Quote-intake calls
Callers ask for a quote, share details over the phone and need a simple, trustworthy way to continue the application online.
Client work · Voice agent audit
Reliant Insurance: making a voice agent tell the truth about what it did
An AI receptionist that promises an SMS it never sent is worse than one that says nothing. We audited Reliant Insurance's ElevenLabs voice agent end to end — SMS consent, failure paths, call handling, privacy rules — and rebuilt the parts that could not be trusted.
12 test scenarios
Consent-based SMS
Safe-failure design
At a glance
- Client Reliant Insurance — insurance advisory, inbound and outbound calling.
- Agent "Victoria", an ElevenLabs voice agent on a dedicated Twilio number.
- Scope Configuration audit, prompt corrections, tool wiring, call-flow testing.
- Deliverable Written audit report with findings and the corrected configuration.
The client
An insurance agency piloting AI phone coverage
Reliant Insurance wanted to test whether a voice agent could handle quote intake over the phone — answering calls, capturing the essentials and sending the caller the approved application link by text. The agent was built and connected to a Twilio number; what it needed was a rigorous audit before it could face real callers.
Real service calls
Existing clients call about documents, cancellations and claims — conversations an unsupervised agent can easily mishandle.
Compliance boundaries
Insurance conversations carry licensing, privacy and truthfulness constraints that must hold under pressure and even during interruptions.
The problem
A voice agent that sounds confident can still be wrong
- The prompt still described SMS as unavailable even after the SMS tool was connected — so the agent refused a capability it had.
- There was no reliable state tracking: after a real tool failure, later sentences could claim the message was sent anyway.
- Consent handling was not explicit enough — sending a text requires a clear yes and a confirmed destination number.
- Email was being collected although no email action existed and the secure application collects it directly.
- Calls had no safe route for service issues or claims once the agent reached its limits.
- Transfer was enabled without a staffed destination — a transfer that rings nowhere is a broken promise.
Why it happens
Voice agents are assembled from prompts, tools and telephony settings that drift apart over time. Each piece can work alone while the combination silently misleads callers — which is why the audit tested the whole system, not just the script.
This describes what the audit found and corrected.
What we did
Audit, correct, test, re-test
The work covered the agent's prompt, its tools, its telephony behavior and its failure states — then proved the corrections with a scenario suite.
Truthful action handling
The prompt no longer claims capabilities are missing, and SMS success or failure is carried in persistent state — so a later summary or closing can never contradict what the tool actually returned.
Consent-first SMS
The quote link is sent only after explicit consent and confirmation of the destination number, with one-call/no-duplicate behavior and careful queued-versus-delivered wording. The sender and message body are locked to the approved branded link.
Dedicated dispatch with both paths
A Twilio dispatch step was added with separate success and failure branches, plus a truthful manual fallback — so a failed text ends in an honest statement and a next step, never a silent drop.
Prompt and call-flow corrections
Removed unnecessary email collection and email-delivery offers, tightened completion handling so callers who are finished get one brief closing instead of another question, and added cancellation warnings about coverage gaps and lender requirements.
Safe escalation instead of dead transfers
Transfer was deliberately left disabled because no safe staffed destination existed. Instead, callers get actionable manual routes — a contact path for service and human requests, and the carrier-contact route for claims.
Audio and turn handling tuned
Background-voice detection on, patient turn-taking, a 12-second response timeout, 45-second silent-call termination, the full opening greeting protected from interruption, and common backchannels ignored as acknowledgements.
Evidence
The audit's verified facts
The delivered audit report documents the configuration and the test suite that proves it.
12
call scenarios in the delivered test suite, from consented SMS to prompt injection
2
SMS dispatch paths wired and tested: success and failure with a truthful fallback
12s / 45s
response timeout and silent-call termination retained through the audit
0
real SMS sent during QA — simulations used per-test mocks
Verified build facts
These facts come from the delivered audit report and its test results. The agent is a controlled pilot, not an unsupervised full-service receptionist — and the report says so plainly. No call-volume or revenue claims are made.
ElevenLabsTwilioVoice agentSMSScenario testing
What was tested
Twelve scenarios, including the uncomfortable ones
The suite covers normal calls and the situations that damage trust if handled badly.
Quote intake and consented SMS
The happy path: details captured, consent given, link sent — and the confirmation matches the tool result.
SMS consent declined
The caller says no, and the agent respects it without lost composure or a repeated ask.
SMS failure fallback
The text fails and the agent says so truthfully, offering the manual route instead of pretending.
Existing-client document request
A service call the agent is not built to fulfil — routed to a real path rather than improvised.
Cancellation and urgent water loss
Sensitive calls handled with the required warnings and without promising coverage outcomes.
Privacy, upset callers and prompt injection
Third-party data requests, an angry premium increase, and attempts to extract secrets or override the rules — all covered by the suite.
Honest limits
What this case study does not claim
Pilot, not production claims
The audit states the agent is ready for a controlled quote-intake pilot only. It is not presented as an unsupervised full-service receptionist, and we repeat that here.
No live-call metrics
Testing used mocks and simulations; no real SMS was sent during QA. We do not claim call deflection, booked policies or savings from live traffic.
Client details kept private
Phone numbers, credentials, account IDs and customer information from the audit are not published. Only the configuration and test facts are shared.
FAQ
Questions about this audit
Why does an AI voice agent need an audit at all?
Because a smooth voice can hide a broken system. In this audit the agent denied an SMS capability it actually had, had no persistent state to keep it honest after a tool failure, and had a transfer enabled with nowhere safe to send callers. None of those are detectable from a casual demo call — they come from testing the prompt, tools and telephony against each other.
What makes an AI receptionist safe in a regulated industry?
Truthful actions, strict scope and clear failure behavior: only claim what tools confirm, never give licensed advice or promise outcomes, capture consent before sending anything, and give callers a real human route for anything the agent cannot do. The audit turns each of those into a tested scenario rather than a hope.
Can the agent transfer calls?
Not yet — and that is deliberate. No safe staffed destination was supplied, so transfer was left off rather than enabled as a broken promise. The moment a staffed destination is available, the transfer tool can be configured and added to the test suite.
Get your plan
Running a voice agent you do not fully trust?
Tell us what your phone agent is supposed to do. On a free strategy call we will map what needs auditing before it handles real customers.
- A free 15-minute AI strategy call — no obligation.
- A scoped blueprint: the riskiest behavior, and the first fix.
- Honest fit check — we will tell you if an audit is not your fastest win.
The fastest way in: answer six short questions and we will map your phone and message flow, find the biggest leak and show the fastest win — before we ever get on a call.
Free strategy call · No commitment · We reply by email
Keep going
Related pages
More client work
Relevant AI services
Industry pages
Ready to trust your phone agent with real customers?
Book a free AI strategy call and we will walk your voice setup against this audit.