Every call our AI makes in a regulated process can be reconstructed afterwards: the recording where one is agreed, the connection and decision logs, the statuses, and the exact version of the script that was live. That is what lets a bank, an insurer or a healthcare provider answer the question every auditor eventually asks: what happened on this call, and why?

A list of calls with durations does not answer that question. Nor does a transcript on its own. This article explains what an auditor actually needs, the four records that make reconstruction possible, and what it looks like to walk through a single conversation step by step.

What an auditor actually asks#

Auditors in regulated institutions are rarely interested in whether a call took place. They want causal reconstruction: what input the system received, under which configuration it was operating, what it said and decided as a result, and whether any person intervened. A generic call log covers the first item at best.

In practice the questions come in a few shapes. A customer complains that they were told something they were not. A compliance review samples calls to check that identity was verified before any debt detail was mentioned. An internal audit asks for evidence that a content change approved in March was actually live in April, and not before. A regulator asks how the organisation knows its AI system behaves as described. Each of these needs a different slice of the same evidence.

The four records#

What each record answers
RecordWhat it containsQuestion it answers
RecordingThe audio of the call, where recording is agreed and consentedWhat did the customer actually hear and say?
LogsConnection data (date, time, direction, numbers, operator) and the sequence of system events: recognised intents, data fields, decisions, utterances playedWhat did the system understand, and what did it do in response, second by second?
StatusesTechnical, dialogue, processing and business statuses at attempt and case levelHow did this call end, and what is the state of the case?
Script versionThe identifier of the approved, versioned scenario and content catalogue in forceWhich approved rules and words was the system using at that moment?

Retention and access follow the project's agreement and the data categories involved.

None of the four is enough alone. A recording without the script version cannot show whether what was said was approved. A log without the recording cannot show how the customer reacted. Statuses without logs show the outcome but not the path. Together they let you trace input, configuration, output and human action, which is the standard a regulated deployment should meet.

Reconstructing a call, step by step#

Take a typical early-collections call. In Vee Analytics, the reconstruction view combines the recording and transcript with time-stamped panels for intents, data fields, decisions and conversation stages. Walking through it looks roughly like this.

  1. Connection. Date and time of the call, direction, the numbers involved and the operator. The technical status shows that a person, not a voicemail, answered.
  2. Introduction. The system introduces itself as AI and names the organisation it is calling from. The log records which catalogue utterance was played and when.
  3. Identity verification. The customer answers the verification questions. The log shows each recognised intent and the data field captured, and a decision such as "need to verify the caller's identity", followed by the moment verification succeeded.
  4. The substance. Only after verification does the conversation reach the amount and due date. The customer says they can pay next Friday. The data field records the date, and a decision checks it against the bank's acceptable window.
  5. Clarification. If the system was unsure what the customer meant, the log shows the clarifying question and the answer that resolved it.
  6. Close. The system confirms the commitment and how to pay. The dialogue status records a promise to pay; the case-level status tells the bank's system what to do next.
  7. Return. The outcome appears in the daily results file the bank receives, mapped to the bank's own business status.

Each step carries a timestamp, so the timeline shows not only what happened but how long it took and where the customer hesitated. That timing is often what explains an outcome: a long pause before verification, an interruption during the amount, a stage that ran longer than the design expected.

Four kinds of status#

Statuses are where reconstruction meets day-to-day operations. We keep four axes separate, because mixing them causes most of the arguments about what a call achieved.

  • Technical: whether the line was reachable, busy, answered by a voicemail, or failed at the operator.
  • Dialogue: what happened in the conversation, for example verification failed, a promise to pay was made, or the customer asked for a person.
  • Processing: where the case is in its lifecycle: open, scheduled for retry, closed, excluded from the next day.
  • Business: the client's version of the truth, in the client's language, configured to match its own systems and not comparable across projects.

When a reviewer asks why a case was closed, the answer can be traced from the business status back through the processing and dialogue statuses to the specific call and moment that produced it.

Why the version number matters#

The script version is the record most often missing from automated contact, and the one that matters most once a process has run for a while. Content changes. A consequence phrase is softened, a verification question reworded, a new path added for customers who say they have already paid. If you cannot tell which version was live on a given call, you cannot tell whether the words the customer heard were approved at that time.

In our regulated configurations, customer-facing utterances come from a closed, approved and versioned catalogue. Each version has an approval owner and an effective date, and nothing changes in production without a versioned release: there is no silent learning in the interaction layer. That is what makes the version number meaningful. It points to a fixed, readable set of rules and words, not to a model state that may have drifted since.

Transfers, stops and human intervention#

Reconstruction has to cover the moments when the AI hands over or stops, not only the moments when it talks. A customer asks for a person. Verification fails in a market where the configuration says to transfer. A customer mentions a bereavement, a legal dispute or financial hardship, or asks for their data to be erased. Each of these maps to a defined route: a live transfer to the client's own staff, a structured stop or a referral. The evidence should show which rule fired, at what second, and where the call went.

This matters because human oversight is only real if it can be shown. Saying that a process has a human route is easy. Showing, for a sample of calls, that the route was offered when the rules said it should be, and that the transfer reached the right queue, is what turns oversight from a line in a policy into a capability. When a person on the client's side takes over, their part of the conversation sits in the client's systems; ours records the handover, its reason and its timing, so the two records can be joined.

The same applies to calls that never became conversations. Attempts that hit a voicemail, a busy line or an unreachable number carry technical statuses and timestamps too. When a reviewer wants to check that calling windows and attempt limits were respected for a given account, the answer comes from those records, not from the conversations.

Reconstruction is a daily tool, not only an audit file#

We do not build this evidence trail only for auditors. It is how we run the service. Our QA team listens to selected calls in every project, every day, especially unusual ones, and uses the reconstruction to see whether speech recognition had a fair chance, where a conversation broke down, or whether something interrupted the customer. Our analysts use the same data to compare dialogue variants and timings.

Every day, in every project, our people listen to selected calls, especially unusual ones, to hear with their own ears what technology still cannot. – Dawid Wójcicki, Founder & CEO of Vee

Because the people who design the process also use the evidence daily, gaps in it are noticed quickly. An evidence trail that is only opened when an auditor arrives tends to be incomplete exactly when it is needed.

Reconstruction also has limits that should be stated plainly. Recordings are kept only where agreed and only for as long as the project's retention rules allow; regulations require old recordings to be deleted, and they are. Access to recordings is restricted to authorised people under the data-processing agreement. A good evidence design balances what an auditor may need against what data protection requires, and that balance is set per project with the client.

What to ask a supplier#

  1. For a call we choose at random, can you show the recording, the event log, the statuses and the script version together?
  2. Does the log show recognised intents, captured data fields and system decisions with timestamps, or only the transcript?
  3. How do you prove which content version was live on a specific date, and who approved it?
  4. Can the system change its behaviour in production without a new version? How would that appear in the evidence?
  5. How are retention, access and deletion of recordings configured for our project?
  6. If a human from our side intervened in a call, where is that recorded?

If a supplier can answer these on a real call, in front of you, the process can be defended. If the answers depend on what the model probably did, it cannot. In the processes we run, we treat reconstructability as part of the product, because for a regulated client it is.

Follow The First Second

Occasional issues on the mechanics and economics of AI-powered customer contact. No noise, no vendor news.

Follow via RSS