Case study · Insurance · Claims
How a leading Indian life insurer aims to cut manual claims effort by up to 70% with ShellKode FDEs.
Leading Indian life insurer · Life insurance
Assessors spent hours on paperwork. Now agents do it, and people make the decisions.
≥80%
Straight-through processing target for eligible claims
50–70%
Targeted reduction in manual effort
80–90%
Field-level extraction accuracy target
120–180 min
Manual handling time per claim, before agents
The challenge
Every death claim was read, keyed in and checked by hand.
For a grieving family, a death claim is the most important conversation they'll have with their insurer. Behind it, an assessor spent two to three hours per claim reading documents, re-entering data across several systems and running verification checks one at a time.
A standard automation project wouldn't fix it. The work ran across a policy admin system, a claims platform, an internal claims system and external verification services, none of which could change. Deciding when an agent should act and when a human must step in also couldn't be specified in advance. It had to be worked out claim by claim with the claims team. That called for Forward Deployed Engineers working inside the claims process, not designing it from outside.
Fully manual handling
Every claim took 120–180 minutes, end to end.
Systems that don't talk
Claim data sat across four platforms and outside services.
Document-heavy reviews
Forms, certificates and policy records were read line by line.
Almost no straight-through processing
Even simple, eligible claims needed someone to touch them.
What we delivered
Agents that do the legwork. Guardrails that know when to stop.
ShellKode FDEs embedded with the insurer's claims and technology teams to answer two questions on every claim: can an agent handle this safely, and how do we know it got it right?
Four agents, one orchestrated flow
Built with Strands Agents on Amazon Bedrock AgentCore, agents validate, register and verify each claim, then hand off.
Rules as configuration, not code
Business rules and document checklists can be refined by the claims team without rebuilding anything.
Every step checked before the next
Amazon Textract extractions and NSDL and IIB verifications are validated, and any gap or failure goes to a person.
An assessor-ready summary
Each claim arrives verified, with what matched, what didn't and why, logged end to end for audit.
The outcome
Assessors now start from a verified claim, not a pile of documents.
≥80% straight-through target
Eligible non-early claims move end to end without manual handling.
50–70% less manual effort targeted
Agents take on validation, data entry and registration.
80–90% extraction accuracy target
Field-level accuracy, measured before a claim moves forward.
Humans keep the final say
Every decision stays with an assessor, with a full audit trail.
The stack
What it runs on.
- Agents
- Strands Agents on Amazon Bedrock AgentCore
- Document extraction
- Amazon Textract
- Verification
- NSDL, IIB
- Systems of record (unchanged)
- Policy admin system, claims platform, internal claims system
What's next
The workflow is in place. Next comes moving more claims onto it.
With intake, verification and summaries automated, adding a claim type is a configuration change, not a rebuild.
- 01
Extending straight-through processing to more claim types
- 02
Tuning rules and checklists from assessor feedback
- 03
Measuring the targets against live claim volume (~83 a day)
Talk to us
Bring one manual workflow. Leave knowing which parts an agent can safely take.
30 minutes on one process, mapped to the agents and guardrails behind it. No slides.