Sapa Group
Industrial group in automotive components.
From 1 hour to 3 minutes per meeting
Result: 95% time savedReport in 3 minutesMeetings up to 3 hours
- The problem
- Their AI transcription tool couldn't reliably tell who spoke or for how long: the action items and reports it produced were wrong and left out what mattered. Useless reports, exactly when they were needed at the ownership table for strategic decisions.
- How we built it
- We built a multi-step agentic automation (N8N, Pinecone, OpenAI) that starts from the raw transcript: multiple agents split the meeting into dynamic, token-based chunks, so even a three-hour meeting never breaks the context window, and reason together to grasp the context and tie every point to who said it. The report then goes through a human-in-the-loop cycle: if the user approves it, it's embedded into a vector database (report and transcript, with date and participant metadata) and stays queryable over time as a decision history; if they reject it, an agent learns from the feedback and rewrites the structure until it's right.
- The outcome
- A long meeting becomes, in three minutes, a readable report with action items, decisions, who they're assigned to, deadlines and next steps, and meeting knowledge turns into a queryable company memory.
Tech stack
N8NPineconeOpenAIHuman-in-the-loop
“We were overwhelmed by too many meetings and the previous software was completely inaccurate. Now in three minutes we have clear action items and make immediate decisions.”
Giovanni Affinita · Executive Director
