The question that makes an SME applicant switch to Twi is the one where a wrong answer has consequences: affordability, ownership, repayment responsibility, or consent. Language choice becomes a test of whether the agent can understand the applicant’s actual situation and confirm it accurately.
In April 1970, Apollo 13’s crew faced a problem that could not be solved by repeating the same instruction more clearly. The spacecraft’s oxygen-tank explosion had forced Jim Lovell, Fred Haise, and Jack Swigert into the lunar module. Carbon dioxide was rising, but the command module used square filters and the lunar module used round openings. NASA engineers had to create an adapter from materials already on board. The crew’s safe return depended on making the message, materials, and constraints fit together.
That is the shape of a difficult onboarding call. An applicant may understand an English prompt perfectly well until the call reaches the question that requires a precise account of their business.
The question arrives after the easy answers
Picture a morning application call with a small business owner in Kumasi. The early questions are routine: business name, location, service offered, contact details. The applicant answers in English. The agent moves on.
Then it asks how the business would meet its repayment commitment during a slow period.
There is a pause. The applicant begins in English, stops, and switches to Twi.
That switch should not be treated as a preference setting being changed halfway through a call. It is a signal. The applicant has reached the part where vocabulary, family roles, seasonal income, informal records, and the difference between “my business” and “our household income” may matter.
A voice agent that continues as though nothing changed can collect words while missing meaning. A stronger flow recognizes the language change, continues in Twi where the applicant can explain, and confirms the important answer in language the applicant understands before recording it.
Understanding must show up in the confirmation
The first answer is rarely the part that needs the most care. Confirmation is.
If an applicant explains a seasonal cash-flow gap in Twi, the agent needs to reflect the meaning back without flattening it into a vague English label such as “irregular income.” The record should capture the answer required for the application, while the conversation gives the applicant room to correct it.
This is especially important when the next step depends on the answer. A question about who can approve a business commitment, for example, should lead to a clear confirmation of the person’s role and consent. A question about a slow trading period should lead to confirmation of the stated plan, not an inference that the business is unable to pay.
The lesson from Apollo 13 is practical: a system can contain the right components and still fail at the connection point. For an onboarding agent, that connection point is often the confirmation after a language switch.
Design the language switch as a controlled handoff
An early-access voice platform should make this behavior testable, rather than promising that every bilingual call will work perfectly on day one.
In Asenda Talk, teams can configure an agent’s persona, first message, and voice, then evaluate conversations using native Twi speech recognition and synthesis developed in-house. The useful test is not simply whether the agent recognizes Twi. Test whether it preserves the applicant’s intended meaning when the call moves between English and Twi, especially at a high-stakes question.
Build the call around a few explicit controls:
- Ask the sensitive business question in plain language.
- Detect or allow a language switch without forcing the caller to restart.
- Confirm the answer before using it in a decision or follow-up.
- Preserve consent, opt-out status, and an audit trail for the call.
- Route uncertain answers to a human review path.
This is closely related to the problem in Bilingual Voice AI Testing: Why Confirmation Must Survive an English to Twi Switch. Recognition alone does not establish that the record is correct.
Keep the operational boundary visible
Asenda Talk currently provides agent configuration, Twi speech capabilities, call-truth tracking through telephony lifecycle webhooks, consent and audit controls, and metered billing with an operator-controlled real-money gate. Its assistant runtime is orchestrated through Vapi.
Outbound calling remains gated behind an explicit telephony-provider decision that has not been made live. Teams can use that boundary to evaluate scripts, confirmations, and language-switch handling before treating an outbound workflow as operational.
Apollo 13’s engineers did not solve a general problem called “spaceflight.” They solved the exact incompatibility in front of them. SME onboarding teams should apply the same discipline: identify the question where misunderstanding creates risk, test the Twi switch there, and require a confirmation the applicant can correct.
Comments
No comments yet.