The Receipt Esi Held, and What a Twi Switch Could Have Cost Her

A bilingual billing agent should keep the caller’s original task active when the caller switches to Twi. The language change is a change in expression, not permission to discard the account, charge, or question already under discussion.

At 4:40 p.m., Esi stands behind the counter of her small shop in Kumasi, one hand on a receipt and the other holding her phone. She called about a charge she does not recognise. The agent has already confirmed the last four digits of the account reference and asked which billing period she means.

Then Esi switches to Twi. She explains that the charge appeared after a payment she thought had failed, and asks whether the same payment was taken twice.

A weak bilingual agent hears the new language and starts again: “How can I help you?” Esi now has to repeat the account context, the disputed charge, and the reason she called. If the agent records the second explanation as a separate request, the billing team may investigate the wrong issue. The original complaint can disappear into a transcript that looks complete but has lost its meaning.

The caller is continuing the same task

Code-switching is ordinary conversation for many Ghanaian customers. A person may begin in English because the first prompt is in English, then move into Twi when the explanation becomes personal, technical, or difficult to phrase.

The agent needs to preserve the active intent across that switch. In Esi’s call, the task remains: verify whether one payment produced two charges. The language used to explain the task has changed. The task itself has not.

That distinction affects every later step. A useful call record should connect the account reference already gathered, the charge under review, the caller’s clarification in Twi, and the next action. A transcript that accurately captures Twi words but detaches them from the earlier English exchange still leaves the support desk with an incomplete case.

This is especially important in billing, where a caller may mention a payment date, an amount, a reference number, and a concern in different turns. The agent should carry those details forward instead of treating each language segment as a fresh conversation.

Context must survive recognition, reasoning, and handoff

Language recognition is only one part of the call. The harder test is whether the system can retain what the caller means as the conversation moves between English and Twi.

For an agent builder, that means defining the billing task before the call begins. What details should the agent collect? Which facts must it repeat back for confirmation? When should it stop and route the case to a human? These rules give the conversation a stable frame when the caller changes language mid-sentence.

Asenda Talk is being built around native Twi speech recognition and synthesis fine-tuned in-house, alongside agent configuration for persona, first message, and voice. Its runtime is Vapi-orchestrated. Those components make it possible to evaluate a Twi and English conversation as one interaction, rather than treating Twi as a decorative opening followed by an English-only workflow.

The evaluation should be concrete. Give the agent a billing scenario in English, switch to Twi while clarifying the charge, then ask it to restate the live issue. Check whether it preserves the disputed transaction and the requested action. One clean response matters more than a long transcript with the wrong thread.

For a related evaluation lens, see Twi Voice Agent Evaluation: Why One Clean Response Matters.

A complete record protects the customer and the team

Esi’s call does not end when the agent understands her. The support desk needs to know what the caller consented to, what the agent said, whether Esi opted out of future contact, and what follow-up is required.

Asenda Talk includes consent, opt-out, and audit-trail handling for every call, plus a telephony lifecycle webhook pipeline designed to track what actually happened during a call. That record matters when a supervisor needs to see whether the agent promised a review, requested further evidence, or ended without a resolution.

The record also needs to preserve the conversation’s order. If Esi first identifies a disputed charge in English and then supplies the reason in Twi, a reviewer should see one connected path. Otherwise, a billing team can mistake an explanation for a new complaint, or close the original complaint before anyone has addressed it.

A bilingual support flow also needs an owner for cases the agent cannot safely resolve. If Esi asks for a refund, disputes account ownership, or gives details the agent cannot verify, the handoff should carry the active billing issue forward. The customer should not have to rebuild the case from the beginning for the next person.

Build for the live call you expect

The practical test is simple: write down the customer’s first request, then add a Twi switch at the point where the explanation becomes more detailed. Review whether the agent can name the same request, the same relevant facts, and the next action after the switch.

Esi eventually hears the right confirmation in the language she chose: the case concerns the same disputed charge, and it has been routed for review. She puts the receipt back under the counter because she no longer needs to call again and explain the whole thing from the start.

Asenda Talk is in active early access. Teams can configure and evaluate agents today, while feature parity with established voice-agent platforms remains in progress. Outbound calling also remains gated behind an explicit telephony-provider decision that has not been made live. That makes pre-launch conversation evaluation especially valuable: find the point where context breaks before a real billing caller has to pay for it.