FNOL triage without leaving ClaimCenter

The problem

35% of calls get escalated because the SharePoint knowledge base answer isn't fast enough. Average handle time 14 min vs. a 9 min target; the top quartile of agents is 45+ min. In a hailstorm rush a queue stretches past 45 minutes. After each call, the specialist manually copies info into ClaimCenter — 12% error rate on that entry. A bolt-on 3rd-party tool was rejected by IT as a 6-month integration.

What we'd ship

A real-time agent copilot embedded inside Guidewire ClaimCenter. Azure Speech-to-Text on the Genesys audio, Azure OpenAI retrieving from the SharePoint KB and PolicyCenter / Oracle data, pre-filling the claim record as the specialist talks. Every pre-fill comes with a confidence score and a source citation — the specialist accepts, edits or rejects per field. Decision authority stays fully human.

Stack

• Azure Speech (real-time transcription on Genesys feed)
• Azure OpenAI (GPT-4o) for intent + KB summarisation
• Azure AI Search indexing the 800-article SharePoint KB
• Guidewire ClaimCenter REST API for claim pre-fill
• Copilot Studio front-end embedded in ClaimCenter

Guardrails & compliance

• EU AI Act high-risk compliant — assist-only, every suggestion has confidence + citation
• Explicit accept / edit / reject per field; decision authority stays with the specialist
• Existing 90-day Genesys retention policy unchanged; no new personal-data pipeline
• KB refresh process documented so stale answers don't slip back in
• ČNB-friendly: every automated decision is explainable and audit-logged

Typical timeline

1. Week 0–2 Knowledge-base cleanup + shadowing

   Sit with specialists, audit the 800-article KB, flag the 1-year-stale articles for refresh, set success metrics with the claims director.

2. Week 2–6 Pilot with 5 specialists, off-peak

   Ship the in-ClaimCenter copilot to volunteers, measure AHT + accuracy + satisfaction, tune retrieval and summarisation.

3. Week 6–14 Full rollout incl. peak handling

   Enable across all 35 FTE, validate behaviour under hailstorm-rush load, wire the ČNB-facing decision audit.