Title

AI Call Transcription & Sentiment Analysis for a Confectionery Business: From 5 Hours to 30 Seconds

Context

Business: Small confectionery

Industry: Food Production (Confectionery / Cake Production)

Problem

Managers spent 5 hours a day analyzing all the feedback from calls.

Solution

Build an n8n workflow that will:

• Check for unprocessed calls in the table every 2 minutes
• Transcribe them using AssemblyAI
• Check the processing status; if the status is "error," record this in the table
• If the status is "completed," have an AI Agent analyze the call (determine sentiment and category)
• Add the sentiment and category to a separate table (Data Table)
• Update the status to "done" in Google Sheets

Process

1. Problem & Data Source Analysis

I looked at how managers were reviewing calls and found they were listening to and analyzing every recording manually to extract sentiment and category — a task that consumed 5 hours a day and didn't scale as call volume grew. The core issue wasn't the analysis itself, but the fact that a human had to sit through raw audio before any insight could be extracted.

2. Designing the Polling & State Logic

Rather than triggering on a webhook (which the calling system didn't support), I designed the workflow to check the tracking table every 2 minutes for unprocessed calls. I built in explicit status states (unprocessed → transcribing → completed/error → done) so the system always knows exactly where each call is in the pipeline, and a failed transcription doesn't get silently lost — it's logged as an error instead of retried blindly or ignored.