Every sales call scored against the script, with the evidence timestamped.

The problem.

The team sells on a structured script, and the difference between a good call and a lost deal is usually one skipped step: no emotional discovery, a price disclosed too early, no next step agreed. The manager had the recordings and the transcripts but not the hours. Which agent is drifting from the script? Where in the call do they lose it? When an inbound was missed, did anyone call back, and how fast? Answering those questions meant listening to hundreds of calls, so they went unanswered.

What we built.

• A daily sync that pulls every recorded call with its transcription from the CRM into an analysis database, joined to the agent and the linked seller lead.
• Pass one: categorise. A fast model reads each transcript and decides what kind of call it was (seventeen categories, from inbound discovery to offer presentation to post-signing), which direction it went, which script phases it contained, and which rubric should apply.
• Pass two: evaluate. A stronger model scores the call against the rubric for that category: eleven sections for discovery, separate rubrics for the two-programme pitch, the offer presentation and post-signing QA. Every section carries a weight and a grading type, and the output is a weighted compliance score, a quality score, the blunders, the highlights worth replaying and the key facts extracted.
• Pass three: verify. Only when a major or critical blunder is found, the model goes back to the timestamped transcript and returns the evidence with timestamps and speaker labels, so a coaching conversation starts from "at 4:12 you said" rather than an impression.
• Missed-call return metrics. Every missed, unanswered or voicemail inbound is matched to the next outbound call to that number within 48 hours, and the return time is recorded.
• A manager dashboard. Summary KPIs with colour thresholds, an agent scoreboard with daily trend drill-down, a phase heatmap showing where in the script the team drops, a blunder breakdown with a one-click false-positive report, a highlights feed, quality trends over time and per-agent activity.
• Prompts under version control in Langfuse, an hourly idempotent scheduler, and access control so only management sees the scores.

How it runs.

1. Calls sync from the CRM daily. The scheduler picks up anything not yet analysed, up to a hundred at a time, and never re-scores a call.
2. Each call goes through categorisation, evaluation and, where needed, verification. Cheap model first, expensive model only where it earns its keep.
3. Return metrics are computed for missed inbounds. The dashboard reads the analysis table live, so there is nothing to rebuild.
4. The manager opens the dashboard, sees who is drifting and where, clicks into the call, and coaches from the timestamped evidence. False positives get reported from the same screen and feed back into the prompts.

What changed.

• Coaching on receipts. Every review starts with the transcript line and the timestamp, not a memory of the call.
• Blunders caught the same day, including the ones that cost deals quietly: skipped discovery, early price talk, no next step.
• The training library builds itself from the highlights feed: real calls with excellent rapport or objection handling, tagged automatically.
• A fair comparison across agents, over time, on the same rubric.

The stack.

• AssemblyAI
• Claude Haiku
• Claude Sonnet
• Podio
• Postgres
• Next.js
• React
• Langfuse

Want this for your team?

Call scoring is one part of our AI Customer Operations offer, alongside chat agents, intake, document processing and ticket triage. If your team sells or supports on the phone and the recordings go unheard, this is the fastest win we know.