For AI voice agents and AI receptionists
Test the failure paths after every change.
CanaryTel places scored test calls against your agent after prompt, model, and tool changes, and gives your customers third-party evidence that it handles the hard calls.
Request the free auditWhy voice agents need an outside check
Silent regressions
A prompt edit, a model swap, or a tool change can break a path nobody re-tested. The demo call still sounds fine while the emergency escalation no longer fires.
Failure paths go untested
Happy-path evals are common. Caller hangs up mid-script, on-call does not answer, wrong number, urgent versus routine: these are the paths that hurt customers, and the ones least often exercised.
Customers want evidence
Your customers are asking how they would know the agent follows their rules. A vendor’s own dashboard is the vendor grading itself.
How Canary Calls works for an AI vendor
Real phone calls into your agent, scored against the customer's script.
Staging and production
Run the scenario library against a staging number before a release, and against production after it. Same scenarios, same scoring, so results compare across versions.
Failure paths first
Scenarios include emergency escalation, interruptions, caller hangs up mid-script, after-hours rollover, on-call not answering, and overflow, each scored step by step.
Third-party evidence for your customers
Scored results with audio, transcripts, and timestamps, produced by a party that is independent of you and of your customer. Share them with your customers’ consent.
Fictional callers
Test calls use fictional callers and fictional patient details only.
Canary Watch via webhook, in early access
Live scoring of real calls is most mature for AI voice agents.
- Webhook ingest of call records and transcripts you send us.
- Scoring for handling time, required disclosures, and script adherence against a rubric agreed per campaign.
- Alerts on SLO burn and anomalies at the cadence of the feed.
- Deploy and version events let us detect regressions after a prompt or model change. This works only for agents that send those events.
- Where calls involve patient health information, any plan includes a BAA.
Questions AI vendors ask
How does Canary Watch connect to our agent?
By webhook. It is the most mature of the early-access integrations: we ingest call records and transcripts you send, and score them against a rubric agreed per campaign. We scope the payload and retention with you first, and sampling with human review of failures is included.
How do you detect regressions after a prompt or model change?
If your platform sends us version or deploy events, we line up score changes against those events, so a drop after a release is visible as such. Without those events we can still run scheduled Canary Calls and compare results over time, but we cannot attribute a change to a specific deploy.
Can you test our staging environment?
Yes, if it is reachable by phone. We call a staging number the same way we call a production one, and compare the results.
Are you a competitor to our own evals?
No. Your internal evals test what you know to test. CanaryTel is an independent outside-in check, and its evidence is something your customers can rely on because you do not control it. Results are never edited at anyone’s request.
Free one-week audit on one campaign
Seven days of Canary Calls against one inbound campaign, happy and failure paths, and a scored audit report with evidence. No integration, no cost, no commitment.
Request the free audit