Drowning in vague requests that needed several rounds of back-and-forth before work could start.
Situation
Incoming requests from clients were often a sentence long. Staff spent their first hour on each one asking what the client actually meant.
Fully automating the response was not acceptable. A person had to stay in control of what got built and when.
What was built
- Maintained a short, structured map of each client’s applications and domain terms. This map is what makes the AI’s questions relevant rather than generic.
- Added an AI clarifier that reads a new request, asks at most two or three focused questions grounded in that map, and produces a confirmed, structured draft ticket.
- Drafts queue for a human to approve in batches, so people make every decision and the AI only removes the ping-pong.
What held up
- Tickets arrive with the detail needed to start work, and the clarification happens without staff time.
- Clients get faster responses because the first questions come back immediately.
- The human approval step kept trust with both staff and clients.
What I would do differently
- Assign an owner for the domain map from the start and surface “area: other” picks as the staleness signal, rather than discovering the map was behind the product from the questions getting vaguer.
- Add the visible decline reason in the first version. Without it the review gate reads as a black hole.
- Size the model to the job earlier. It is a small-model task with a short context; the first version used a bigger model out of habit.
Stack
Claude API, Supabase, Next.js