- Problem
Two assessor teams needed an AI risk signal on an already dense claims page. - What I changed
A risk tag in the bar they already use, with the full assessment right before the decision. - Key result
0 new patterns for assessors to learn. - My role
Sole designer, across both models and both teams.
Fraud, waste and abuse inflate the costs paid on insurers' behalf, and manual review can't keep up as claim volume grows. The AI team built two models: one scores clinic behaviour on panel claims, the other scores member behaviour on non-panel claims. My job was to design where and how the risk signal appears to the two teams who act on it.
Two teams, two definitions of suspicious.
Same question for both: where does an AI risk signal belong on a dense claims-review page? But two handoffs, and two ideas of what looks wrong.
- Panel claims, provider management team. Reviews diagnosis and medication mismatches, duplicate prescriptions and excessive quantities. Acts by contacting clinics to amend claims.
- Non-panel claims, claims assessors. Reviews receipt authenticity and how often a member submits. Acts by asking members to resubmit documents.
Before this, I spent a year building the drug-to-diagnosis tagging in the provider admin portal. That structured data is what the panel model now scores against.
Five directions, one that could ship.
AI pattern libraries had nothing for an always-on signal inside a dense professional tool, so I looked for patterns with the same behaviour: visible while scrolling, supporting action without demanding it.
- Inline section. Only visible when scrolled to, so assessors could judge before seeing the AI. Rejected.
- Floating chat widget. Always present, but detached from the claim. Early feedback: "feels like an ad." Rejected.
- Side drawer. Good depth, but covered the claim when open. Assessors were reluctant to open it. Rejected.
- Corner widget. The same detachment problem as the floating widget. Rejected.
- Slide-up panel. Integrated and expandable on demand. My preferred direction, cut by the sprint timeline.
Working within a tool they already trust.
1. A new AI widget→The bar they already use
What changed
The risk tag sits in the sticky claim-status bar, so it stays visible while scrolling. No new pattern to learn, and the workflow stays untouched.
2. Score at the bottom→Assessment before the decision
What changed
The full assessment sits right before Change status, in the assessor's decision path. Clicking the tag jumps straight to it.
3. A green "low risk" badge→No tag when nothing is wrong
What changed
Only High shows today. A green badge invites complacency on a professional judgment, so low-risk claims get no tag at all. Medium arrives as the model improves.
Designing the signal to scale.
- Grows with the model. Risk shows as High only for now, on purpose. As accuracy improves, High and Medium tagging follow, so the interface scales with the model's confidence.
- Feedback goes back in. I structured how daily audit feedback, confirmed fraud, false positives and new behaviours flow back into retraining, so the model keeps improving after launch.
What I took from it.
- Restraint, not spectacle. AI in internal professional tools should support the workflow, not announce itself.
- Shared patterns, not identical screens. When one system serves two teams, consistency comes from the pattern.
- Design doesn't end at launch. Structuring the model's feedback loop was a design decision too.
- Build for tomorrow's confidence. The interface should scale with the model, not freeze at launch-day accuracy.
