Sample project for a fictional business, made to show what you get. Make-Believe Escapes is not a real company.
Make-Believe Escapes runs escape rooms in three places: Nowhere Springs, Elsewhere Junction and Someday Hills. Most messages ask the same few things (is there room on Saturday, how young can kids play, where do we park), and group bookings mean checking rooms at all three locations at once. The front desk answered them between games, so replies waited for hours, and big groups booked elsewhere.
Try the sandbox yourself: chat with the agent, see each tool it uses, and approve the bookings it hands to a person. Nothing is booked or sent.
What the agent does
It answers on the website, around the clock, using the business's own information: rooms and capacities at each location, the booking calendar, prices and policies. It has four tools: search the FAQ, check availability across all three locations, draft a quote, and hand a conversation to a person.
The recording shows five real conversations:
- "Room for 14 people this Saturday afternoon?" It checks every location and time, finds two rooms side by side at Nowhere Springs, and gives the total and the deposit rule for groups.
- "How young can kids play?" Straight from the FAQ, including the kids' editions of two rooms.
- "Which location has parking?" All three locations compared in one answer.
- A team-building event for 35: five rooms at once plus the party room, with a draft quote. Groups over 20 need a person's approval, so it goes to the staff queue.
- "My son is sick, can we cancel and get a refund?" Inside 48 hours that's a person's decision, so it says so kindly and hands it over.
A person stays in charge
The agent never promises what the business hasn't agreed to. Big bookings, refunds and anything it isn't sure of land in a staff approval queue with a summary and a draft reply; staff approve with one click. Every answer can be traced to the tool call behind it, which the sandbox shows beside the chat.
What the Done for you tier includes
- Specification and test cases: what the agent may and may not do, written with you, and the conversations it must get right.
- One working agent connected to up to two systems, here the booking calendar and the FAQ.
- Hand-off documentation and a recorded walkthrough, like the one above.
In the sandbox the "understand the question and write the answer" step is a scripted stand-in, so it runs in your browser without an AI model; try the suggested questions. A client build uses an AI model there. The availability search, quotes, FAQ and approval rules are the real logic.
Want an assistant that answers your customers while you work? See AI Agent Development or get a fixed-price estimate.
