Agent Registryclaims-investigation-agent

claims-investigation-agent

AWS Bedrock AgentCore

Multi-step agent for investigating complex insurance claims, coordinating policy lookup, claim history retrieval, photo analysis, and adjuster interaction.

OwnerCatherine O'Brien·In CertificationHigh under EU AI Act · Critical under SR 11-7
Last sim pass
87.4%
running now
Open findings
4
3 critical
Tools
6
3 read · 2 call · 1 write
Sub-agents
2
2 specialists
Days to cert
21
re-cert window
Task completion
92.4%
last 30d
Production
Cert
in queue

Behavior summary

Business purpose
Reduce average claim investigation time from 2.4 days to under 6 hours, while preserving adjuster oversight on payout decisions.
Decisions affected
Claim triage prioritisation, evidence requests, payout-amount recommendations. Mapped to EU AI Act Annex III §5(c) and SR 11-7.
Data subjects
Claims adjusters · policyholders · claimants (incl. third-party).
Human oversight
Human-in-the-loop for any payout > $5,000. Human-on-the-loop for all other agent decisions with daily sampled review.
Escalation
3 failed tool calls → claim team queue. Detected adversarial pattern → security on-call. Cross-jurisdiction ambiguity → senior adjuster.

Capability surface

Tools6 (3R · 2C · 1W)
MemoryLong-term per-claim · 90d
Sub-agents2 specialists
Token budget50K / task
Wall-clock5 min / task
Cost cap$0.50 / task

Recent behavior

Behavioral drift +3.4% vs v17
  1. Simulation 200-scenario behavioral test started· Catherine O'Brien
    38m ago
  2. F-2026-04-1247 escalated to critical· indirect-injection via RAG documents
    2h ago
  3. v18 promoted to Lab· Vikram Shetty · 2 tools added
    1d ago
  4. Continuous red-team — multi-claimant variant — pass 89%· Arjun Iyer
    1d ago
  5. Cross-version diff vs v17: no behavioral regressions on 188/200 scenarios· auto
    1d ago