A caller switches to Twi. Their real question goes unanswered.

A clean English response after a caller switches to Twi leaves the question unresolved. Support teams need to evaluate whether the agent understood the caller’s intent in the language the caller chose, then confirm the next step was captured.

In 1999, NASA lost contact with the Mars Climate Orbiter as it approached Mars. The investigation found that one part of the navigation work used pound-seconds while another expected newton-seconds. The spacecraft had received data that looked usable until the units carried a different meaning.

A support lead hears a similar mismatch on Monday morning. The call recording is clear. The agent’s English voice is calm. The caller had switched to Twi to explain the actual problem, then received an English answer that addressed a nearby issue. There is no dramatic error sound. No obvious system failure. The caller simply leaves without the help they called for.

A fluent voice can still miss the request

The first unanswered Twi question is often easy to overlook because the agent keeps talking. It may recognize that the caller is speaking, select an English response, and follow the script smoothly. That is a conversation-shaped failure, not a resolved conversation.

For a Ghanaian support desk, the useful question is more demanding: did the system recognize the Twi words, retain the meaning when the caller mixed in English, and route the caller toward the intended outcome?

That means reviewing calls for more than silence, duration, or whether the assistant completed its prompt. Check the caller’s original request against the agent’s response and recorded next step. If someone asks to be called back, change an account detail, explain a payment issue, or stop future calls, the record should show that intent clearly. What Happens When Twi-English Calls Lose the Customer’s Intended Next Step? explores why this is where a call can fail even when both parties stayed on the line.

The call record has to show what happened

NASA’s Mars Climate Orbiter did not fail because its teams lacked data. The problem was that the data was interpreted across incompatible assumptions. The Mars Climate Orbiter Mishap Investigation Board documented the unit mismatch and its role in the loss.

Voice support has its own version of that risk. A transcript can say one thing, an agent response can imply another, and a billing record can show a completed minute. None of those alone proves the caller got help.

Asenda Talk is being built with a telephony lifecycle webhook pipeline and call-truth tracking so operators can inspect what occurred across the call lifecycle. It also records consent, opt-out requests, and an audit trail for each call. Those controls matter most when the call contains a meaningful instruction, especially one expressed after a language switch.

The platform is in early access. Native Twi speech recognition and synthesis are built and fine-tuned in-house, while more African languages remain in progress. The assistant runtime is Vapi-orchestrated. Outbound calling is still gated behind an explicit telephony-provider decision and is not live as a general capability today.

Test the moment the script stops helping

A support lead should create a small Twi-English test set from real, approved support scenarios before relying on an agent for customer calls. Include the questions that create downstream work when misunderstood: follow-up requests, account changes, payment questions, complaints, and opt-outs.

Listen for the turn where the caller stops following the first-message language. Does the agent recognize the switch? Does it answer the actual question? Does it preserve the requested next step in the call record? Can an operator see enough context to correct a mistake?

This work should happen before a team treats a successful English demo as evidence of Twi call quality. Asenda Talk’s Twi accuracy test. A missed instruction can break the call. sets out the practical problem: one missed instruction can change the outcome of an otherwise polished call.

Make unresolved language switches visible

The useful Monday report is short: calls where the caller switched to Twi, the question detected, the response language, the recorded intent, and the final disposition. A support lead can then review the calls where those fields disagree.

This is also where consent and opt-out handling must remain explicit. If a caller says they do not want future calls, that instruction deserves the same care as a payment or support request. A fluent response that fails to record the opt-out creates a different kind of unresolved call.

Mars Climate Orbiter is a reminder that a system can keep receiving signals while losing the meaning that makes those signals safe to use. For Twi voice agents, the test is equally concrete: when a caller changes language, verify the question, the answer, and the next step before calling the interaction complete.