Sales Call Analytics With AI: Transcription, Scoring, Insights
Practice3 minInnovialabs editorial team
In short
Calls are pulled from the PBX, transcribed by ASR, scored by an LLM against a checklist with objections extracted as JSON, and the result is written to the CRM deal, so the sales lead sees every conversation instead of a random sample.
How it works
AI call analytics is a five-step pipeline:
- Fetch the recording from your PBX or cloud telephony along with metadata: agent, customer number, duration, deal ID.
- Transcribe speech to text (ASR) with speaker separation.
- Score the conversation against the script checklist with an LLM.
- Extract objections, agreements and the next step.
- Write the result to the CRM and build reports.
The main value: the sales manager sees every call, not the handful they had time to listen to.
Steps 1-2. Recording and transcription
- Get recordings via the PBX API or webhook right after the call ends.
- If the PBX records in stereo (agent and customer on separate channels), use it: speaker separation becomes more accurate.
- For Uzbek and mixed Russian-Uzbek speech, test ASR quality on your real calls before launch.
- Skip very short calls and missed calls to save resources.
Step 3. Checklist scoring
Turn your sales script into specific, verifiable items. Bad: "the agent was polite". Good: "the agent stated their name and company at the start of the call".
A sample prompt with JSON output:
You are analyzing a sales call. Score the transcript against the items.
For each item return: id, passed (true/false), quote - a short quote
from the transcript as evidence. If there is no evidence, passed = false.
Items:
1. greeting - introduced themselves and named the company
2. needs - asked at least one question about the customer's needs
3. next_step - agreed on a specific next step with a date
Also return: objections (list of customer objections), summary (2-3 sentences).
Respond with valid JSON only.
Requiring an evidence quote sharply reduces made-up scores and lets the agent see exactly why an item failed.
Step 4. Objections and insights
LLMs are good at extracting:
- objections: "too expensive", "need to discuss internally", "we already work with someone";
- how the agent handled them — and whether it worked;
- mentions of competitors and products;
- commitments: what was promised to the customer and by when.
Map objections to a fixed list of categories, otherwise the model invents new wording each time and reports will not add up.
Step 5. Writing to the CRM
Useful fields on the deal or contact:
| Field | Why |
|---|---|
| Short call summary | Anyone grasps the context in seconds |
| Checklist score | Quality control and coaching |
| Objections | Analytics on loss reasons |
| Next step and date | Automatic task for the agent |
| Recording link | Quick review of disputed cases |
Most CRMs let you update fields and create tasks via API.
Common mistakes
- Vague checklist — the model scores by gut feeling and results are not comparable.
- No calibration: before launch, compare AI scores with the sales manager's scores on a few dozen calls and refine the wording.
- Using scores to punish without review — agents stop trusting the system.
- Ignoring privacy: customers must be informed about recording, and access to transcripts should be restricted.