Teaching an AI health coach to think like a clinician.

Before Ask Allie, SMART goal-setting at Codex Health depended entirely on a clinician being present. They'd walk the patient through the process during a time-limited visit, or email a worksheet and hope the patient remembered to fill it out and return it. There was no way for a patient to set a goal on their own, track it in the app, or have Allie respond to their progress.

The opportunity was to change how care is delivered — not just improve a UX flow. If Allie could hold the goal-setting conversation the way a skilled clinician does, patients could set meaningful goals on their own time, track them independently, and give their clinician something real to work with at their next visit.

This was also the first AI feature in the platform to update a patient's care plan. The stakes were higher than a typical design problem.

We started by interviewing our clinical team — what do they believe makes a great SMART goal? They described a thorough, structured process: walk the patient through each letter, make sure every component is covered. That became our first training approach.

We also ran mock goal-setting sessions between clinicians and employees, designed to simulate what a health coach would do. Those sessions were helpful — they captured the warmth and adaptability of a real coaching conversation — but they were also long, unhurried, and focused entirely on goal construction. They didn't reflect the reality of a clinical visit.

What clinicians actually do in a time-limited patient visit is something different. Faster. More intuitive. They acknowledge barriers early. They don't walk through every letter. They focus on getting the patient to a goal that's close enough and achievable — something that gets them moving — rather than constructing a perfect SMART goal from scratch.

I made a mistake early in this project. I trusted the clinical team and the data scientist to structure the conversation based on what they believed was clinically correct — and I didn't push hard enough on whether that translated to a good patient experience.

The result was a conversation that walked patients through every letter of SMART in sequence. Technically thorough. Practically too long. Patients dropped off before finishing. The output wasn't meaningfully better than a much shorter conversation would have produced.

The final conversation was shorter, warmer, and barrier-aware. It acknowledged what the patient was up against before asking them to commit to anything. It produced goals that patients actually tracked — because they felt like their own, not something filled out on a worksheet.

The shift wasn't just about a better UX flow. It fundamentally changed the relationship between patients and their care team.

Cutting the conversation to S · M · T reduced the time to set a goal significantly and increased the number of patients who completed goal-setting. We had a clean cohort to study: whether setting and completing a SMART goal correlated with health outcomes — that analysis was in flight when I left.

The full metrics aren't public, but I'm glad to walk through them in person.