An automated voice call should state, before the first substantive question, whether it is collecting information or assessing eligibility. It should also explain how each answer will be used, whether the call is recorded, and how the applicant can pause, opt out, correct the record, or reach a person.
Consider an illustrative applicant named Efua, a seamstress in Kumasi who has applied for a small business grant. At 2:18 p.m., between customers, she answers an unfamiliar number with a measuring tape still around her neck.
The voice greets her in Twi and asks about her monthly income. Efua pauses. Is the agent documenting what she entered on the application, checking whether her figures are consistent, or deciding whether she deserves the grant?
She gives a cautious answer. The next question asks how many people depend on her income. Now every word feels like it could move her closer to approval or quietly disqualify her.
The call sounds natural. Its purpose remains unclear.
Natural Twi cannot fix an unclear decision process
Language quality matters. An applicant should be able to answer in the language that lets them express dates, income changes, family obligations and uncertainty accurately. A system that loses meaning when someone switches between Twi and English can damage the record before any eligibility rule is applied, as explored in Twi and English Code-Switching: Why Voice Agents Must Preserve Context.
But accurate speech recognition solves only one part of Efua’s problem.
She needs to know the function of the conversation. “We need to ask you a few questions” leaves too much hidden. An interview may gather facts for later review. An assessment may score or classify answers during the call. A verification call may compare new statements with an existing application. Those processes create different risks for the person answering.
If Efua thinks the agent is only recording facts, she may speak casually about a good sales month. If she suspects an assessment, she may shorten, qualify or second-guess the same answer. Ambiguity changes behaviour, which can change the data the system receives.
The agent must disclose what each answer can affect
Before asking about income, dependants or business activity, the voice agent should give Efua a plain explanation in Twi or English:
“This call will collect information for your grant application. Your answers will be recorded in your application file. The voice agent does not make the final eligibility decision. You can ask to repeat a question, correct an answer, stop the call or request human help.”
That wording would need to change if the automated system scores answers or influences eligibility. The disclosure should say so directly. A vague consent prompt cannot carry that burden.
Question-level context also matters. “This answer helps us confirm your current business income” gives Efua a clearer basis for responding than a sequence of unexplained prompts. When the subject changes, the agent can state why.
The system should preserve the applicant’s exact response, any correction, the consent event and any opt-out request. A summary alone may remove hesitation or qualification that matters later. For a practical governance framework covering language access, human escalation and auditable records, see the governance checklist for African public-service voice agents.
Uncertainty should trigger a safe exit
Halfway through the imagined call, Efua asks in Twi, “Will this answer decide whether I qualify?”
The agent repeats the previous question.
That failure changes the stakes. Efua must either continue without understanding the consequences or end the call with an incomplete application record. The grant could be lost, and she has no reliable way to know which choice creates the greater risk.
A safer design treats questions about purpose, scoring and eligibility as escalation triggers. The agent should pause the interview, acknowledge the uncertainty, explain what it knows, and offer a human route when it cannot answer. It should never improvise a policy explanation.
The same rule applies when the applicant disputes a transcript, says an earlier answer was misunderstood, or asks whether changing languages will affect the record. Continuing the script may complete the call while weakening the evidence behind it.
Build the disclosure before opening the phone line
Asenda Talk is in active early access. It supports configurable voice agents, native Twi speech recognition and synthesis developed in-house, consent and opt-out records, call lifecycle tracking, and auditable call events. Vapi orchestrates the assistant runtime.
Live outbound calling remains gated while the telephony-provider decision is unresolved. That gate matters here. A team should define the call’s purpose, eligibility influence, escalation path and record-retention rules before enabling real-money calling.
Return to Efua at 2:26 p.m. In the safer version of the scene, the agent pauses when she asks whether her answer affects eligibility. It explains that it is collecting information for human review, records her correction to the income figure, and confirms that she can end the call without losing what she has already provided.
Efua takes the measuring tape from her neck and answers the next question without guessing what the machine has decided about her.
Comments
No comments yet.