Agentic AI Insurance Claims

Designing the trust layer for AI-driven decisions

→ Read the full case study here

Where Trust Breaks (The Problem)

Insurance claims are slow and fragmented. AI improves speed but it also creates a trust gap. Decisions are made instantly but users are left without context, reasoning or control.

  • Drivers don’t understand decisions

  • Adjusters don’t trust AI outputs

Outcome: Decisions happen faster than humans can verify them

What Needed to Change (The Goal)

The goal wasn’t automation, it was making AI decisions understandable, accountable, and usable by humans. Design a claims experience that:

makes AI decisions transparent

keeps adjusters in control

reduces uncertainty for drivers

System Design Moves (What I did)

  • Exposed the “black box” failure in AI decision-making

  • Separated driver and adjuster workflows because their trust requirements and decision responsibilities were fundamentally different:

    • drivers needed reassurance and visibility

    • adjusters needed auditability, override capability, and operational speed

  • Turned AI outputs into explainable, editable decisions instead of presenting them as final answers

    • surfaced reasoning and confidence signals inline

    • created override paths for adjusters

    • framed AI recommendations as collaborative inputs, not automated conclusions

  • Balanced transparency with operational usability

    • too little reasoning reduced trust

    • too much detail increased cognitive load and slowed decision-making

    • optimized for inspectability, not full model transparency

  • Design Trade-Offs:

    • expose enough AI reasoning to build trust without overwhelming adjusters with model-level complexity

    • more transparency increased confidence but too much detail slowed decision-making and added cognitive load

    • make AI reasoning inspectable, not exhaustive

  • Reframed prioritization around regulatory deadlines, not model confidence

    • the system originally prioritized based on AI confidence scores

    • operationally, claims urgency was driven more by regulatory deadlines, legal exposure, and customer escalation risk than model certainty alone

  • Defined trust operationally through:

    • override frequency

    • adjuster acceptance behavior

    • claim escalation patterns

    • confidence visibility

    • and time-to-verification.

Outcome: The system was designed not just to improve AI usability, but to reduce verification friction, support regulatory compliance, and increase operational confidence in AI-assisted claims handling.

What Changed (The Impact)

1. Increased trust in AI decisions by exposing reasoning and confidence

2. Reduced adjuster cognitive load by replacing fragmented tools with a unified system

3. Improved decision clarity by turning AI outputs into explainable, editable inputs

4. Created a human ↔ AI feedback loop through override and audit mechanisms.

Key Insight

AI doesn’t fail when predictions are imperfect but fails when humans lose the ability to inspect, challenge, and confidently act on its decisions.

Designing for explainability, overrideability, and operational visibility turns AI from a black box into a usable decision system.

Back to the top?