The Setup
This hotpot brand runs 18 stores in a tier-3 city in eastern China, average ticket 90 to 120 yuan, targeting community and family diners.
The owner, Mr. Chen, has been in F&B for 12 years. By late 2025 he was anxious — new hotpot competitors kept opening, price wars were unsustainable, repeat purchase was stuck at 22%, and ingredient waste sat at 8%.
He found a local AI vendor, spent 150,000 yuan, and deployed three things: AI member profiling, smart ordering recommendations, and ingredient waste prediction. The stack runs on Doubao LLM plus Volcano Engine HiAgent.
What They Did
1. AI Member Profiling: Flavor Tags for 200K Members
Previously the member system only had phone numbers and spend totals. Now AI analyzes order history and auto-tags: spice-lover, mild, adventurous eater, family gathering, date night, beef-favoring or lamb-favoring.
Push campaigns stopped being blasted to everyone. Instead they match flavor profiles — spice lovers get the new numbing broth, mild folks get tomato or mushroom, adventurous ones get limited drops.
2. Smart Ordering: Broth Plus Dish Combos
The ordering iPad now shows AI recommendations based on member tags and order history. A 4-person family gathering gets split pot plus a kids combo and two signature beef plates.
Basically, the veteran server's instincts got digitized. New staff skip the menu memorization — AI hands them the plan.
3. Ingredient Waste Prediction: How Much to Buy Daily
AI factors historical sales, weather, holidays, and store location to predict each store's daily procurement. Store managers used to eyeball it; now AI suggests, managers review and adjust.
Two Pits
These two nearly killed the project, but they're also what made it work in the end.
Pit One: Cold Start Pushed Spicy Broth to a Non-Spice Eater
Week one, the system defaulted every new member to the best-selling numbing broth. A family visiting from the north, explicitly non-spicy, got pushed the spicy pot. The kid tried it and cried on the spot.
Bad review came in that day. The problem? New members had no order history, so AI had no flavor data and fell back to store-wide defaults.
The fix: a 15-second flavor quiz on first scan-to-order — pick 3 preferences, and only then does AI have something to work with. Quiz completion hit 68%. Most people don't mind.
Pit Two: Forecast Ignored Holidays and Weather
Mother's Day, pouring rain. AI predicted beef procurement at 80% of normal. Foot traffic spiked 40% instead. Beef ran out by 2pm. A hotpot restaurant with no beef is basically closed.
Root cause — the model's training data had no Mother's Day label, and weather was connected but weighted too low. They re-tuned holiday weights and added a real-time weather alert that auto-bumps procurement 15% on storm days.
The Results
Six months in, the numbers:
- Member repeat purchase: 22% → 38%
- Ingredient waste: 8% → 3.5%
- Front-of-house labor: down 25% (AI recommendations replaced part of manual upselling)
- New-item recommendation conversion: 12% → 27%
- 150K yuan invested, payback in 5 months
The Real Talk
Chen's verdict: AI didn't make the hotpot tastier, but it made the business smarter.
The core here isn't model strength — it's feeding the right data. Hotpot is a business where every variable bites: flavor, weather, holidays, neighborhood radius. AI can crunch those for you, but only if you've already mapped your own business logic. Don't expect AI to think for you.
One more thing: for tier-3 city restaurant AI, don't jump straight to "agents." Nail the basics first — member tags, recommendations, procurement. Small spend, visible return.