Context
"Солодка Мрія" is a local bakery that receives many customer phone orders and feedback. The business accumulated thousands of call recordings, but nobody had the time to listen to them and analyze customer complaints.
Problem
- No Data: Management couldn’t see the real reasons behind customer dissatisfaction.
- Wasted Time: Manually listening to a single 10-minute call took a manager 15 minutes.
- No Sorting: It was impossible to track whether issues were caused by delivery or product quality.
Solution
I built a two-part automation system in n8n that completely automates call analysis:
1. Processing Pipeline
- Audio to Text: Call recordings are automatically sent to AssemblyAI for transcription.
- AI Analysis: OpenAI reads the text, detects if the call was positive or negative, and tags the specific issue (e.g., "Delivery", "Cake Quality").
2. Interactive AI Analyst
- A custom AI Agent connected to the internal Data Table.
- The manager can just type in plain English (like "What's the statistics for this week?"), and the agent automatically calculates dates and generates a full text report.
Process
- Receive audio: Call recordings enter the n8n processing pipeline and are sent to AssemblyAI.
- Transcribe the conversation: AssemblyAI converts audio into text. Speaker diarization identifies separate speakers in the conversation; this transcription stage is distinct from the later AI interpretation.
- Interpret the transcript: OpenAI analyzes the text for positive or negative sentiment and assigns issue categories such as Delivery or Cake Quality.
- Store structured results: The workflow saves the analysis to Google Sheets for reporting and backup, and to n8n Data Tables for the conversational analyst.
- Answer reporting questions: An AI Agent queries the internal Data Table. It interprets the requested date range and produces a text report, such as a weekly breakdown of complaints.
Results
- Reported processing comparison: The project materials report under one minute of automated processing compared with 15 minutes of manual review for a 10-minute call, described as a 15× speed-up. This is a project-reported comparison, not an independently verified production benchmark.
- Conversational reporting: The AI analyst generates weekly complaint breakdowns from the stored call data.
- Automated analysis: Transcription, classification and storage run as a background workflow, reducing the need to manually review each recording.
Tech Stack
- Automation: n8n (Advanced AI, Chat Trigger, Data Tables)
- AI Models: AssemblyAI (STT), OpenAI API (LLM)
- Backend: Google Sheets, Webhooks
- Infrastructure: VPS, Linux, Domain & DNS management
Visuals
Watch the Loom walkthrough
