Role
Care Admin
Owns the strategy but couldn't build it. Depended on engineering for every change.
Clinical Engineering
Translated intent into JSON. The bottleneck for every strategy, edit, and fix.
Care Manager
Receives the patients a strategy assigns. Downstream of every decision made here.
Patient
Indirectly impacted. Who qualifies and who's missed - is decided upstream.
How the strategy worked
Owned by care admins, built by engineers - fully understood by neither side at the same time.
Confidentiality Notice
To respect confidentiality agreements, certain visuals have been blurred or omitted. If you'd like to learn more, I'd be happy to walk you through the original screens, design decisions, and outcomes during 1:1 conversation.
What it cost-
Two phases. One problem, evolving
Phase 1
Making strategy creation accessible
Every filter and rule becomes an addressable, debuggable step. Workflow logic becomes spatially legible.
Every node must validate before the user can move forward. Broken logic cannot reach dry run; broken dry runs cannot reach production. Errors surface at the node, not in production.
So why wasn't this enough?
We made the interface easier - but the thinking was still difficult.
Care administrators could now configure workflows visually - but they still relied on institutional knowledge to decide:
Which patients should qualify?
Which interventions should trigger?
Which assignment logic should be used?
Which rules were missing?
The challenge shifted from configuration to decision-making. That's where AI became meaningful.
Phase 2
AI isn't the product here - it's the next logical step: a decision-support layer on top of a builder people could already use. What follows is the research that revealed what to build, and the five design decisions it led to.
To support how admins decide, I first had to see how they decide today. I shadowed care admins building strategies and traced the JSON tickets and PDFs they'd been sending engineering for months.
The thinking was already happening - just off-platform, in flowcharts and institutional memory.
Five observations from that work became five design decisions:
Different users arrive in different mental states. Some know exactly what they want. Some need a starting point. Some are reacting to an operational alert. Some are working from organizational patterns.
Care admins were already drawing strategy flowcharts and sending them to engineering as PDFs. I designed the AI to accept those directly as chat attachments - extract the logic, ask clarifying questions, finalize the configuration.
Same input, same output, every time. In healthcare, an auditor asking "why was this patient assigned?" can't get a different answer next Tuesday. No freeform LLM at runtime.
The AI scores the prompt, shows what it understood, asks only what's missing (max 4 questions), and presents a tagged Blueprint. The canvas opens only after the user confirms - with every node pre-configured.
Every refusal redirects to something productive. "I can't predict outcomes, but I can run a test on current data." "I can't bypass validation - want me to help fix the issues?"
After a strategy goes live, the AI tracks performance against its KPI and suggests refinements grounded in real data. Create → Publish → Measure → Suggest → Refine.
Prototyped in code with Claude in VS Code, integrated with the design system, deployed as a live product. Leadership reviewed it as an interaction, not a mockup.
What the AI refuses shaped this product more than what it offers. Every guardrail was a design decision, not a limitation.
Same input, same output, every time. That's not a tradeoff against intelligence. It's the requirement that earns operational trust.
Care admins were already drawing flowcharts. Designing the AI to accept those - instead of asking users to restate their thinking in a chat - was a small choice with the biggest impact.
Most AI products optimize for the first interaction. Designing performance tracking so the AI proves its value over time - not just at the moment of creation - was the move nobody else on the team had considered.
