An outdoor gear company was growing fast. Too fast. Their support team couldn't keep up. But every chatbot solution they looked at wanted to store customer data on third-party servers. That felt wrong. So they built something different.
The Problem Was Simple
You're doing 400 orders a day. Your support inbox is overflowing. You need help. The options:
- Hire three more support staff: $150,000+ per year
- Buy enterprise chatbot software: $6,000 to $12,000 per year
Cost aside, there was a bigger issue. Every platform wanted customer data. Names, addresses, order history. All sitting on someone else's servers.
Data breaches cost companies $4.44 million on average. GDPR fines can hit €20 million. And more importantly, customers trust you with their information. Handing it to a third party just felt wrong.
They needed something that worked without the risk.
How It Works
Most chatbots work like this: send everything to a server, let the AI process it, send back a personalized response. Simple, but it means your customer data lives on someone else's infrastructure.
This system flips that around. When someone asks "Where's my order?", here's what happens:
- The browser grabs what it needs. Customer info from Shopify sits in browser memory temporarily. Just long enough to personalize the response. Then it's gone.
- Only an ID goes to the server. Not a name. Not an address. Just a customer ID.
- The server asks Shopify for order data only. Order numbers. Shipping status. Dates. No personal details.
- The AI writes a response with blanks. Something like: "Your order {ORDER_NUMBER} is shipped and will arrive at {SHIPPING_ADDRESS} by {DELIVERY_DATE}."
- The browser fills in the blanks. Only in the customer's browser does the system swap those placeholders for real info: "Your order #1432 is shipped and will arrive at 742 Evergreen Terrace by Friday."
The personal information never leaves the browser. The AI never sees it. OpenAI never sees it. It's not stored anywhere.
This isn't a privacy policy. It's how the system is built. The code doesn't send personal data to the server because it's written not to. Can't leak what isn't there.
What It Costs
Most platforms charge per seat or per conversation. The bill grows as you grow. With serverless, you pay for what you actually use.
At 30,000 conversations per month, the cost is:
- Compute and API calls: $3
- Search infrastructure: $26
- Storage: $1
- AI model: $2
Total: $32/month
That's for everything. Order tracking, product search, personalized recommendations, the works. Traditional platforms would charge $500 to $1,000 for the same volume.
Double your conversations? You'd pay maybe $60 or $70. Traditional platforms would jump to $1,000 or more.
Why so cheap?
- GPT-4o-mini costs 99% less than GPT-4
- No over-provisioned servers sitting idle
- You only pay for what you use
Security
Privacy keeps data safe. Security keeps the whole system safe.
Every request from Shopify comes with a signature. It's proof the request actually came from Shopify. The system checks that signature. Wrong signature? Request gets rejected.
Each request also has a timestamp. Anything older than five minutes is rejected. So even if someone intercepts a valid request, they can't reuse it.
On top of that:
- All traffic is encrypted
- Database is encrypted
- API keys are encrypted
- Each component only gets access to what it needs
When you're not storing much personal data, security gets simpler.
Better Search
Most e-commerce search is bad. You type "comfortable winter gear" and get nothing unless a product uses those exact words. Type "warm jacket" and you get different results than "winter coat."
This system understands what you mean, not just what you type.
Ask "What should I wear for winter hiking in Colorado?" and it knows you need cold-weather gear for high altitude. It suggests insulated layers that breathe.
The search adapts based on what you're asking:
- "What's trending" shows what's actually selling
- "What's new" shows recent arrivals
- Personal recommendations look at what you bought before
Buy hiking boots and a rain jacket? You'll see suggestions for trail maps, moisture-wicking socks, waterproof pack covers. The kind of thing a good salesperson would recommend.
Customers find what they want on the first try. No more guessing the right keywords.
What Gets Automated
This handles the repetitive stuff that eats up hundreds of support hours:
- Order tracking. 40 to 60% of support tickets are just "Where's my order?" Now that's automatic.
- Product search. Customers find what they need without opening a ticket.
- Recommendations. No one needs to manually review purchase history to suggest related products.
- Follow-up questions. The system remembers the conversation. Customers don't repeat themselves.
- Login detection. Features unlock automatically based on whether you're logged in.
Support teams can focus on the complicated issues that actually need a human. Everything else just works.