Absa’s reported AI push shows the scale customer-service automation can reach: its chatbot fields 100,000 queries monthly. But English-centred AI could leave a familiar service gap untouched for customers who explain financial problems most clearly in Twi.
Consider an illustrative scene. At 8:14 on a humid morning in Kumasi, Adwoa stands beside a provision shop with her phone pressed to one ear and a handwritten transaction reference in her other hand. She needs to explain a debit she does not recognise before the money is needed that afternoon.
The automated service understands her opening sentence in English. Then she switches to Twi to describe what happened, who sent the money and why the timing matters. The exchange begins to fracture. She repeats herself, translates part of the story back into English, then starts again after the system follows the wrong detail.
If she cannot make the problem understood, the payment she planned to make that afternoon may fail. For one tense minute, there is no clear path forward.
High query volume can hide a language gap
Absa has said that 1,400 developers use AI coding tools and that its chatbot handles 100,000 queries each month. Those figures show substantial AI activity. They do not tell us which languages the chatbot understands, how well it handles code-switching or whether a Twi-speaking customer can complete a sensitive support conversation without moving back to English.
That distinction matters beyond Absa.
Large banks often set expectations for customer service across a market. When a bank makes AI support familiar, customers begin expecting insurers, clinics, delivery companies, campaigns and smaller financial providers to offer similar access. If the model behind that support remains strongest in English, businesses may copy the automation while copying its language boundary too.
A dashboard can still look healthy. Calls connect. Conversations last several minutes. The system produces transcripts. Yet a customer may spend that entire interaction repairing misunderstandings.
For Adwoa, the failure is not silence. It is a system that appears to understand until the conversation reaches the detail that matters.
Twi support requires more than translation
A voice agent can recognise English reliably and still struggle when a caller moves between Twi and English inside one thought. Names, locations, transaction references and borrowed financial terms make the task harder. Literal translation may preserve words while losing intent.
This is why the useful question is deeper than “Does the agent support Twi?” Businesses need to evaluate what happens when a customer corrects a detail, changes language halfway through a sentence, speaks quickly under pressure or describes a problem without using the company’s preferred terminology.
The difference between hearing words and understanding intent is explored further in Does your AI agent truly understand Twi, or just hear the words?. It is especially important in banking, where confusing “money was sent” with “money was deducted” can send a conversation down the wrong path.
Native speech processing offers a more direct route. Asenda Talk’s Twi speech recognition and synthesis are fine-tuned in-house rather than passed through a general third-party voice layer. The platform is in active early access, with more African languages in progress. It is still reaching feature parity with established voice-agent platforms.
That caveat belongs beside the capability. Language coverage should be tested in real conversations, not inferred from a language label.
The real test happens at the point of consequence
A useful pilot should begin with moments like Adwoa’s, not polished demo scripts.
Build scenarios around disputed debits, missed deliveries, appointment changes and consent withdrawal. Include speakers who code-switch naturally. Test background noise, corrections, interruptions and the phrases people use when they are worried. Then review the call outcome alongside the transcript.
Call-truth tracking matters here. A completed call does not prove that the caller completed the task. Teams need lifecycle records showing whether the call connected, what state it reached, whether consent was recorded, whether an opt-out occurred and how the interaction ended.
Asenda Talk has a telephony lifecycle webhook pipeline, consent and opt-out records, per-call audit trails and metered per-minute billing behind an operator-controlled real-money gate. Voice agents can be configured with a persona, first message and voice, while Vapi orchestrates the assistant runtime.
Outbound calling is not live by default. It remains gated behind an explicit telephony-provider decision. That limitation should be part of any evaluation plan, especially for teams comparing the platform with Vapi, Retell AI or Bland AI.
Build language access before automation becomes the default
The opportunity for Ghanaian businesses is not to imitate the largest AI deployment. It is to decide which customers their automation must understand before call volume grows.
Start with twenty high-consequence conversations from your support desk. Remove personal information, mark every language switch and identify where meaning changes if a phrase is translated literally. Test the agent against those moments. Count successful outcomes, corrections and human handoffs instead of relying on transcript accuracy alone.
In the illustrative scene, Adwoa’s turning point comes when the service recognises the intent behind her Twi explanation and confirms the disputed transaction before moving forward. She lowers the paper in her hand. The payment later that afternoon is still possible.
That is the standard worth testing: not whether an AI voice answers, but whether the person can finish the conversation in the language they reach for when the outcome matters.
Comments
No comments yet.