The hard part of a voice agent is knowing when the call should stop being automated_
Zurich Airport is procuring a multilingual AI voice agent that must automate routine enquiries while handing complex, sensitive and security-relevant calls to staff. That requirement exposes the production boundary that matters: handoff is part of the architecture, not a fallback prompt.
- Voice AI
- Contact Centers
- Systems Integration
- Workflow Automation
- Human Handoff
Voice agents are usually demonstrated on calls that stay inside the automation boundary. Production contact centres are defined by the calls that leave it.
Zurich Airport is currently procuring an AI voice-agent solution for its Customer Contact Center. The requirement is deliberately split: publicly available routine and standard enquiries should be handled around the clock and in multiple languages, while complex, sensitive or security-relevant matters must be forwarded to staff. Questions close on 21 September and submissions on 28 September 2026.
That is a more useful specification than a target such as “automate 80% of calls.” It defines classes of work the system may own and classes it must not.
Handoff should be a first-class state_
A production call flow should not treat human transfer as the final catch block. The conversation state needs an explicit handoff decision with a reason, destination and payload.
Useful triggers include policy boundaries, security-sensitive intent, repeated misunderstanding, authentication failure, unavailable authoritative data, customer request and operational exceptions that the automated workflow is not permitted to resolve.
Those triggers should be deterministic where possible. A language model can help classify an utterance, but the permission to continue, authenticate, disclose information or execute a protected action belongs to the workflow policy.
Transfer the context, not just the audio_
A cold transfer makes the customer repeat the problem and forces the agent to reconstruct what the automated system already learned. A useful handoff package can include verified identity state, detected language, concise call summary, intent, information already collected, actions attempted, failure reason and the next permitted step.
The destination matters too. Billing, lost property, security, accessibility and general information may require different queues, operating hours and escalation rules. Routing should therefore be based on the operational owner of the unresolved work, not simply on the last phrase spoken.
Measure why automation gives work back_
Containment rate is useful, but it is not enough. Track handoff rate by intent, handoff reason, repeat-contact rate, time from transfer to resolution, authentication failures, abandoned transfers and cases where the human agent had to recollect information the voice system already possessed.
Those traces tell the team where autonomy can safely expand. A high volume of transfers caused by one missing backend lookup is an integration problem. Repeated transfers after misunderstood names may be a speech-recognition problem. Security-sensitive calls being transferred exactly as policy requires are not automation failures.
This boundary is familiar in Parallaxis voice work. Our published PBX interview integration preserves the existing telephony path while adding speech recognition, controlled dialogue, text-to-speech, call events and structured results. Our bilingual loan-qualification work similarly treats captured intent as input to a wider sales workflow rather than pretending the conversation model owns the business process.
A voice agent is ready for production when the team can explain not only what it can answer, but when it must stop, where the call goes next and what context survives the transfer.