The Problem: Too Many Steps = Nobody Does It
Before the bot, logging inventory meant:
- Open Airtable → find the right table → fill the form
- Select ingredient type → enter weight → enter bag count
- Review → Submit
Sounds manageable. But in reality, staff receive deliveries at 6 AM surrounded by boxes of meat — nobody wants to fill out forms.
A Hard Truth
The best system is the one people actually use. If staff won't use it, it doesn't matter how good the system is.
The Solution: A Chat Bot That Understands Human Language
Instead of forms, I built a bot that accepts plain chat:
Staff: received sirloin 5 bags 12.5 kg
Bot: ✅ Logged
📦 Sirloin — 5 bags, 12.5 kg
🏷️ Bag ID: BEEF000758 - BEEF000762
📊 Total stock: 23 bags (48.7 kg)
What the Bot Does Behind the Scenes
📝 Receives message from staff
↓
🧠 Parses: ingredient name + bag count + weight
↓
🔍 Validates against ingredient database (273 items)
↓
🏷️ Auto-generates Bag IDs (1 per bag)
↓
💾 Saves to database
↓
📊 Updates real-time stock levels
↓
✅ Sends confirmation with summary
From 5 Minutes → 10 Seconds
| Step | Before (Form) | After (Bot) |
|---|---|---|
| Open app | 15 sec | 3 sec (open chat) |
| Enter data | 2-3 min | 5 sec (type 1 line) |
| Review | 30 sec | Bot validates |
| Submit | 15 sec | Automatic |
| Total | 3-5 min | 10 seconds |
Multiply by 40 entries per day = 2-3 hours saved daily
What Makes the Bot Actually Usable
1. Natural Language Understanding
No rigid format required. Staff can type multiple ways:
- "received sirloin 5 bags 12.5 kg"
- "sirloin 5 bags 12.5"
- "sirloin 5/12.5"
2. Asks When Unsure
Staff: received beef 5 bags
Bot: 🤔 Which type of beef?
1. Sirloin
2. Tenderloin
3. Flank steak
Staff: 1
Bot: ✅ What's the total weight?
3. Automatic Validation
- Single bag over 5 kg → asks for confirmation
- Ingredient not in system → alerts + offers to add
- Weight doesn't make sense → asks again
Tech Stack
| Component | Tool | Why |
|---|---|---|
| Chat Interface | Telegram Bot API | Staff already use it |
| AI Processing | Claude API | Understands Thai, accurate parsing |
| Database | Airtable + SQLite | 2-layer (frontend + backend) |
| Hosting | Agency server | Near-zero cost |
Lessons Learned
UX Beats Features
The old system had every feature. Nobody used it. The bot has fewer features but is so easy there's no excuse not to use it.
Start with One Flow
Don't build a bot that does everything. Start with just "receiving" — once people are comfortable, add "withdrawals" and "transfers."
Bot Data > Form Data
Because staff log more frequently (every item vs. some items) → data is more complete → analysis is more accurate.
Results
Stock entries went from ~50/month to 757/month — not because we received more, but because staff actually logged every single item.
Can You Build This?
Yes, if you have:
- A Telegram account — create a bot via @BotFather
- An ingredient list — in Google Sheets or Airtable
- AI — Claude/ChatGPT to parse messages
Start with a simple prototype: bot receives message → sends to Google Sheet. Upgrade later.
