Automated motor insurance claim intake.
Within 5 months of introduction to an insurer, Voxworks designed and deployed an AI voice agent to handle after-hours motor claims calls. Voxworks runs the full first notice of loss conversation, writes a structured FNOL into the claims management system and reads the claim reference back to the claimant before the call ends, saving the local team hundreds of hours of call handling time.
- 400+
- data points captured
- 15-20 min
- claims-officer time returned per lodgement
- 93%
- deflection rate
- 3
- motor brands covered
Claims administrator for a global motor insurer
An Australian third-party claims administrator handling motor claims for a global Fortune 500 insurer, headquartered in the USA, that writes motor policies in Australia under three brands (client and brands anonymised to protect commercially sensitive information).
Three brands, one claims line, and claims that arrive at any hour.
The insurer’s three motor brands each serve a different group of drivers, and all of them report claims to the same claims line. Car accidents don’t keep business hours, so neither can claims intake.
A motor first notice of loss takes ten to fifteen minutes of structured questioning. Covering that around the clock with people means staffing a phone line overnight for calls that arrive unpredictably, or letting claims wait until the next working day.
Average motor claim costs rose 42% in five years (Insurance Council of Australia, 2025). The claims administrator needed intake that scales with claim volume: every claim captured to the insurer’s standard and lodged straight into the claims system, without adding claims staff for every new brand.
From first meeting to live claims in under six months.
Claims automation has to clear legal, compliance and underwriting review before a single live call. Working alongside the claims administrator, Voxworks went from a discovery meeting in February to underwriting sign-off and live claims in August, an exceptionally fast timeline for insurance.
- 9 days
- from kick-off to the first script in testing
- 5 months
- from signed agreement to live claims
- Feb 2026Introduction and discovery meeting
- Mar 2026Enterprise agreement signed
- Mar 2026Project kick-off
- Apr 2026First script tested
- Jun 2026Five Sigma integration completed
FNOL lodged into the claims system from the call.
- Jul – Sep 2026Script improvement and quality control
Agent-to-agent persona testing and human testing.
- Aug 2026Underwriting sign‑off and go‑live
Live after-hours claims intake for all three brands.
Every claims call carries risk.
Claims handling is a licensed financial service carrying legal and financial risk, directly related to what the voice agent hears and how it responds.
The agent must never create cover that doesn’t exist
One sentence that sounds like confirmation, and the insurer may have to honour it. Estoppel, or misleading conduct, can follow from a single spoken line.
- Speech recognition: jargon, slang and word error rates
- Language processing: hallucinations, intent handling and variability
- Speech synthesis: pronunciation and abbreviations
Indemnity can’t be decided on corrupted facts
FNOL is the input to a coverage decision, not a record of one. Wrong facts in means a wrong decision out, and the error surfaces at reserve review rather than on the call.
- Speech recognition: accents, demographics and jargon
- Language processing: interpreting a bad transcription
A vulnerable customer has to be recognised
The General Insurance Code of Practice obliges insurers to identify and support vulnerable customers. Those disclosures arrive obliquely, mid-sentence, from a caller who is not asking for help.
Every call has to be defensible afterwards
Each interaction has to stand up at AFCA, in a dispute or in a Code audit. The insurer must be able to show what was said, and that it was said consistently from one call to the next.
After-hours claims intake, end to end
Voxworks answers the call, runs the claim conversation, writes a structured First Notice of Loss into the insurer’s system of record, and provides a claim number before the caller hangs up.
What this system does
- 01
Answers after-hours claims calls for all three brands
- 02
Runs the full motor FNOL conversation, not a message-taking script
- 03
Lodges a structured FNOL in Five Sigma during the call
- 04
Reads the claim reference back before the caller hangs up
- 05
Uses authored, deterministic responses wherever cover is at stake
- 06
Hands off live with a spoken summary when a caller needs a person
- motor FNOL
- three motor brands
- no one on shift
- intent pathways · mid-call redirects confirmed
- discrete classifiers · authored conditions
- routes minimised by Karnaugh mapping
- structured FNOL lodged
- reference read back
- claim ready for the team
Global handoff at any step. When a caller needs a person, the agent escalates live and gives a spoken summary of the claim so far.
Online claim form by text, as an option. When a claimant would rather finish online, the agent texts them a link to the insurer’s online claim form.
Emergency tow truck handling. When a vehicle needs towing, the agent hands the caller live to the tow truck operator, and keeps recording and transcribing the call after the handoff.
Australian-built AI technology, handling real claims from real customers.
The claims administrator signed an enterprise agreement in March 2026, and every script was tested agent-to-agent and by people before Voxworks went live on out-of-hours claims intake for all three brands in August.
In the first nine weeks live, the agent answered 205 calls on the claims line and lodged 69 new claims end to end in Five Sigma. 93% of calls were handled without a transfer to a person, with tow calls handed to the tow operator by design.
The integration with Five Sigma allows the output of every call to be lodged in the system of record, rather than a transcript or voicemail that someone has to re-key the next morning.
Unlimited claims capacity without new headcount
Every minute the agent spends on a claims call is a minute the claims team spends on the claims that need a person.
- 205
- claims-line calls answered
- 13.3 h
- of calls handled by Voxworks
- 69
- new claims lodged in Five Sigma
- 93%
- handled without a transfer
Minutes handled each week
Call minutes taken by the voice agent, by call type
- New claim
- Existing claim
- Tow
- Complaint
- Other
Deflection rate
Calls handled end to end, without a transfer to a person. Tow handoffs to the tow operator are by design and count as handled.
Where the minutes went
Share of all call minutes, by call type
- New claim596 min · 75%
- Existing claim68 min · 8%
- Other50 min · 6%
- Tow43 min · 5%
- Complaint39 min · 5%
The charts cover nine weeks of calls, from 19 August to 20 October 2026. Deflection is calls handled by the agent as a share of calls either handled or transferred to a person; tow handoffs to the tow operator are by design and count as handled.
What a regulated claims line taught us.
We specifically engineered Voxworks to address the risks of using AI voice in regulated industries like insurance. Every layer of the stack is built to detect the caller’s intent and align with policy before the agent speaks a word.
- 01
Transcription is always the fundamental challenge in voice AI
Most risk arises with the agent mishearing what the claimant said, therefore there is getting transcription correct is the most important element. Voxworks developed its own multi-layered audio processing and transcription algorithm: sub-second reflex models decide whether the turn has ended, whether the caller is interrupting and whether this is an emergency, while slower, deliberate models re-read what was heard with the whole conversation as context before routing, extraction and compliance decisions are made.
- 02
Determinism is the safety feature
Anything that could sound like a coverage decision is an authored response, chosen by routing against conditions the insurer has signed off. Borrowing Karnaugh mapping from electrical engineering, Voxworks collapses the combinations of caller intent into a minimal set of routes, so full coverage doesn’t need thousands of hand-written branches. Generated responses are only used where authorised, and are checked before any audio is played.
- 03
Australian callers need Australian speech models
Overseas models fail on Australian accents and on everyday nuances like “yeah, nah”, and a mis-heard answer on a claims call becomes a wrong fact on the claim. Keyterm tuning, mispronunciation detection and multilingual handling are built into the Voxworks engine, and the platform is Australian-hosted.
- 04
After hours is the right proving ground
Four months of agent-to-agent and human testing came before the first live call, and underwriting signed off before go-live. Starting with out-of-hours intake meant every call Voxworks took was a claim that would otherwise have waited until the next working day.
- 05
The time comes back through the system of record
The fifteen to twenty minutes returned per claim depend on the FNOL arriving in Five Sigma complete and structured. A voice agent that only takes a message shifts the work rather than removing it. Writing the claim directly into the claims system is what makes it a lodgement.