Operations-heavy businesses automating a review workflow where mistakes are expensive and a human still has to own the outcome.

Every insurance claim was reviewed by a person. Every reviewer made errors, and each error either paid a fraudulent claim or delayed a legitimate one. The medical expense ratio was climbing. The team had the models; what they did not have was a workflow that let humans and models share the work without either one becoming a bottleneck.

Mapped the existing claims workflow end to end with the operations team, then classified every decision point by confidence and consequence - which decisions the model could own outright, which needed a human, and which needed a human plus an audit trail.

Designed the exception queue as the primary interface rather than an afterthought. Reviewers now spend their day exclusively on ambiguous cases, with the model's reasoning and confidence surfaced next to each one so a human can accept, override or escalate in a single view.

Designed fraud anomaly flags to be explanatory, not just red. A flag that says only 'suspicious' gets ignored within a week; a flag that says why gets acted on.

Built real-time status tracking for both the operations team and policyholders, which removed a large volume of "where is my claim" contact.

Designed the override path deliberately: when a reviewer disagrees with the model, that disagreement is captured as training signal rather than lost.

Three months as design lead, working alongside the claims operations team, data science and engineering. I sat with claims reviewers doing their actual job before designing anything - most of the important constraints were things nobody had thought to write down.

Claims processing 48% faster

Human error rate down 64%

Medical expense ratio reduced by 23%

NPS improved from 45 to 63

Figma

Workflow architecture

Human-in-the-loop AI patterns

Fraud detection UX

Human-in-the-loop workflow design for an operations team introducing AI. Typically 4–8 weeks: workflow mapping, confidence-and-consequence decision model, exception queue design, override and audit patterns.

Typical investment: $5,000 – $10,000 depending on workflow complexity. Discovery-only engagement: $2,000.