CASE STUDY / Automations
Reply drafting with the human kept in the loop
FUGU GMBH · 360 PROGRAM · 2026
SITUATION
Participants reply to program mail with real concerns, and those replies arrive across weeks while a cohort is running. Each one needs the thread's history, the program's rules and a tone that does not read as automated. That is a lot of context to rebuild by hand, several times a day.
APPROACH
The workflow reads the inbox, drops anything from an address that is not in the program, and rebuilds the conversation from the log rather than trusting the quoted text below the reply. It reads the sentiment, retrieves the relevant program documents, and writes a draft in Outlook. Then it stops and asks a person, in the channel they already work in, and waits for the answer.
RESULT
Across the programmes 1,600+ people were reached and about a fifth of them asked something. All of them were answered, none of the answers was sent without a person reading it first, and drafting them is worth an estimated 10 to 15 hours a month. The system does the expensive, repetitive half, which is assembling context. It does not touch the half that carries the relationship.
- A contact who is not on the program is ignored, not answered generically.
- Conversation history is rebuilt from the log, so a trimmed quote cannot mislead the model.
- The program's documents re-index themselves whenever a source file changes.
- Approval happens in the team chat and blocks the workflow until it comes back.
n8n · Microsoft Outlook and Teams · Supabase vector store · OpenRouter



Figures read 2026-09-04.