The Twi Withdrawal the Transcript Could Miss, and the Calls It Could Allow

Consent can change during a call, and a compliant voice agent must act on the caller’s latest clear intent in whichever supported language they use. An English “I agree” cannot override a later Twi withdrawal such as “Mennpɛ bio.”

At 4:40 p.m. in Kumasi, Abena is standing behind the counter of her small provisions shop, holding a supplier receipt against the breeze from the doorway. She answers an automated call and agrees in English to hear about a service. Thirty seconds later, a customer enters and points to an item on the top shelf.

Abena wants the call to stop. “Mennpɛ bio,” she says.

In this invented scenario, the agent has a decision to make. If it treats the opening “I agree” as permanent, it may continue the pitch, schedule another call, or preserve Abena as contactable. Her withdrawal disappears at the exact moment it matters. The bad ending is concrete: Abena receives more calls after she has told the system, in Twi, that she does not want them.

Consent is a changing state

A consent record often looks deceptively simple: yes or no, accepted or declined, timestamped once. A real conversation moves differently.

Someone may agree to continue, then change their mind after hearing the purpose of the call. They may consent to one action but reject another. They may begin in English and switch to Twi when they want to speak plainly, quickly, or with less room for misunderstanding.

The system therefore needs to track intent across the full call. Abena’s state changes when she says “Mennpɛ bio.” Every later action should use that updated state.

This distinction also matters when reviewing an incident. A record showing that consent existed at the start does not answer whether it still existed when the agent continued speaking or initiated a follow-up action. The useful question is: what was the caller’s latest clear instruction before the system acted?

That is why a dial list cannot prove permission by itself, as explored in Automated Twi Call Consent: Why a Dial List Cannot Prove Permission. Permission must survive contact with the actual conversation.

The transcript can preserve the wrong truth

Suppose Abena’s English opening is transcribed correctly, while her Twi withdrawal is omitted, mistranscribed, or reduced to an uncertain fragment. The stored transcript may look clean. It may also tell the opposite story from the call.

That creates two failures.

First, the runtime may miss the opt-out and continue. Second, the audit record may later support the wrong conclusion because the most important utterance never reached the consent state.

A voice agent serving Ghanaian callers must treat language recognition as part of the compliance path. Native Twi speech recognition matters here because the system needs to hear the words that change what it is allowed to do. Asenda Talk’s Twi recognition and synthesis are fine-tuned in-house rather than passed through a generic third-party speech wrapper. The relevant evaluation is still practical: does the agent detect the withdrawal under real accents, code-switching, background noise, and ordinary call conditions?

A polished transcript cannot compensate for a missed opt-out. The same problem appears in What If a Clean Transcript Hides an Earlier Opt-Out?.

The opt-out must change system behavior

Catching the phrase is only the first step. The withdrawal needs to move through the system as an event with consequences.

In an Asenda Talk agent configuration, consent and opt-out handling can be attached to the call record and audit trail. The telephony lifecycle pipeline is designed to preserve call truth across events, including what happened and when. That record should make the sequence visible:

Abena agreed to continue. Abena later withdrew in Twi. The agent stopped the consent-dependent action. Future contact eligibility changed.

The operator also needs a clear boundary between testing and real calls. Asenda Talk is in active early access. The assistant runtime is orchestrated through Vapi, while live outbound calling remains behind an operator-controlled real-money gate and an explicit telephony-provider decision that has not yet been made live. No outbound deployment should be described as operational until that decision, the provider path, and the applicable testing gates are complete.

That constraint is useful. It creates space to evaluate the hard case before money is spent on calls: a caller consents in one language, withdraws in another, and expects the second answer to control what happens next.

Test the reversal, not only the opening

Near the end of the evaluation call, Abena repeats herself. This time the agent recognizes the Twi withdrawal, updates the consent state, ends the relevant flow, and records the change for review.

The important result is not that the conversation sounded natural. It is that the system changed course before another consent-dependent action occurred.

A practical test set should include reversals in both directions, code-switched sentences, interruptions, repeated opt-outs, and phrases delivered after the agent has already started its next turn. Review the audio, transcript, state transition, call event, and suppression behavior together. If one layer says “opted out” while another still permits contact, the test has failed.

Back at the shop counter, Abena lowers the phone and serves the waiting customer. Her first answer remains in the record. Her last clear instruction governs what happens next.