How we put a clinician copilot in patient rooms in nine weeks

Nine weeks is fast. We know it sounds like a pitch. But the timeline isn't the interesting part. What's interesting is the sequence of decisions that made it possible, and the ones that would have killed it if we'd got them wrong.

The brief

Lattice Health had a documentation problem. Clinicians were spending 30 to 40% of their in-room time charting, summarising, filling in structured fields. Time that should have been spent with the patient was going to the EHR.

The ask: a copilot that listens to the consultation, surfaces relevant patient history, flags missing documentation fields, and generates a draft note for the clinician to review before they leave the room.

The goal was not to replace clinical judgment. It was to take the clerical load off it so there was more room for the actual judgment.

What week one actually looked like

We spent the first five days at two clinical sites. Not coding. Watching. We needed to understand the consultation rhythm: when the clinician looks at the screen, when they don't, what information is actually useful in the room versus what's only relevant after.

By Friday of week one we had a hypothesis: the highest-value thing the system could do was surface the last two relevant entries from the patient record right at the start of a consultation. Everything else was secondary. That single insight shaped the entire architecture.

The build

Weeks two through six were core build. We used a specialised language model fine-tuned on clinical notes rather than a general-purpose model, because in a clinical context, "almost right" isn't a category that exists. The pipeline was two stages: transcription, then structured extraction with a human-review step before any note goes into the record.

The interface was one screen. No onboarding flow, no tutorial, no help modal. If a clinician needed to be trained to use it, we'd built the wrong thing.

Five hospitals

By week nine: a live system running in a controlled pilot across five sites in two health networks. Not a demo, not a PoC. Real clinicians, real patients, real consultations. Documentation time dropped by an average of 22 minutes per clinician per day. Across a hundred clinicians, that's 37 hours of direct care capacity added every single day.