A polished English greeting proves that a voice agent can start a call. It does not prove that the agent can understand a billing problem explained in Twi, preserve the account context, or help the caller reach a resolution.
At 8:03am, the first caller hears a clear welcome and begins explaining an unexpected charge. She starts in English, then switches to Twi when the details become difficult to describe. The voice still sounds natural, but the responses stop matching the problem. After one repeated question, she hangs up.
The dashboard may show a connected call. The greeting may pass a demo. The customer still received no help.
The interface can work while the system fails
On September 23, 1999, NASA’s Jet Propulsion Laboratory in Pasadena expected the Mars Climate Orbiter to enter orbit around Mars. The spacecraft passed behind the planet, and the team waited for its signal to return.
It never did.
The spacecraft had travelled from Earth and operated for months. The failure came from a mismatch buried inside the navigation process. Software supplied by Lockheed Martin produced thruster data in pound-seconds, while the navigation system at JPL expected newton-seconds. The values moved through the system, but they did not carry the meaning the receiving side assumed.
Arthur Stephenson chaired NASA’s Mars Climate Orbiter Mishap Investigation Board. Its Phase I Report documented the unit mismatch and the failures in verification and communication that allowed it to persist.
That mechanism matters for voice AI. A call can connect, audio can travel, and a fluent greeting can play while the language layer fails to carry the caller’s actual meaning. The surface looks operational. The substance has already broken.
A greeting tests presentation, not understanding
Voice-agent demos often begin with the easiest part of the interaction: a scripted opening. The agent says the company name correctly, pauses in the right place, and asks a familiar question. That can confirm synthesis quality and basic call flow.
A billing conversation places different demands on the system. The caller may switch languages mid-sentence, pronounce an account detail in English, describe the dispute in Twi, and expect the agent to remember what came before. Each turn depends on recognition, context and response generation working together.
Teams should therefore test full tasks, rather than isolated phrases. Give the agent a billing dispute with several turns. Introduce code-switching. Correct one detail halfway through. Ask the same question in different words. Then inspect whether the final response reflects the complete problem.
An account match alone cannot establish comprehension. Bilingual voice-agent testing should check what the agent understood, what it missed, and what action followed. When the same detail is misheard twice, the test should examine recognition output, conversation state and recovery behavior, as described in this practical testing guide.
Call records need to reflect customer outcomes
A connected status answers a narrow question: did the telephony system establish a call? It cannot tell you whether the caller was understood, whether consent remained valid, or whether the issue was resolved.
A useful review record should connect the telephony lifecycle to the conversation itself. It should show when the call connected, what language appeared, where recognition confidence fell, whether the caller asked to stop, how the agent responded, and why the call ended. A hang-up after repeated misunderstanding belongs in a different category from a completed support interaction.
Asenda Talk is being built around that distinction. The platform currently supports configurable voice agents, native Twi speech recognition and synthesis fine-tuned in-house, Vapi-orchestrated assistant runtime, call lifecycle webhooks, call-truth tracking, consent records, opt-out handling and audit trails. It is in active early access. More African languages remain in progress, and live outbound calling is still gated behind an explicit telephony-provider decision.
Those limits should remain visible during evaluation. A working component deserves credit for what it proves. It should not become evidence for capabilities that the test never exercised.
Test the problem the caller came to solve
Before opening traffic to real callers, define a small set of complete bilingual tasks. Use realistic account references without exposing customer data. Include clean audio, interruptions, corrections and natural switches between Twi and English. Review transcripts beside recordings and lifecycle events.
Most importantly, score the outcome. Did the agent identify the billing issue? Did it retain the corrected detail? Did it ask a useful follow-up? Did it record an opt-out when requested? Did the final call status describe what happened?
The Mars Climate Orbiter reached the point where one system’s output had to mean the same thing to another system. That assumption failed, and months of apparently successful operation could not compensate for it.
The equivalent voice-agent check happens after the greeting. At 8:03am, when the caller leaves the script and explains the real problem in Twi, the system has to carry her meaning all the way through.
Comments
No comments yet.