The brief
Four studios, three inboxes each, and a front desk that only exists during class hours. Enquiries about class times and trial pricing landed at 10pm and sat until the morning, by which point the person had usually joined whichever studio replied first.
What they bought
Votel AI Agent Setup
Conversation design, knowledge base, calendar booking, escalation rules, pre-launch testing, plus the monthly maintenance engagement
View the full scopeIntegration Specialist
Instagram DM, WhatsApp, and web chat routed into one agent; four studio timetables connected
View the full scopeWhat we built
- 01
Scoped the agent tightly
Three jobs only: answer class and pricing-range questions, qualify, book a trial. Memberships, refunds, and injuries were out of scope from day one.
- 02
Built the knowledge it answers from
The class list, published price ranges, timetable, and cancellation policy — rewritten into answerable form rather than dumped in raw. Anything not in there, it says it does not know.
- 03
Connected the four timetables
Live availability per studio, so it offers real slots and writes the booking into the right diary, then handles the confirmation and reminder itself.
- 04
Tried to break it before launch
Replayed a month of real DMs plus pricing pressure, off-topic messages, and prompt-injection attempts. Fixed the wrong answers, then got sign-off from the agency and each studio.
What they got
- One AI agent live on three channels, scoped to three jobs: answer, qualify, book
- A knowledge base built from the studios' real class list, pricing, and cancellation policy — documented and editable
- Trial classes written straight into the correct studio timetable, with confirmations and reminders
- A pre-launch test log: 60 replayed enquiries, including the awkward ones, and what we changed after each
- A monthly transcript review with a one-page report on what it handled and what it passed on
What their client sees
- An answer within a minute at 11pm, in their studio's voice
- Trial classes appearing in the timetable overnight
- Anything about injuries, refunds, or memberships going to a named human, with the transcript attached
What we'd tell you on a call
A narrow agent that books is worth more than a broad one that chats. Almost every bad AI deployment we are asked to fix was scoped too wide, not modelled too cheaply.
How we document projects: white-label engagements are confidential (our name never appears on the client side), so each write-up covers the scope, the sequence, and the result by vertical rather than naming the agency behind it. Screenshots are taken from the real build and redacted accordingly.