Case study · The Fellas Group
Scope confirmedSix AI agents for a food and beverage group in Shanghai, where Western AI cannot go
How we mapped the work of a food and beverage group in Shanghai and scoped six AI agents around the rules of the Chinese platforms it runs on.
The Fellas Group is an Italian-founded food and beverage group on The Bund in Shanghai, with rooftop venues among the city's best known.
The Fellas Group runs rooftop venues that are among the best known in Shanghai, and it has built them on something that is hard to copy: a personal touch. Regular guests are recognised, the best ones are invited back by the founders themselves, and every venue holds the standard the group is known for. Success brings its own weight. Guests book through many channels, every venue publishes on several platforms, and the supplier paperwork grows with every busy night. The team carries all of it, and carries it well, because it matters too much to hand to tools that do not understand how the group works. On top of that, the group operates in one of the most demanding digital environments in the world. WeChat, Douyin, Xiaohongshu and Dianping each set their own rules on what software may do: some are open, some are partly open, and some are closed entirely. So the question was never whether AI could help. It was how to bring it in without touching what makes the group special, and without promising anything the platforms do not allow.
We started with a discovery call with the three principals, on their own operation, in their words: where the guests come from, where the bookings land, where the team's time goes and how long each job takes. Then, before proposing anything, we checked each Chinese platform against its own current documentation: what it lets software read, what it lets software publish, and what it closes entirely. Every opportunity in the blueprint we delivered was tested against that map first.
From the blueprint, the agents were scoped one by one and handed to the client on a page they could mark up line by line: confirm, change the wording, ask a question, or say no. They went through all of it. Their questions were answered, their edits went into the scope, and the six agents below are what they confirmed, in the order they chose.
- Reservation concierge. Booking messages answered at any hour on WeChat and in the group's Mini Program, every booking in front of the staff the moment it lands, and anything nobody confirmed escalated.
- Supplier invoice processor. Invoices and delivery notes read however they arrive, photo, paper, PDF or fapiao, everything that matches posted, and nine hours a day of typing turned into a short queue of exceptions.
- Content library and post packager. Every photo and video filed by venue and date, the post drafted against what is working that week, and published directly on Douyin and the WeChat Official Account.
- Online presence monitor. Every listing checked every week, new reviews read as they land, and a reply drafted to each.
- Paid campaign operator. Campaign drafts built and the spend watched, with every change approved by the marketing manager.
- Operations controller. Every recurring task held in one place, the owner reminded, each one checked as done, and only what slipped reported.
Six agents, each taking one kind of repeating work off the team, scoped with the client and confirmed line by line. Every agent is built so a second venue is a configuration, not a rebuild.
The hardest part of automating a business in China is not the AI. It is knowing which doors are open. WeChat Channels has no way in for any software, at any price. Personal WeChat accounts and groups cannot be automated without risking the account itself. Dianping lets you read a review but not answer it from outside; Douyin lets you do both. And WeChat requires a person to tap send on every outbound broadcast, by design. We did not treat that as a limitation to hide. We built the human tap into the design, the same way we would have chosen to anyway: the machine prepares everything, a person decides what goes out.
This engagement is at the scope stage: the six agents are agreed with the client and nothing has been built yet. This page shows the work that comes before the build, and it grows into the full case study once the agents are live.
Let's talk
Let's find what AI can do for your business.
Every company is different. The best place to start is a chat.