CASE STUDY / Automations
Support Assistant — nobody answers a question they should not
FUGU GMBH · 360 PROGRAM · 2026
SITUATION
A leadership program generates the same twenty questions from every cohort: which link is mine, who did I nominate, when does this close. Answering them by hand is a tax on the program owner. Answering them with a bot is worse if the bot will tell anyone anything, because the questions are about other named people.
APPROACH
Identity first, and no exceptions. The assistant walks a state machine held outside the conversation: greet, classify whether you are a participant or an organizer, ask for the email, look it up against the roster, ask for the password that belongs to that role, and only then move into a state where answering is allowed. A separate guard checks each question is about the program before the answering agent ever sees it.
RESULT
Across six complete programmes and half of a seventh, sixteen participants each ask continuously about their 360 status, their missing raters and everything else, on top of organizer questions about overall progress. Those questions stop reaching the programme owner, which is worth an estimated 20 to 30 hours a month, and the participant gets an answer in seconds rather than at the owner's next free moment.
- Participants and organizers authenticate through different paths and see different things.
- Conversation state lives in a key-value store, so a refused login cannot be talked around.
- Answers are retrieved from the program's own documents, not from the model's memory.
- A scope guard runs before the answering agent, so an off-topic question never reaches it.
- Faster answers are a participant-experience gain as much as a time saving, which is the half nobody counts.
n8n · Redis · Supabase vector store · Google Sheets · OpenRouter
Figures read 2026-09-04.