AI governance in Ghana must protect a person’s ability to access an automated service in the language they can use, including when they switch from English to Twi mid-call. Rules for consent, storage and audit trails matter, but they cannot repair a service that stops understanding at the sentence where the person explains what they need.
Consider an invented composite: Esi, a market trader in Kumasi, is standing behind her counter late on a Friday afternoon, holding a supplier’s handwritten reference number. She answers an automated support call in English. Halfway through, she switches to Twi to explain that the delivery location mentioned by the agent is wrong.
The agent keeps going.
It repeats the English address and asks Esi to confirm. The shop will close soon, the delivery could go to the wrong place, and she cannot tell whether saying “no” in Twi has registered at all. Her data may be encrypted. The call may be logged. Every internal policy could still appear satisfied while the service fails at the exact moment access depends on language.
The untranslated sentence is a governance event
A language switch can change the meaning of an entire call.
People do not always choose one language at the start and remain inside it. Esi may use English for the greeting, Twi for the problem, then English again for a reference number. She may switch because the agent used the wrong word, because the issue is easier to describe in Twi, or because the consequences have become serious.
If the system recognises the English sentences and loses the Twi correction, the failure reaches beyond transcription quality. The automated service has given Esi a narrower route to resolution than it gives someone whose explanation stays in English.
That is a question of access.
A governance review should therefore ask more than where audio is stored, who can retrieve it and how long it remains available. It should also ask which speech the service can understand, what happens when confidence drops, and whether a person can reach a human or another safe path before an incorrect action continues.
A clean record can conceal this kind of exclusion. The transcript may show a coherent English exchange while omitting the sentence that changed Esi’s request. What happens when a Twi request disappears from an English transcript? examines that evidence gap more closely.
Consent requires comprehension and control
Consent has little practical value when a caller cannot tell what the agent understood.
An automated call should identify itself, explain its purpose and provide a usable way to opt out. Those controls must survive code-switching. If Esi says in Twi that she does not agree, asks to stop, or corrects a destination, the system needs to preserve that moment as part of the call truth.
Asenda Talk is being built with consent, opt-out handling and an audit trail for every call. Its telephony lifecycle pipeline is designed to track what happened across the call rather than treating a generated transcript as the whole truth. Native Twi speech recognition and synthesis are fine-tuned in-house, instead of routing Twi through a generic third-party voice layer.
These are current product foundations, not proof that every accent, noisy connection or English-to-Twi transition has been solved. Asenda Talk remains in active early access. Outbound calling is also held behind an operator-controlled real-money gate while the live telephony-provider decision remains open.
That distinction matters. Responsible governance needs evidence from evaluated behaviour, not confidence borrowed from a feature label.
Test the switch, then test the consequence
A bilingual demo can sound convincing for several minutes and still fail on one decisive sentence. Testing should place the language switch where it can alter an outcome.
Give the agent an English introduction, then put the correction, refusal or urgent detail in Twi. Add ordinary call conditions: background conversation, a repeated number, a pause while someone finds a receipt. Check what the system heard, what action it took, whether an opt-out stopped the workflow, and whether the audit trail preserved the disagreement.
Then test recovery. When recognition confidence falls, does the agent ask the person to repeat the sentence? Does it confirm the corrected detail in the language used? Can it pause an action that may cause harm? Is there another route when automation cannot complete the job safely?
One polished transcript cannot answer those questions. A previous Asenda Talk test illustrates why one code-switch can invalidate the demo.
Govern the doorway to the service
Return to Esi at the counter. The safer version of the call does not pretend certainty. After her Twi correction, the agent pauses, repeats the destination back in Twi and asks for confirmation. If it cannot resolve the mismatch, it stops the delivery change and records why the call needs review.
Her sentence now changes the outcome.
That is the standard Ghana’s AI debate should carry into procurement, testing and regulation. Ask whose speech reaches the decision system, which language transitions have been evaluated, and what happens when understanding fails. Before approving an automated service, put the first untranslated sentence near a real consequence and watch whether the person still has control.
Comments
No comments yet.