A bilingual campaign can fail even when every Twi sentence is understood correctly. If consent context disappears during a transfer, the next agent may continue a call the person already limited, redirected, or asked to stop.
At 4:46 p.m. in Accra, Kojo was holding a cold cup of tea and listening to the final campaign test before sign-off. The composite campaign manager had configured an English and Twi voice agent for a customer follow-up campaign, and the conversation sounded natural enough to make the room relax.
Then the caller switched to Twi, agreed to discuss one service, and refused questions about another. The voice agent acknowledged the limit. Seconds later, the call moved to the next stage of the workflow, where the consent restriction was missing.
The receiving agent resumed the excluded questions.
Kojo stopped the test. If that handoff reached a real call, the campaign could override a caller’s stated boundary and leave the review team without a reliable account of why the conversation continued. The Twi worked. The transfer did not carry the meaning that mattered.
Consent must travel with the conversation
A transcript records words. A safe handoff needs more: what the person permitted, what they declined, whether they opted out, and which instruction remained active when control changed.
In Kojo’s test, the caller had not withdrawn all consent. She had narrowed it. That distinction matters because a transfer cannot safely reduce a specific boundary to a generic status such as “consent captured.”
The next agent needs structured context that answers practical questions:
- What did the caller agree to discuss?
- What did the caller refuse?
- Which language carried the instruction?
- Did the agent acknowledge it?
- What action should the receiving agent take?
Without those answers, a technically correct transcript can still produce the wrong action. The risk grows during bilingual calls because a language change may also mark a change in comfort, emphasis, or intent. A caller might begin in English, set a firm boundary in Twi, then return to English after the agent acknowledges it.
That sequence belongs in the audit trail. [Bilingual call review](\/blog\/bilingual-call-review-why-language-transitions-belong-in-the-audit-trail-fcd3031d\/) depends on preserving the transition alongside the instruction it carried.
A successful transfer needs call truth
Campaign teams often test whether audio survives a handoff and whether the next agent starts speaking. Those checks reveal connection failures. They do not prove that the receiving stage inherited the caller’s choices.
Kojo rewound the test and inspected the lifecycle events. The transfer event showed that control had moved, but the receiving context lacked the refusal attached to the earlier Twi turn. The campaign had one record of what was said and another record of what the next agent believed it could do.
That split created the dangerous moment. If the team approved the workflow based on conversational quality alone, they could launch a bilingual experience that sounded capable while behaving carelessly.
Call-truth tracking should connect the conversation, transfer, consent state, opt-out state, and resulting action. When those records disagree, the campaign should pause for review. A completed handoff is not evidence of a valid handoff.
This is also why language evaluation must extend beyond transcription accuracy. [A correct bilingual transcript can still trigger the wrong action](\/blog\/bilingual-voice-agents-why-a-correct-transcript-can-still-trigger-the-wrong-action-a0b44be7\/) when downstream logic receives an incomplete version of the caller’s intent.
Test the boundary, then test the handoff
Kojo changed the acceptance test. The team would no longer ask only whether the agent recognized Twi and produced an appropriate reply. Each scenario would follow the instruction through every stage that could act on it.
They tested a full opt-out, consent limited to one topic, a mid-call language switch, and a transfer after the caller corrected the agent. For every case, the reviewers checked the original utterance, the agent’s acknowledgement, the stored consent state, the transfer payload, and the receiving agent’s first action.
The campaign remained paused while the failed path was corrected. That pause mattered. Asenda Talk includes consent, opt-out, audit trails, telephony lifecycle webhooks, and call-truth tracking, but teams still need to configure and evaluate how campaign logic uses those records.
The platform is in active early access. Native Twi speech recognition and synthesis are built and fine-tuned in-house, while more African languages remain in progress. Vapi orchestrates the assistant runtime. Outbound calling also remains behind an explicit telephony-provider decision and an operator-controlled real-money gate, so a configured campaign should not be treated as approved for live dialing.
The next review starts after the transfer
The following morning, Kojo returned to the same test call. This time, the receiving agent inherited the topic restriction and stayed within it. The audit record showed the Twi instruction, its acknowledgement, the handoff, and the action taken afterward.
He did not mark the campaign ready because the voice sounded natural. He marked one scenario as passed because the caller’s boundary survived the point where the earlier version had lost it.
That is the review standard worth carrying into every bilingual campaign: follow consent until the final system that can act, then confirm the action matches what the caller allowed.
Comments
No comments yet.