- Client
- MedTech company based in Chicago, IL, serving consumers and organizations such as hospital systems, insurers and employers
- Function
- Product, clinical and member engagement
- Headline result
- Members completing a recommended preventive action within 90 days rose from 24% to 41%.
The challenge
Preventive care fails less from lack of knowledge than from lack of follow-through. Most people do not know which screenings they are due for, and advice rarely arrives at the moment they can act.
The company had a working platform for surveys, reminders and prescriptions, but three gaps limited its impact:
- Static engagement. Long questionnaires drove drop-off, and recommendations were not tailored enough to feel personal.
- No one to ask. Members with a question between appointments had nowhere to turn, so they often did nothing.
- An assistant not ready for production. An early AI prototype showed promise, but it lacked the reliability, evaluation and safety controls that health guidance demands.
The question for leadership was not whether AI could answer health questions. It was whether AI could change what members actually did, safely and at scale.
Our approach
We rebuilt the platform around a simple loop: learn about the member, recommend the next action, prompt at the right moment, and answer questions along the way. Rules decide what is recommended. AI explains it and supports the member.
The work had four parts:
- Adaptive intake. Rule trees decide who is asked what. Surveys are capped at 10 questions and skip what is already known, so each session stays short.
- Deterministic recommendations. Explicit rules trigger each preventive action, with deduplication and cooldown periods to avoid fatigue. A simulator shows administrators exactly why a rule fired. A language model never decides what care a member should receive.
- An AI health assistant. Members chat on the web or by text. Answers draw on a curated clinical knowledge base and remember the member's context.
- Everyday medication support. Members capture a label with a photo, receive dose reminders and log doses as taken, late or missed. Interaction checks run automatically, and family profiles let caregivers help.
Rules decide what to recommend; AI helps members act on it
We then began a six-week production-readiness program for the assistant. Foundation and reliability work is complete, covering observability, streaming, durable jobs and rate limiting. Quality evaluation, clinical review, compliance hardening and launch readiness are the next stages.
Impact
In the months after relaunch, engagement and follow-through improved on every tracked measure.
| Measure | Before | After | Change |
|---|---|---|---|
| Recommended preventive action completed within 90 days | 24% | 41% | +17 points |
| Members current on age-appropriate screenings | 52% | 68% | +16 points |
| Medication doses logged as taken on time | 71% | 86% | +15 points |
| Survey completion rate | 38% | 74% | +36 points |
| Monthly active members | 22% | 47% | +25 points |
The assistant handles about 60% of member questions without a human. For organizational customers, closed preventive-care gaps give a clear measure of value.
What made it work
- AI explains; rules decide. Keeping recommendation logic deterministic made every decision explainable and auditable.
- Timing mattered more than content. A short prompt at the right moment outperformed longer education.
- Reliability came before scale. Monitoring, cost tracking and abuse controls went in first, and clinical review has its own milestone in the plan.