Responsible deployment

What should an AI phone agent do when it does not know?

A good system does not hide uncertainty: it detects it, reduces scope, and gives control to a person.

Canadian French + English • Configurable human handoff • Keep your number when supported by your carrier

At a glance

Defined by the workflow, not only the voice

Voice agents can mishear a name, use stale knowledge, encounter an integration outage, or receive an out-of-scope request. A natural voice does not remove those risks.

Reliability comes from assuming failure: thresholds, validation, permissions, fallback messages, handoff, and human review.

Outcomes

Failure modes to test

Misunderstanding

Noise, accent, proper names, correction, overlapping speech, or language change.

Insufficient knowledge

Missing question, stale policy, or conflict between sources.

Impossible action

Calendar downtime, denied permission, duplicate, or incomplete data.

Higher-risk request

Emergency, professional advice, complaint, threat, or sensitive data.

Implementation

Design safe recovery

  1. 1

    Recognize

    Detect uncertainty, an out-of-scope request, or technical failure.

  2. 2

    Clarify once

    Ask a simple question without forcing endless repetition.

  3. 3

    Reduce scope

    Move from action to message capture when confidence is insufficient.

  4. 4

    Transfer and record

    Give context to a person and retain the event for correction.

Scenarios

When to require a person

Sensitive decision

Advice, exception, refund, contractual commitment, or professional judgement.

Uncertain identity

No account disclosure or sensitive action without suitable authentication.

Emotion or danger

Intense anger, threat, reported danger, or emergency-services request.

FAQ

Limitation questions

Can AI hallucinate on a phone call?

A model can generate an incorrect answer. Limited scope, approved sources, action rules, and handoff reduce but do not eliminate that risk.

How many times should it ask a caller to repeat?

The workflow should limit attempts and then offer another path, such as transfer or structured message capture.

Must every call be reviewed?

Not necessarily, but new scenarios, higher-risk calls, failures, complaints, and quality samples should have a defined review process.

What does a person receive during handoff?

When supported, they should receive the reason, captured information, and handoff cause so the caller does not start over.

Request a demo that includes failure

We will test an unknown question, integration outage, and transfer request—not only the ideal path.