Asenda Talk’s busy-hour handoffs. Consent and call records must hold.

A busy-hour test should show whether the agent keeps the caller’s intent, language choice, consent status, and next step intact when Twi and English appear in the same conversation. A polished demo can prove that a voice sounds good; it cannot prove that the full call path holds up when several real callers change pace, wording, and language without warning.

In 1999, NASA’s Mars Climate Orbiter approached Mars after a journey from Earth. The mission was lost when navigation data passed between systems using different measurement units. NASA’s investigation traced the failure to a mismatch between pound-seconds and newton-seconds, documented in its Mars Climate Orbiter mishap report.

The lesson applies to a busy voice-agent hour. Each part of a call can appear to work on its own: speech recognition, agent logic, voice synthesis, telephony events, billing, and call records. The risk appears at the handoffs. If one component loses the language switch, misunderstands an address, records consent incorrectly, or marks a completed call differently from the telephony provider, the call experience can fail even when the voice itself sounds convincing.

Test the switch, not two separate languages

A useful Twi and English test is not a Twi call followed by an English call. Ask callers to switch naturally in the middle of one task.

A customer might begin in Twi, use English for a product name or account reference, return to Twi to explain the problem, then answer a confirmation question in English. The test should reveal whether the agent carries the same request forward, or treats each language shift as a new conversation.

Listen for more than pronunciation. Check whether the agent:

• Preserves names, numbers, locations, and product terms when the language changes.
• Confirms uncertain details instead of confidently continuing with a wrong interpretation.
• Keeps the agreed task intact after a switch, such as booking a follow-up, collecting a preferred callback time, or escalating to a human.
• Uses language that matches the caller’s choice instead of forcing every response into one language.

For farm-input, support, or campaign calls, confirmation is often the difference between a useful interaction and a harmful one. Twi Speech Recognition: Why Farm-Input Agents Must Confirm Before Advising explores why retaining the caller’s original words matters before an agent acts on them.

Asenda Talk is built around native Twi speech recognition and synthesis, fine-tuned in-house. Teams evaluating it should still treat the busy-hour test as an evaluation exercise, not assume a capability from a short recording. Early access is the right time to find where a specific vocabulary, call flow, or code-switching pattern needs work.

Check the call trail beside the transcript

A transcript can look acceptable while the operational record tells a different story. During a busy-hour test, review the lifecycle events alongside selected recordings and transcripts.

For each call, can the team answer simple questions without guessing? Did the call connect? Did the agent receive the caller’s speech? Was consent captured? Did the caller opt out? Was the requested handoff created? Did the call end normally, or did the provider report a failure?

Asenda Talk includes a telephony lifecycle webhook pipeline with call-truth tracking, plus consent, opt-out, and audit records for every call. Put those records under pressure during the test. Compare a small sample of completed calls, dropped calls, opt-outs, and ambiguous interactions. The states should agree across the call record, the agent outcome, and the provider event.

This matters most when a campaign team is deciding who receives the next call batch. A fluent voice does not fix an unreliable event trail. What Happens When Consent, Opt-Out, and Call Records Disagree? covers the operational cost of letting those records diverge.

Measure recovery when the conversation gets messy

The first busy hour should include calls that a demo usually avoids: background noise, pauses, interrupted answers, repeated questions, mixed-language phrases, uncertain names, and callers who want to stop.

Evaluate recovery, not perfection. Does the agent ask a clear follow-up question? Does it repeat back the important detail? Does it offer a handoff when confidence is low? Does an opt-out end the intended contact path and leave an audit record?

The Mars Climate Orbiter did not fail because spaceflight lacked impressive individual components. It failed because a critical mismatch crossed a system boundary without being caught. Voice-agent teams need the same discipline: test the boundary between what was heard, what the agent understood, what action it took, and what the call system recorded.

Keep the commercial and provider decisions visible

Busy-hour testing also needs guardrails. Metered per-minute billing should remain behind an operator-controlled real-money gate until the team is ready to incur charges. Use a known call set, define who can approve the gate, and record which results count as a pass or a failure.

Asenda Talk uses Vapi for assistant runtime orchestration. Its outbound calling path remains gated behind an explicit telephony-provider decision that is not live yet. That constraint should be part of the evaluation plan. Teams can validate agent configuration, Twi and English conversational behavior, records, consent handling, and internal readiness now, while keeping any live outbound rollout contingent on the provider decision.

The first busy hour earns its value by exposing where the system needs another decision, confirmation, or safeguard. Run it with real call scenarios from Accra, Kumasi, or the locations your team serves, then turn every missed handoff into a concrete test case before the next session.