A high-stakes underwriting workflow where AI support had to remain legible, challengeable, and human-led.
- Year
- 2026
- Focus
- Risk workflow clarity
- Role
- Product design, AI workflow definition, interaction models, prototyping
- Themes
- AI workflows, Insurance, Decision support, Enterprise UX, Explainability
How can an expert trust, question, and act on recommendations in high-stakes work?
Case specimen
Risk workflow clarity
- Messy inputs
- Submission signals · Risk exceptions · Model confidence · Audit trail
- Working frame
- Separate AI observations from recommendations so underwriters can inspect the evidence before accepting or challenging an action.
- Decision enabled
- Decide whether to proceed, question the recommendation, or override the agent with a documented rationale.
- What changed
- The prototype shifted the conversation from automation speed to accountable expert judgement.
The situation
Context
Risk assessment is a knowledge-heavy workflow with many signals, exceptions, and institutional practices. The product opportunity was not simply to automate review, but to make expert decision-making more legible and supported.
Problem
The core challenge was helping experts understand why an AI agent recommended an action, what evidence it used, and where human review remained essential.
The work
My role
I shaped the workflow narrative, mapped agent-human handoffs, designed confidence and evidence patterns, and built prototypes to make the concept testable.
What made it hard
The hardest part was balancing speed with trust. Too much automation could feel opaque; too much explanation could slow the workflow back down.
Process
I started by mapping underwriting moments of judgement, then translated them into states: intake, triage, evidence review, recommendation, challenge, decision, and audit trail.
Key design decisions
Key decisions included separating AI observations from AI recommendations, showing confidence as a qualified signal, and giving the human reviewer explicit ways to accept, question, or override outputs.
What it changed
Outcome
The prototype created a concrete conversation around the future of underwriting, agentic systems, and the design requirements for trustworthy AI adoption.
What I learned
AI workflows need interaction patterns for disagreement. Trust grows when the system is clear about what it knows, what it assumes, and what remains unresolved.
What I would do differently
With more time, I would test the confidence language with underwriters and compare how different explanation depths affect speed and decision quality.