Ghana’s future AI rules should require a qualified human to review any high-risk call where meaning moves between Twi and English before deployment. The review should examine the audio, each language segment, the translated meaning, the agent’s action and the audit trail, because a clean English transcript can hide a dangerous Twi error.
In 1999, NASA’s Mars Climate Orbiter was approaching Mars when its navigation team faced a growing discrepancy in the spacecraft’s predicted path. The mission depended on numbers passing correctly between systems built by different teams. One system produced thruster data in pound-force seconds; another expected newton-seconds.
The spacecraft was lost on 23 September 1999. Arthur Stephenson chaired the Mars Climate Orbiter Mishap Investigation Board, whose report documented the unit mismatch and the processes that failed to detect it before the final approach.
The numbers travelled. Their meaning did not.
A language transition can change the action
Picture a supervisor in Accra reviewing a recorded customer-service call before a voice agent goes live. The caller begins in English, switches to Twi while explaining the sensitive part of the problem, then returns to English when the agent confirms the next step.
The final transcript looks orderly. The English summary says the customer agreed to proceed.
The supervisor plays the original audio and hears something more complicated. During the Twi segment, the caller qualified the request, corrected a detail or withdrew permission. The agent’s English response treated that statement as confirmation.
This example is illustrative, but the failure pattern is concrete. Speech recognition can mishear a phrase. Translation can preserve the topic while losing a condition, negation or change of intent. The agent can then take an action that appears reasonable when judged only from the English transcript.
That is why a high-risk language transition needs review at the boundary where meaning changes form. Reviewing the agent’s final answer alone is like checking the Mars Climate Orbiter’s trajectory without checking whether both systems used the same unit.
High risk depends on consequence, not call length
A ten-minute call about opening hours may carry little risk. A thirty-second exchange about consent, payment, health, eligibility, debt, account access or a service cancellation may carry much more.
Future rules should define review requirements around what the agent can cause. Human review should apply before deployment when a Twi-to-English or English-to-Twi transition can influence:
- whether a person has consented or opted out;
- whether money, credit or access could be affected;
- whether a complaint enters the correct queue;
- whether an urgent statement receives the right escalation;
- whether the system records a binding decision or instruction.
The reviewer should have access to the original audio, language-labelled transcript, translated text, agent response and resulting system action. A single English summary cannot show where recognition ended, translation began or a decision rule fired.
This distinction also matters after launch. If a customer disputes what happened, the operator needs a call-level audit trail showing consent, opt-out events and the telephony lifecycle. A clean transcript may hide an earlier opt-out, particularly when the disputed words occurred in another language.
The reviewer needs authority to stop deployment
A required review becomes ceremonial if the supervisor can flag a problem but cannot block release.
The rule should require documented approval from someone with the language ability and operational authority to evaluate the transition. That reviewer should be able to mark a test as failed, identify the exact audio span, record the expected interpretation and prevent deployment until the defect is corrected and retested.
A useful review record would answer five questions:
- What did the caller say in the original language?
- What meaning did the system record?
- What did the agent say or do next?
- Could the difference affect consent, money, access, safety or legal rights?
- What evidence shows the corrected version passed a new test?
The test set should include code-switching, interruptions, corrections and opt-outs, since real callers rarely keep each language in a separate, tidy block. One transition can invalidate an otherwise convincing demonstration, as this bilingual voice-agent testing example shows.
Build the evidence before calling is enabled
Asenda Talk is in active early access. It currently supports configuring voice agents, including persona, first message and voice, alongside native Twi speech recognition and synthesis fine-tuned in-house. It also has call-truth tracking, consent and opt-out records, audit trails, metered billing controls and a Vapi-orchestrated assistant runtime.
Those controls create evidence for review. They do not prove that every high-risk language transition is safe. Outbound calling also remains behind an explicit telephony-provider decision and an operator-controlled real-money gate.
That boundary should remain visible. Before a high-risk workflow reaches live callers, an operator should select representative bilingual calls, review every consequential transition, record failures and repeat the test after correction. Product teams should publish what has passed, what remains under evaluation and which calling capabilities are still gated.
The Mars Climate Orbiter board did not treat the lost spacecraft as the result of one bad number in isolation. Its report examined why the surrounding process failed to detect the mismatch. Ghana’s eventual AI rules should apply the same discipline to bilingual voice systems: inspect the transition, preserve the evidence and give one accountable human the power to stop the call.
Comments
No comments yet.