reliable.
Reliable on day one
- Phone numbers — "My number is 555-0199."
- Addresses — "We're at 4827 Oak Street, Springfield."
- Dates & times — "Can we do Tuesday afternoon?"
- Names — "It's for Maria Chen." (Spelling out helps.)
- Email addresses — if the caller spells it.
Gets sharper with use
- Industry-specific job types — "The boiler is short-cycling," "the GFCI keeps tripping," "we need a Tier 2 inspection." The AI doesn't ship knowing your vocabulary.
- Product names and SKUs — anything unique to your menu, catalog, or service list.
- Internal codes — work-order numbers, policy numbers, the system your CRM uses.
What trips it up
- Two people talking over each other.
- Heavy regional accents on uncommon proper nouns.
- Callers who change their mind mid-sentence ("Tuesday at three — actually no, Wednesday").
Make it better
Correct facts at the moment of capture, not at wrap-up. Each correction adds a training signal that shows up in the Steward's overnight runs and tightens future captures. Add "say-to-confirm" phrases to fields in your workflow editor — gives the AI a literal anchor to match against.
Turning it off for a field
Workflow editor → click the field → Field Settings → Disable AI capture. Staff types it manually from then on. Useful for fields where the legal cost of a wrong value beats the time saved by auto-fill.