Two things only you know, and the analysis needs both.
A conversation is evidence of what was said. It is not evidence of which campaign the agent was on, which offer was in force that week, or what was typed into your CRM afterwards. Elite AI lets you define your own reported metrics and your own context blocks, so the model answers the questions your operation actually asks and scores against the rules that were actually in force.
What you get
Your own metrics, reported on every conversation
Define up to ten questions of your own — contact reason, outcome, product mentioned — each asked only about the teams it applies to, and the model answers each one from a closed list you write. A closed list is what makes the answer countable: it charts, it filters, and it stays comparable across a year.
Questions only the recording answers
Listening questions have the model hear each transcribed call and answer what no transcript can show — was the agent clear, courteous, reading from a script — from your own options. It is told never to grade accent, every answer quotes the moment it was based on, and the answers are context for the scoring rather than a score of their own.
Context the recording cannot contain
Up to three named boxes per conversation — the campaign, the disposition, the notes pasted from a ticket — filled in by the evaluator and read by the analysis alongside the transcript.
Questions that used to be unanswerable
'Did the agent record the parameters correctly?' cannot be answered from a transcript, so the model used to mark it unverified on every call. With the context in front of it, the question gets a definite answer.
Treated as background, never as evidence
The prompt states that context is somebody's account of what happened, not a record of it: where it disagrees with the conversation, the conversation wins. Context is masked exactly like a transcript, and the agent can read what the analysis read.
How it works
Define them once, for the whole company
Metrics and context blocks are set at company level rather than per scorecard, so a scorecard being rewritten never changes what a stored answer meant.
Fill the context in as you go
The evaluator completes the boxes on the conversation. Retiring a block later does not erase what was already recorded.
Read them back as data
Your metrics and listening answers appear on the conversation, on the evaluation, and as a breakdown in the dashboard, where every value is a cross-filter like any other.
Questions people ask about this
Metrics are scoped to teams, so a company can hold ten while any single conversation is asked only the handful that matter for that agent's work. A context block costs whatever somebody pasted into it, on every analysis of that conversation, so three is the number that stays useful without burying the scorecard under background. Listening questions have their own ten: they ride on a separate pass over the recording and never compete with the scorecard at all.
Whatever your listening questions ask about how the call sounded: tone, clarity, pronunciation, courtesy. It is instructed never to assess accent or origin, and the settings screen warns you if a question looks like it does. The recording goes to the same Google enterprise platform that transcribes it, a second time; what the model writes back is masked with your rules. Listening is not charged to your AI allowance.
That is precisely the failure we design against. Context is presented to the model as background rather than as evidence, the conversation takes precedence wherever the two disagree, and every text field written in our forms is screened for instructions aimed at the model before it is saved.
Related capabilities
See it on your own calls
Book a walkthrough of the software. Send us a few recordings and your current evaluation form; we load them into Mira scores so you can watch the platform score your own calls, side by side with the scores your team assigned.
A live walkthrough of the platform with a person, usually within one business day.