Roger Boisjoly’s O-ring warning. The decision system failed.

Absa’s AI customer service ambitions require audit records that preserve what customers actually said, in the language they used, along with consent, opt-outs, system actions and call outcomes. An English transcript alone can conceal recognition errors in Twi or another African language, leaving compliance teams to review a translated interpretation rather than the original interaction.

In January 1986, engineer Roger Boisjoly and colleagues at Morton Thiokol warned that the Space Shuttle Challenger’s O-ring seals could perform poorly in unusually cold conditions. During a teleconference before launch, the company initially recommended against launching at the expected temperature. Managers later reversed that recommendation.

Challenger broke apart 73 seconds after launch on January 28, killing all seven crew members. The Rogers Commission documented the technical failure and the decision process surrounding it. The warning had existed. The system for carrying that warning into the final decision had failed.

Banking voice AI presents a smaller-stakes version of the same structural problem. A record can exist while still failing to carry the meaning that mattered.

An English transcript can create false confidence

Absa has said its chatbot handles 100,000 queries each month, while 1,400 developers use AI coding tools. At that volume, individual interactions quickly become operational records: evidence for complaints, quality reviews, fraud investigations and regulatory enquiries.

In African markets, those records may begin with customers moving between English and a local language during the same conversation. A customer might state a name in English, describe a disputed transaction in Twi, then return to English for an account detail. If the audit trail retains only an English rendering, reviewers cannot easily distinguish the customer’s words from the system’s interpretation.

That distinction matters when a model misses a negative, changes a date, confuses a person’s name or fails to recognise an opt-out. The resulting transcript may read cleanly while representing the call poorly.

The problem is examined more closely in The English-Only Transcript That Nearly Closed Adwoa’s Dispute Incorrectly. The practical lesson is simple: readability does not prove fidelity.

Compliance starts before transcription

A useful audit trail should reconstruct the call from initiation to final outcome. That means recording which agent configuration ran, which voice and language settings were active, when consent was captured, what the caller said, what the system recognised, what response it generated and how the call ended.

Opt-outs need the same treatment. A transcript containing “do not call me again” offers limited protection if the suppression action was never recorded or applied. Teams need evidence connecting the customer’s request to the system response and the resulting contact status. What Happens When a Voice Agent Ignores a Customer’s Opt-Out? explores that failure path.

Call-truth tracking also matters. A dashboard should distinguish an answered call from a failed connection, a completed conversation from an early hang-up, and a billed minute from an attempted call that never reached the customer. Without lifecycle events, teams may reconcile polished summaries against an incomplete account of what happened.

Native-language evidence changes the review process

Language support should extend through recognition, synthesis, evaluation and audit. Adding Twi at the conversation layer while keeping evaluation English-only pushes the uncertainty downstream. Compliance reviewers then see the translation without the evidence needed to challenge it.

A stronger review package preserves the original audio under an appropriate retention policy, the language-specific recognition output, any translation used for review, timestamps, model or configuration versions, consent state and telephony events. Sensitive access should be controlled and logged. Secrets should remain masked and write-only rather than appearing in operational records.

This also changes testing. Teams should evaluate code-switching, names, numbers, negation, consent phrases and opt-out language with speakers who use Twi naturally. Aggregate accuracy can hide the handful of errors that determine whether a complaint is upheld or a customer is contacted again.

Asenda Talk is being built around native Twi speech recognition and synthesis fine-tuned in-house, with Vapi orchestrating the assistant runtime. Its current platform includes configurable agents, telephony lifecycle webhooks, call-truth tracking, metered billing controls, consent records, opt-outs and audit trails. It remains in active early access, and live outbound calling is gated while the telephony-provider decision remains unresolved. More African languages are in progress rather than available today.

Build the evidence path before call volume grows

The Rogers Commission’s record shows why preserving a warning is insufficient when the decision system cannot carry its meaning forward. For Absa and other banks expanding AI-assisted service, the equivalent test is whether a reviewer can reconstruct a multilingual interaction without trusting a single English summary.

Before increasing call volume, define the evidence package for one disputed conversation. Confirm that it links the original language, recognition output, generated response, consent status, opt-out action and final call event. Then test whether a reviewer who was absent from the call can determine what happened and where uncertainty remains.

That review should happen while traffic is manageable. Once 100,000 monthly queries become a reference point for broader automation, gaps in language evidence will scale with the service.