The challenge
A telecalling team that records hundreds of calls a day can only review a tiny fraction of them by hand. This sales team had exactly that problem. Supervisors did not have time to listen to every call, so quality checks ran on a small random sample, and almost every recording went unheard.
The feedback that did arrive came days or weeks after the call. By then it was too late to change behaviour on the next call. Two reviewers could also score the same call differently, so agents had no stable picture of how they were doing. Most of what they heard was correction after something went wrong, with very little said about what they did well.
- Manual review did not scale past a handful of calls a week per agent.
- Feedback came too late to change behaviour on the next call.
- Scoring was subjective and varied between reviewers.
- Agents got almost no positive reinforcement.
The raw material was messy as well. Recordings landed on an FTP server from the team’s call recorder, in more than one filename format, and many came from phones that save M4A audio with a mislabeled header.
What we built
Amatec built two connected n8n workflows: one scores every call recording within the hour, and the other turns those scores into personal coaching that agents can ask for in Slack. Both workflows are also published as MCP-callable, so an AI assistant can trigger them.
The first workflow runs every 60 minutes and works through each new recording in order:
- Lists new files on the FTP server and keeps audio only, WAV or M4A.
- Reads the call time and call details from the filename with a code node.
- Checks Postgres and skips any file already processed, so no call is scored twice.
- Matches the recording to the right agent by the folder it was saved in.
- Repairs the audio header where needed, then transcribes the call with OpenAI Whisper.
- Drops transcripts too short to be a real conversation.
- Scores the call with GPT-4o-mini and saves the transcript and score to Postgres.
- Deletes the recording from the FTP server to keep storage clean.
The header repair is the detail that made the system work on real phones. Many handsets save M4A recordings with a 3GP header that OpenAI rejects. A small code node reads the start of each file, rewrites the 3GP label to a standard MP4 one and passes the patched audio on to Whisper. That one fix lets the workflow handle real phone recordings without breaking.
Scoring uses a fixed rubric returned as strict structured JSON, so every call produces the same fields: a call score out of 10, talk ratio, tone (positive, neutral or aggressive), an objection handling score out of 10, key issues, suggestions and a short summary. The model runs at a low temperature of 0.2 and must return every field, every time.
The second workflow is the coaching bot. An agent sends it a Slack DM, for example “how am I doing” or a question about a specific call. The bot confirms the agent is registered, then pulls their last 10 scored calls, their total call count and their recent chat history from Postgres. Anyone not registered gets a polite reply asking them to have their admin add their Slack ID.
That data goes to an AI Agent running GPT-4o-mini, with a live Postgres tool for looking up older calls. Because the agent’s real scores are already in context, the bot never asks them to describe a call. It is instructed to be warm and specific, to celebrate strengths and to frame weaknesses as things to work on. A new agent with no scored calls yet gets encouragement and three practical telecalling tips. Every exchange is logged, so the next conversation picks up where the last one ended.
The results
Every call the team records is now transcribed, scored and stored within the hour it was made, which means 100% of calls are reviewed.
- 100% of calls reviewed, where a supervisor once sampled a few.
- Agents get an objective score within the hour instead of waiting days for a supervisor.
- Coaching is available on demand. Any agent can ask “how am I doing” and get a real answer instantly.
- Coaching replies reference real scores and real moments from the agent’s own calls.
- Corrupted headers, short or silent recordings and duplicate files are handled without breaking the run.
The bigger shift is in who starts the feedback loop. Agents no longer wait for a review to come to them. They ask the bot when they want to know, and the answer comes from the same scored data a supervisor would see in Postgres.
This setup suits any team running telecalling, sales or support calls that wants consistent quality scoring and same day coaching without hiring a QA team to listen to every call.