Every call produces a record: the turn-by-turn transcript, a summary, and whatever actions the agent took. This is your primary tool for improving an agent and your primary evidence when someone disputes what was said.

​
Transcripts

The transcript is the literal exchange, turn by turn, with each turn attributed to caller or agent. Where a caller mixed Arabic and English, the transcript preserves the mix rather than normalising it. Use transcripts for debugging. When a call goes wrong, the failure is nearly always one identifiable turn — a misheard detail, a question answered before it was asked, an action fired too early.
When someone reports “the agent is bad”, ask for the call. The general complaint is rarely actionable; the specific turn almost always is.

​
Summaries

The summary is a short structured account of what the call was about, what was established, and what happened as a result. It is written for a human scanning many calls, not for an archive. Summaries are produced in both Arabic and English, so a bilingual team reads whichever it prefers without translating.
Summaries are generated from the conversation and can be imperfect on unusual calls. For anything contested — a disputed order, a compliance question — the transcript is the record, not the summary.

​
Reviewing at volume

Once a line handles real traffic, reading every call stops being possible. A workable rhythm:
1

Read everything for the first week

Every call, daily. Failure patterns show up fast and are cheapest to fix early.
2

Then sample

A fixed number per day, chosen randomly. Enough to notice drift.
3

Always read the exceptions

Every escalated call, every failed action, every call that ended unusually early. These are where problems concentrate.

​
What to look for

  • Repeated questions — the agent asked for something the caller already said. Usually an instruction problem.
  • Premature actions — fired before collecting enough. Tighten the action description.
  • Loops — the same exchange two or three times. Add an escalation trigger.
  • Language drift — replying in the wrong language after a code-switch.
  • Over-apologising — the most common reason an agent reads as artificial.

​
Retention

How long records are kept, and who can read them, is a policy decision with legal weight — call recordings and transcripts are personal data in most jurisdictions, and voice recordings attract specific obligations in several.
Settle retention and access before a line goes live, not after. Decide how long records are kept, who can read them, and what your callers are told. Involve whoever owns data protection in your organisation.