Voice agents quality.
Calls that fail and nobody knows why. Voice agents book visits, refill prescriptions and call patients.
A core focus: we deliver this as a service today.
Where voice agents breaks.
Voice agents book visits, refill prescriptions and call patients. When a call fails, the cause could be the microphone, the transcript, the prompt or the model, and logs rarely say which.
Mishears medical terms
A drug name or dose is transcribed wrongly and the agent acts on it.
Breaks on real callers
Accents, background noise and interruptions that never appeared in testing.
Writes the wrong thing back
The call went fine, but the appointment or refill entered in the chart doesn't match what was said.
Three steps, in the open.
Replay and component-swap attribution; fitted caller simulators; escalation and medication gates; call QA.
- 01
Simulate callers
Callers with accents, noise and interruptions run scripted scenarios many times.
- 02
Require every repeat to pass
A scenario passes only if all repeats succeed, since one lucky run proves little.
- 03
Find the faulty part
We replay failed calls with one part swapped (speech model, prompt, model) to see which fixes it.
Call QA and failure attribution.
Caller simulator, replay diagnosis.
Who builds it, who buys it.
Hippocratic ($3.5B), Assort ($1.2B Series C, Jun 2026), Hello Patient, Prosper, Third Way
The stakes
Horizontal voice-testing tools exist. The gap is clinical judgement: knowing which transcription error matters and which doesn't.
Rules and reading
Measure voice agents on your own data
What we measure for voice agents.
Workflow pass rate (all repeats pass)
Scenario success across repeated calls
Interruption recovery and naturalness
Composite
ASR error and its clinical impact
WER weighted by clinical meaning; LLM judge κ ≈0.71 and 0.50 vs two clinicians (clinician-clinician κ ≈0.51)
Turn latency, call completion, handoff-to-human rate
Operational
Clinical-safety label error rates
Nurse-labelled errors in labs, vitals, medications, human navigation, with adjudication
EHR write-back correctness
Scheduling, refill and intake entries match intent
Common questions.
What is pass^k?
The share of tasks that pass on every one of k repeated runs. It's stricter than average accuracy and closer to what callers experience.
Can you test Indian languages?
Yes, with native-speaking clinicians. See our language evaluation service.
Do you need call recordings?
Not necessarily. Simulated callers can create the audio, which avoids privacy issues.
Know your healthcare AI is safe to trust
before it reaches a patient.
Book a 30-minute call. We'll look at your product and tell you what we would measure first.
hello@precision.example