The Bank Engineer’s “Completed” Call. The Auditor Needs Proof of Who Ended It.

A call summary cannot reliably tell an auditor which system ended a call. The answer must come from timestamped telephony lifecycle events that distinguish an assistant decision, customer hang-up, provider disconnect, transfer failure, and infrastructure error.

In 1999, the Mars Climate Orbiter reached Mars after a journey of hundreds of millions of kilometres. Then NASA lost contact. The mission’s software had produced numbers that looked usable, but two parts of the operation did not describe those numbers in the same units.

NASA’s mishap investigation, chaired by Arthur Stephenson, found that one team had supplied thruster data in pound-force seconds while another expected newton-seconds. The spacecraft approached Mars at the wrong altitude and was lost. The final output had concealed a disagreement much earlier in the chain.

A bank call marked “completed” presents the same class of evidence problem at a far smaller scale. A tidy status can hide conflicting events underneath it.

The summary reaches its limit

It is Friday at 4:17pm in Accra. An engineer working through an audit sample opens the record for an outbound customer call. The AI summary ends with two words: “call completed.”

The auditor asks a narrower question: which system ended it?

The answer determines what happened next. Did the customer hang up after hearing the disclosure? Did the assistant terminate after detecting an opt-out? Did the telephony provider disconnect? Did the runtime stop while the underlying call remained active for several seconds? A summary written for quick review cannot settle those questions by itself.

The engineer turns to the telephony events. They show the call was created, connected, and later disconnected. An assistant event appears close to the final provider event, but ordering by display time leaves room for doubt. One timestamp came from the assistant runtime. Another came from the provider webhook. Delivery delay means arrival order may differ from event order.

“Completed” describes a broad terminal state. It does not prove who initiated that state, why it occurred, or whether every system agreed.

Call truth lives below the transcript

Transcripts and summaries matter. They help reviewers understand what the customer and agent discussed, including a switch between Twi and English. They remain interpretations of the conversation layer.

Call-truth tracking answers a different set of questions:

• Which provider event confirmed that the call connected?
• Which component requested termination?
• What reason code accompanied the disconnect?
• When did each event occur, and when did the platform receive it?
• Did the assistant record an opt-out before the connection closed?
• Can the original event payload be tied to the reviewed call?

Asenda Talk’s telephony lifecycle webhook pipeline is built to retain this operational sequence. Consent, opt-out records, and an audit trail sit alongside it so a reviewer can trace the call rather than rely on a generated sentence.

That distinction matters in financial services. A completed status does not establish consent, and it does not explain whether an opt-out was respected. The related audit problem is examined in Financial Call Compliance Audits: Why a Completed Call Does Not Prove Consent.

A defensible answer needs ordered evidence

The bank engineer’s useful response is specific and limited: the provider recorded the final disconnect, while the assistant had issued a termination request shortly beforehand. The retained events support that sequence. They do not support a stronger claim about anything the systems failed to record.

That answer gives the auditor something testable. It also exposes the controls the team should inspect:

• Store both the event’s source timestamp and its receipt timestamp.
• Keep provider reason codes in their original form.
• Give every call and event a stable identifier.
• Record assistant termination requests separately from provider confirmations.
• Treat summaries as review aids, never as the sole audit record.
• Preserve opt-out evidence even when a call ends unexpectedly.

The same discipline applies when language affects intent. A Twi phrase that signals refusal must survive transcription, classification, and call control. The Missed Twi Opt-Out That Made Adwoa Pause Monday’s Campaign explores that failure path.

Asenda Talk supports native Twi speech recognition and synthesis fine-tuned in-house, with Vapi orchestrating the assistant runtime. It is in active early access. Outbound calling remains behind an operator-controlled real-money gate while the live telephony-provider decision is unresolved. Those boundaries belong in any current evaluation because audit claims must match the system actually running.

Test the disagreement before an auditor finds it

The Mars Climate Orbiter investigation did not stop at the fact that contact was lost. It traced the loss through interfaces, assumptions, and units until the disagreement became visible.

Voice AI teams need the same habit. Before enabling a real campaign, create test calls where the customer hangs up, the assistant ends the conversation, the provider rejects the connection, a webhook arrives late, and an opt-out occurs immediately before disconnection. Then ask one person who did not build the flow to reconstruct each outcome from stored evidence.

If that person can see only “call completed,” the audit trail is unfinished. The next practical step is to make every terminal state answer three questions: what ended, which system ended it, and which retained event proves the answer.