Call Quality Agent

Overview

The Call Quality Agent reviews recorded inbound calls, distinguishes live conversations from IVR and voicemail, and scores eligible calls against operating rules. The work combines speaker-aware transcription, model review, an audit trail, and QA reporting.

The challenge

An inbound recording may contain a real conversation, an automated menu, voicemail, or a caller who disconnects before reaching an agent. Treating every recording as a conversation can produce misleading scores. Reviewers also need to understand why a score was assigned and inspect the underlying call, while teams need a consistent way to compare calls across campaigns.

Our approach

The pipeline transcribes recordings with speaker and channel information, then saves the transcript and its analysis trace. It marks short, single-speaker recordings as having no conversation. Longer recordings continue to model review so a diarization error does not automatically discard a real call.

A structured model response assigns speaker roles, intent and confidence scores, a campaign-specific qualifier, evidence, and a summary. The campaign rubric and qualifier travel with the transcript, connecting each review to the rules for that campaign. The system retains the model prompt, response, parsed scores, and usage metadata for audit.

In the reviewer app, teams filter and select calls, inspect scores and reasons beside synchronized audio and transcript, and open date-based reports by campaign or buyer.

The outcome

The reviewer workflow connects call classification, campaign scoring, transcript and audio inspection, audit traces, and reporting. Reviewers can inspect the evidence behind a score, return to the recording, and compare results against the campaign’s rules.