The brief: modernise the whole patient journey — without breaking the clinic
The MD didn't want to buy a new EMR. He didn't want a two-year enterprise migration. He wanted his patients to have the experience they get from every other consumer app — and his staff to stop losing an hour a day to phone triage that a well-designed intake form could handle. He'd seen three healthtech vendors quote 12–18 months and £400k+ engagements. He wanted 8 weeks and something his practice managers could actually use on the Monday after go-live.
"Our patients live on their phones. Our portal was still 2019 desktop. Growvate spent a week in our clinics before they touched Figma. That's why what shipped actually fits." — Practice MD, Northline Health
Week 1: UX research — sitting in reception
Before any design, we spent five days across the four clinics. Fourteen patient interviews (in the waiting room, before their appointment) and four full days shadowing reception staff. Three findings reshaped the roadmap:
- Patients trust plain English, not medical language. The 2019 portal asked "presenting complaint" — 71% of interviewees didn't know what that meant. The new portal asks "what's bothering you today?" and lets people type or record a voice note.
- Reception is the real product surface. 62% of bookings still came through the phone. The AI triage had to feed reception, not replace them. So we built a reception-side console that shows the AI's suggested urgency and category, and lets staff override with one click.
- No-shows are predictable. Shadowing showed us the pattern — same-week bookings on Monday mornings, patients under 30, first-time visits. The dashboard now surfaces high-risk appointments and auto-schedules a targeted reminder.
We ran a card-sort exercise with 8 reception staff to build the symptom taxonomy. Six categories, twenty-eight sub-categories, mapped to AAFP + USPSTF guideline references so the AI could cite its reasoning to the clinician.
Week 2–3: design system + clinical RAG
Two parallel sprints. The design team built an accessible-by-default component library (WCAG 2.2 AA, 16px minimum body text, high-contrast state changes, keyboard-first) packaged as Figma tokens and mirrored in code as CSS custom properties + React components. Colour, spacing, radius, type, motion, elevation — all tokenised so the practice-manager dashboard and the patient portal share a source of truth without looking identical.
The AI team built a clinical RAG index over AAFP + USPSTF guidelines, the clinic's own care pathways, and a red-flag symptom catalogue. Every AI response cites the specific guideline it drew from. Every response is reviewed by a clinician before the patient sees a clinical answer — the AI never diagnoses, only triages urgency and suggests the right clinician.
Week 4–5: building the loop
This is where most healthtech projects stall — the AI works in demo, breaks in production. We built an eval suite of 340 real (anonymised) intake cases and ran the triage agent against every one, checking urgency assignment against the clinician's after-visit conclusion. Baseline accuracy at week 4: 78%. After two rounds of prompt tuning + RAG index expansion: 94%.
The practice-manager dashboard was the surprise winner. What started as a small "capacity view" became the manager's Monday-morning cockpit — a single screen showing the week's appointments, no-show risk score for each, live capacity across all four clinics, and a triage volume trend.
Week 6–8: rollout across four clinics
Rolled clinic-by-clinic, a week apart. We onsite'd each rollout — the design team ran a 90-minute session with reception staff walking them through the AI triage console, and the AI team pair-worked with the practice manager for the first two days after go-live.
By week 8, all four clinics were on the new stack. By month three, no-shows had dropped 38%, patient portal weekly-active-users had grown 62%, and reception staff were reclaiming a documented 60 hours per clinic per week.
What we'd do differently
Two things. First, we'd have built the reception console before the patient portal — the clinic runs on reception, and their buy-in was the hard part, not the tech. Second, we'd have invested in the eval suite in week 1, not week 4 — the AI accuracy problem is unsolvable without one, and the earlier you have it the tighter the loop.
Both are now standard parts of our healthcare engagement playbook.