Case studies
Clients are anonymized. Each entry names what was failing, what we changed, and what we measured afterwards.
Clinical AI diligence for a venture firm
An investor was about to back an oncology decision-support product that drafted treatment plans and prior authorization requests. Model updates were changing recommendations without a version change, and a clinician who wanted to override a plan had to file a support ticket. The investor closed with safety covenants in the term sheet, and the company shipped an inline override within three weeks.
Regional care network
- What was failing: escalations were handled differently on every shift, and staff stopped trusting them.
- What we changed: mapped where the work piled up and rebuilt the escalation ladder around it.
- Result: escalation decisions 27% faster and 34% fewer after-hours handoff errors within 8 weeks.
Public-sector operations team
- What was failing: an audit was coming and nobody could say who owned each exception path.
- What we changed: ran working sessions with each team and set up a short decision memo the board reviews on a fixed schedule.
- Result: every exception path had a named owner within 6 weeks.
AI operations program
- What was failing: operators were burning out on repeated emergency overrides.
- What we changed: ran safety drills and added decision checkpoints before the points where overrides usually happened.
- Result: 41% fewer emergency overrides the following quarter.
Field notes
Shorter write-ups from teams we learned from, each with one thing you can try this week.
- Rebuilding the escalation ladder at a benefits helpline: three triggers and a relief rotation cut stalled cases by 42%.
- Pausing an AI launch until every partner could sign off: one steward per partner, one evidence log, one restart checklist.