An English-only entry flow can make Twi-speaking demand disappear before it reaches a support queue. If the system records only routed calls, the business may read a language-access failure as a lack of demand.
During the Second World War, statistician Abraham Wald studied damage on aircraft returning from combat. The visible bullet holes seemed to show where armor should be added. Wald identified the missing evidence: aircraft hit in other areas had not returned to be counted.
His analysis, preserved in a 1943 Statistical Research Group memorandum and discussed by Marc Mangel and Francisco Samaniego in the Journal of the American Statistical Association, became a defining example of survivorship bias. The sample contained survivors. Treating it as a picture of every aircraft would produce the wrong decision.
A support queue can hide the same measurement problem.
What disappears before the queue
Consider a hypothetical customer calling a Ghanaian business at 8:07 AM. She begins in Twi. The automated entry flow responds in English, presents options she does not follow, and waits for an input that never arrives.
The call ends before routing.
No support agent sees a missed interaction. No ticket records her billing question. A dashboard built around queue entries shows no Twi request because the customer never crossed the point where language demand becomes visible.
The business may record the attempt as abandonment, invalid input, a short call, or no routed outcome. Each label describes part of the system event. None captures the customer’s intent.
This is the support equivalent of examining only the aircraft that returned. The queue contains people who understood enough of the entry flow to reach it. The callers excluded by that flow remain outside the sample used to make language decisions.
That creates a circular conclusion: the business sees little measured Twi demand, keeps the English-only entry flow, and continues filtering out the calls that would challenge the conclusion.
Measure the path before routing
Queue analytics answer useful questions. How long did callers wait? Which issue types reached agents? How many routed calls were resolved?
They cannot explain what happened before routing unless the telephony system records that stage deliberately.
A better measurement model follows the full call lifecycle:
- Record when the call connected and when the entry flow began.
- Capture the language used in the caller’s first response.
- Preserve recognition confidence and fallback events.
- Distinguish caller hang-ups from platform failures and system-ended calls.
- Track whether the call reached an agent, automation, transfer, or consent checkpoint.
- Keep the final call status tied to the provider event that confirms what happened.
This is call-truth tracking: measuring the call as it moved through the system, rather than treating a queue entry as proof that the customer journey began there.
Language detection also needs care. A failed recognition event does not prove silence or irrelevant speech. It may show that the model could not understand the language, accent, code-switching, or audio conditions presented to it. Repeated fallback prompts can turn that model limitation into a customer exit.
The same distinction matters after routing. As discussed in Bilingual Voice Agent Testing: Why an Account Match Cannot Prove Understanding, a successful transaction step does not establish that the caller understood the exchange. Measurement must cover both system completion and human comprehension.
Build language access into the first turn
A Twi-capable support experience cannot begin after an English-only gate. The first message, consent language, recognition model, fallback prompts, and routing rules all affect who becomes visible in the data.
Asenda Talk is being built around native Twi speech recognition and synthesis fine-tuned in-house, with English support and more African languages in progress. Teams can configure an agent’s persona, first message, and voice, while the calling runtime is orchestrated through Vapi.
The platform also includes consent, opt-out, and audit records for each call, plus a telephony lifecycle webhook pipeline designed to preserve call outcomes. These controls matter because recognizing Twi is only one part of the job. A business also needs to know whether the caller consented, where the interaction stopped, and which event supports the reported result.
Asenda Talk remains in active early access. It has not reached feature parity with established platforms such as Vapi, Retell AI, or Bland AI. Outbound calling is also held behind an operator-controlled real-money gate while the live telephony-provider decision remains unresolved. Those boundaries should shape pilot scope.
Test the callers your dashboard excludes
Start with first-turn tests in Twi, English, and natural code-switching. Include unclear audio, pauses, corrections, repeated questions, opt-out requests, and callers who hang up during the entry flow.
Then compare three counts: calls that connected, calls where speech was detected, and calls that reached the intended route. Investigate the gaps by language and failure stage. Do not collapse them into one abandonment rate.
Review the actual sequence for selected failures: first prompt, caller response, recognition result, fallback, routing decision, and final provider-confirmed status. Where consent wording changes by language, check meaning as well as pronunciation. The English Disclosure Ama Had, and Why It Failed in Twi explains why a stored disclosure can still fail the caller.
Wald’s lesson was about the evidence missing from the sample. For a support desk, the missing evidence may be a Twi-speaking caller who never became a queue statistic. Instrument the first turn, preserve the complete lifecycle, and treat pre-routing exits as customer demand that requires investigation.
Comments
No comments yet.