InsureSight set out to answer routine policyholder servicing calls without a person on the line — and without a menu tree. QyrosCloud built an agentic servicing platform on AWS where an orchestration agent decides what a caller needs, calls live carrier systems through a governed tool catalogue, and hands off to a human the moment a question leaves its scope.
It runs on voice and web from one implementation, in English and Spanish, for multiple carriers whose books are separated by IAM rather than by application logic. Between 25% and 45% of in-scope servicing calls are now resolved autonomously, and servicing no longer stops at close of business.
What was at stake.
InsureSight sells policy servicing to carriers and Managing General Agents. Its market is concentrated in non-standard personal auto, where policies carry lower premium, payment schedules are frequent, and servicing contact rates per policy are correspondingly high. That combination makes servicing cost per policy a first-order economic problem rather than an operational detail.
Routine questions with a single correct answer were being handled end to end by licensed representatives, only during staffed hours, and any platform that served more than one carrier had to prove the separation between their books before a carrier would put a policyholder on it.
“InsureSight partnered with QyrosCloud to build an agentic policy servicing platform on AWS. Amazon Connect and Amazon Q in Connect orchestrate an agent that calls a ten-tool catalogue over the Model Context Protocol through Amazon Bedrock AgentCore Gateway, with Amazon Nova Sonic answering in English and Spanish on one number. Two carrier tenants are live in production, each behind a boundary enforced by AWS IAM rather than by application code.”
How we solved it.
QyrosCloud built an agentic servicing platform on AWS where the model decides what to do and the architecture decides what it is allowed to do. The agent plans across the call and selects its own tools; its authority is an explicit, versioned catalogue, and the boundary between carriers is enforced by IAM rather than by application code.
Amazon Connect answers the line and an Amazon Lex bot fronts an Amazon Q in Connect orchestration agent that plans across the call — pre-identify from the calling number, verify identity, then resolve the actual question.
Turn-by-turn tool selection belongs to the model. There is no scripted flow behind it.
Amazon Nova Sonic on Amazon Bedrock carries the conversation as speech-to-speech over a bidirectional stream, rather than chaining speech-to-text, text generation and text-to-speech — each hop of which adds latency on a channel where silence reads as failure.
Nova Sonic detects English and Spanish, which removed the language menu entirely rather than shortening it.
The agent acts through ten tool schemas published over the Model Context Protocol by Amazon Bedrock AgentCore Gateway: identification, verification, policy, billing, ID card, payment link, callback, alerts, documents and claim status.
The catalogue is versioned in source control, so what the agent may do is reviewable rather than an emergent property of a prompt.
Each carrier has its own servicing function, its own execution role and its own credential in AWS Secrets Manager. The web tier resolves the carrier from the request host with no default — an unmapped host receives a 501 rather than another carrier's book.
The boundary is a permission, not a code path, so a mistake fails closed.
The web surface runs on Amazon ECS Fargate behind an Application Load Balancer, serving the same capability from the same tool definitions, so the phone and the web cannot drift apart in what they will tell a policyholder.
Every read passes through a shared wrapper that rejects an exception, a non-200 response, and a 200 whose body lacks the expected field. A contract test exercises five failure modes against every read tool discovered by reflection, so a tool added later is covered the day it is written.
The agent says it cannot check, rather than reporting absence.
Amazon Bedrock Guardrails are provisioned as infrastructure code in a per-tenant, per-intent matrix, so each carrier's configuration is independently auditable. Blocked interactions hand off to a human rather than dead-ending the caller.
Latency to the carrier's system of record was measured at a 0.55-second median with a 2.3-second cold-connection worst case, so reads run on a six-second timeout with one retry — replacing a transport default that could leave a live caller in silence.
“This engagement required us to balance speed with compliance rigor. We deployed infrastructure-as-code from day one, automated evidence collection across the environment, and delivered a production-ready architecture that passed security review on the first attempt.”
The results speak for themselves.
Between 25% and 45% of in-scope servicing calls are resolved without a human on the line. Two properties of regulated personal-lines servicing hold that figure below the deflection rates quoted in simpler self-service domains, and both are design decisions rather than defects.
Policy data is not disclosed until the caller is verified, and a caller who cannot complete verification is routed to a person rather than pressed further. And a policyholder calling about a balance frequently raises something outside scope — an endorsement, or a coverage change requiring underwriting judgement — at which point the agent hands over the whole call rather than answering half of it and transferring a partly-served caller.
The calls this agent declines are the ones a carrier would not want it to take.
InsureSight LLC provides agentic policy servicing to carriers and Managing General Agents in the United States, answering policyholder questions on the telephone and the web against each carrier's own system of record. Its market is concentrated in non-standard personal auto, where servicing contact rates per policy are high and servicing cost per policy is a first-order economic concern.
Related stories.
Ready for results like these?
Let's talk about your AWS environment.
Book a discovery call