InsuranceMid-sized general insurer, ~1,800 handlers

Claims-handler AI copilot for a mid-sized insurer

A governed AI copilot embedded in the claims-handler workspace — drafting responses, summarising case files and flagging escalations under a strict maker-checker model.

Tools / product used · Vestval Flow + Custom AI agents

Challenge

Claims handlers spent most of their day stitching context across PDFs, emails and the policy admin system. Response quality varied widely…

Solution

A handler-side copilot built on Vestval Flow that summarises case files from retrieval over policy and claim documents, drafts response…

Architecture

We profiled three handler personas over two weeks, designed a copilot that drafts but never sends, and instrumented every suggestion with a…

Timeline

4-phase implementation · Vestval Flow + Custom AI agents

Impact

Average handle time reduced materially across pilot cohorts (qualitative)

Challenge

Claims handlers spent most of their day stitching context across PDFs, emails and the policy admin system. Response quality varied widely between handlers and supervisors had no leverage.

Objectives

Cut average handle time without eroding response quality
Standardise response language across handlers without making it robotic
Give supervisors a defensible review surface for AI-assisted decisions
Keep every AI output reviewable, explainable and overridable

Approach

We profiled three handler personas over two weeks, designed a copilot that drafts but never sends, and instrumented every suggestion with a reviewable trail before rolling out cohort by cohort.

Solution

A handler-side copilot built on Vestval Flow that summarises case files from retrieval over policy and claim documents, drafts response language, and routes anything with escalation signals to a supervisor queue with full reasoning visible.

Implementation approach

1
Handler shadowing first
Two weeks shadowing three handler personas before any UI was sketched. The copilot was designed around their real day, not the documented one.
2
Draft-not-send by design
Every AI output is a draft that requires a human action. No outbound communication is sent by the model. This is encoded in the workflow, not in convention.
3
Supervisor escalation queue
Escalation signals (regulatory keywords, vulnerable-customer markers, dispute language) are routed to a supervisor queue with the model's reasoning attached.
4
Cohort rollout with kill-switch
Rolled out one handler cohort at a time, with per-cohort metrics and a per-tenant kill-switch the operations lead can pull without engineering.