Answering more calls creates little value when a bank cannot prove who consented, who opted out, and who left without a complete answer. At 8:07 a.m., the useful metric is therefore verified call outcomes, not queue volume.
In 1999, NASA’s Mars Climate Orbiter was approaching Mars when the navigation data stopped matching the spacecraft’s expected path. One engineering team had produced impulse data in pound-seconds. Another processed that data as newton-seconds.
The spacecraft passed too close to Mars and was lost.
Arthur Stephenson chaired the investigation into the failure. NASA’s Mars Climate Orbiter Mishap Investigation Board Phase I Report documented the unit mismatch and the process failures around it. Teams had received data. Systems had processed it. Yet the chain could not reliably establish what the measurements meant before the consequences became irreversible.
A bank handling automated calls faces the same operational problem at a smaller scale. More activity does not repair ambiguous evidence.
The dashboard looks healthy at 8:07 a.m.
Consider a composite operations lead at a bank in Accra. She opens the morning dashboard and sees a queue spike. Customers are calling about account access, payment questions, and unresolved requests from the previous evening.
The voice agents are answering. Average wait time begins to fall. The completed-call count rises.
Then compliance asks three ordinary questions:
Which callers clearly consented to continue?
Which callers asked the bank to stop calling?
Which conversations ended before the agent gave a complete answer?
A call status marked “completed” cannot answer any of them. It may mean the call connected and ended normally. The caller may have heard the required disclosure, interrupted it, switched between Twi and English, asked to opt out, or hung up while the agent was still explaining the next step.
This distinction matters because telephony events describe what happened to the connection. They do not automatically establish what happened in the conversation.
That is why a green status cannot prove customer contact. The operations lead needs a record that connects the call lifecycle to consent, opt-out handling, and answer completion.
Call truth requires separate, reviewable states
A useful voice operation records more than “answered,” “failed,” and “completed.” It preserves the states a reviewer will need later.
Consent should have its own event. The record should show whether the disclosure was delivered, whether the caller agreed, and whether the system could confidently interpret the response.
An opt-out should become a durable instruction, not a phrase buried in a transcript. The system should record when it occurred and prevent another eligible campaign from treating that person as callable.
Answer completion also needs a defined standard. A connected call can still produce an incomplete outcome. The agent may fail to verify an account detail, misunderstand a switch into Twi, or end before explaining the required action. “The call ended” and “the customer received the answer” are two different facts.
Asenda Talk is being built around this separation. Its telephony lifecycle webhook pipeline tracks call events, while consent, opt-out, and audit records preserve the evidence needed for review. Metered per-minute billing also sits behind an operator-controlled real-money gate, so testing activity does not silently become live spend.
The platform remains in active early access. Voice agents can be configured with a persona, first message, and voice. Native Twi recognition and synthesis are fine-tuned in-house, with more African languages in progress. Vapi orchestrates the assistant runtime. Live outbound calling remains gated until an explicit telephony-provider decision is made.
Those boundaries matter. A credible operational record should show what the system did today, not imply that a roadmap item already happened.
Bilingual calls make weak records weaker
Consent and opt-out logic must survive the language people actually use.
A caller might begin in English, answer a verification question in Twi, then return to English for the final request. If the system recognizes the account number but misses the surrounding meaning, a successful field match can create false confidence.
The same risk applies to “stop calling,” especially when the request appears in Twi, in code-switching, or after the agent has started another prompt. Teams should test the full interaction and inspect the resulting audit trail. The practical standard is simple: a reviewer should be able to determine what the caller said, how the agent interpreted it, and what action followed.
The review process described in what a call review must prove after a caller says “stop calling” offers a useful starting point.
Build the evidence before increasing the volume
Before expanding a campaign, define the states that matter: consent confirmed, consent unclear, opt-out captured, answer complete, answer incomplete, and human review required. Then place sample calls in English, Twi, and natural code-switching.
Inspect the webhook history alongside the conversation record. Confirm that each opt-out changes future calling eligibility. Confirm that an interrupted explanation cannot appear as a completed customer outcome. Keep the real-money gate closed until those checks pass.
NASA’s investigators could trace the Mars Climate Orbiter failure because the underlying mismatch eventually became visible. The lesson for a call operation is to make critical mismatches visible while the next call can still be stopped.
At 8:07 a.m., the queue will demand speed. The operations lead should first be able to prove what every status means.
Comments
No comments yet.