An illustrative reference architecture, not a customer
case study. No organisation is named and no deployment is described - this is
what the layers look like for a insurance organisation,
so you can map it onto your own estate.
Channelswhere the conversation starts
Inbound numberspublished lines
SIP / contact centreexisting estate
Outboundcampaigns and callbacks
In-appbrowser WebRTC
▼
VoiceKernelthe agentic voice layer
Conversationagent, voice, turn taking
Knowledge groundinganswers with citations
Tool callsreads and writes back
Transferto a human, with context
Tenant isolationownership registry
Audit and erasureper call, per mutation
▼
Integrationyour boundary, your rules
REST + typed SDKscoped API keys
Signed webhooksretried, replayable
Live event streamSSE
▼
Insuranceyour systems of record
Policy administrationcover, excess, endorsements
Claims systemFNOL, status, payments
Document storephotos, PDS, correspondence
CRMrenewal and retention
Paymentsexcess collection
Only the bottom layer is specific to insurance. The
three above it are identical in every deployment, which is the point: the
voice layer does not need to know what industry it is in, only which systems
it is allowed to reach and what it may say.
What it changes
Mechanisms, not projections. Each of these follows from something structural
about the work, so you can check whether it holds for your own queues.
FNOL is captured completely, first timeA fixed question set asked identically means fewer callbacks to fill gaps that stall the claim.
Claims start moving immediatelyThe claim exists in the system before the call ends, rather than after a queue of voicemails is worked.
Advice risk is bounded by designThe agent cannot recommend, because the tools to do so are not available to it.
Renewal conversations happen at scaleRetention outreach becomes a volume exercise rather than a staffing decision.
What the agent actually does
First notice of lossCapture what happened, when, where and who else was involved; create the claim; send the upload link before the call ends.
Claims statusAnswer from the claims system rather than a queue, including what is blocking and what the customer must do next.
Renewals and retentionExplain the premium change against last year, and transfer to a licensed human the instant the conversation turns to advice.
What keeps it deployable
General-advice guardrailsThe agent states factual information and refuses to recommend, with the boundary in the prompt and enforced by tool availability.
Evidence trailEvery claim field the agent captured, with the utterance it came from.
Consent captureRecording consent taken and stored before anything else.
FNOL calls are long, highly structured, and almost entirely fixed questions - the shape of work a voice agent handles without judgement.
NextRead the source
Everything above is in the open-source Community edition. Nothing on
this page needs a licence.
NextRun it yourself
Point it at your provider account and put one queue through it before
you talk to anybody.
NextThen buy the guarantees
An Enterprise agreement adds the SLA and the
support contract when it carries real volume.