How to Connect Your Restaurant's Data to AI

Everyone is telling you to use AI in your restaurant. Ask it about your sales. Have it write your prep order. Let it find the store that's bleeding labor. Good advice, all of it.

Here's the part nobody mentions: the AI can't see your restaurant.

Type "how were sales last week?" into Claude or ChatGPT right now and it will give you a confident, well-written answer about a restaurant it knows nothing about. It's guessing. It has no idea what your stores did last week, because your numbers aren't anywhere it can reach them. They're locked inside your POS, your scheduler, your inventory platform, your review tools — a dozen systems that don't talk to each other and definitely don't talk to an AI.

So before any of the AI advice works, one boring thing has to happen first: your data has to be connected to the AI. Here's the honest shape of it up front, so nothing in this piece feels like a bait-and-switch: the AI part is genuinely easy now, the part you can do in an afternoon. The connecting part is real work, and it's the part nobody wants to do. This explains what "connecting your data to AI" means, why it's suddenly possible, what you can do once it's done, and what it actually takes to get there.

What does "connect your data to AI" actually mean?

Picture handing someone a locked filing cabinet and asking them a question about what's inside. They can't answer — not because they're not smart, but because they can't open the drawer. That's an AI without your data connected. It's plenty capable; it just can't reach anything real about your business. Connecting your data is opening the drawer.

The thing that opens it has a name: an MCP, short for Model Context Protocol. Skip the acronym. What it is, in plain English, is a secure doorway between your business's numbers and an AI tool like Claude or ChatGPT. With that doorway in place, when you ask the AI "what were net sales last week by location?", it doesn't guess. It reaches through the MCP, reads your actual numbers, and answers from them.

If you want the slightly more technical version, think of an MCP as a USB port for AI. You've plugged a thumb drive into a laptop a hundred times without thinking about how it works — you trust that anything with a USB plug will fit the slot. An MCP is that same idea for data: a standard plug that lets your business's information connect to any AI tool that supports it, instead of someone hand-building a custom cable for your specific systems. One important difference, though, and it's the catch this whole piece comes back to: a thumb drive is ready to plug in the second you buy it; your restaurant's data is not. It has to be cleaned up and lined up first. The plug is standard. Your data isn't — yet. More on that below.

So that's the idea. An MCP is the difference between an AI that can talk about restaurants and one that can talk about yours.

What can you actually do once your data is connected?

Once the connection is live, you stop exporting spreadsheets and start asking questions.

You open Claude or ChatGPT and type the question you'd otherwise hand to an analyst. What were net sales last week by location? Which three stores are running the highest food cost this period? Show me labor as a percent of sales, ranked worst to best. You get a plain-English answer, grounded in your real numbers, in the time it takes to read it.

You can also ask it to build things. Build me the weekly sales spreadsheet by store. It does. Draft the one-page recap for my franchise meeting. It drafts it. No exports, no formulas, no waiting on the one person who knows the reporting tool.

Which of your systems this draws from depends on what you run, and Expo is built to be agnostic to your stack: POS like Toast, Square, and NCR; back-office and accounting like Restaurant365 and CrunchTime; reviews and guest feedback like Yelp, Google, and SMG. If a number lives in a system, the goal is to get it in — Expo has pulled data from places most operators don't expect, down to phone systems and in-store safes. If your particular stack isn't obvious, that's a short conversation, not a dealbreaker.

It's worth being precise about what this does, because the category is full of people promising the moon. Connected to your data, the AI reads and reports — it answers questions and builds the spreadsheet, the ranking, the recap. By design it does not reach back into your POS to change a price or push an order; reading your business and doing an analyst's work is the job, and keeping it read-only is a deliberate safety choice, not a limitation waiting to be lifted. For most operators that read-and-report job is the one that was eating their Sunday nights.

Is it safe to connect AI to my numbers?

This is the right question to ask, and there are really two worries hiding inside it. One: can the AI mess up my data? Two: once the AI can see my numbers, where do they go? Both have good answers.

Can the AI change anything? No. The connection is read-only: the MCP only has permission to read, not write. There is no path through it back into your POS or scheduler to change a price, delete a record, or push an order. The agent can tell you your food cost is up in twelve stores; it cannot touch the systems where that number lives. And the connection is yours — granted to specific people on your team, not an open door anyone can walk through.

Where do my numbers go once the AI sees them? Be clear-eyed about this part, because it's the one that actually matters. When you ask a question, your data leaves your systems and travels to a third party — the AI vendor, Anthropic or OpenAI — so their model can read it and answer. That's not unique to Expo; it's true of any AI tool you connect to your business. So before you point AI at anything sensitive, treat it like any other vendor you send data to, and get a few things right:

• Use a business or enterprise account, not a personal one. This is the single biggest lever. The business tiers (Claude Team and Enterprise, ChatGPT Business and Enterprise) come with a commercial data agreement: by default they do not use your data to train their models, retention is controlled by your admin, and data is encrypted in transit and at rest. The personal plans are a different story — as of late 2025, the consumer versions of Claude (Free, Pro, Max) train on your conversations unless you turn it off, and consumer ChatGPT has long had a similar default. The account you sign up for changes the default, so sign up for the business one.
• Loop in whoever owns your data and vendor agreements. Sending business data to an outside AI vendor can intersect with your franchisor's data rules or your own vendor contracts. It's usually fine on a business plan, but this is a "check with the person who'd know" item, not a thing to wave through.
• Check the data-sharing setting anyway. Even on the right account, open the privacy or data-controls settings once and confirm "improve the model for everyone" (or whatever your tool calls it) is off. It usually is on a business plan, but a 30-second check is cheaper than an assumption.
• Know that retention is a setting too. Business plans generally let an admin control how long conversations are kept. If your data is sensitive, set it deliberately rather than leaving the default.

None of this is restaurant-specific; it's the same hygiene any business should apply before pointing AI at its books. It's also part of what "set it up properly" means, and part of what a partner like Expo helps you get right instead of leaving you to read terms of service on a Sunday night.

So the honest two-line answer: the connection itself can only read, never change, your data; and where your numbers go after that is controlled by your AI account settings, which is why using a business account and checking the sharing toggle matters.

Why can't I just set this up myself?

The honest version is more useful than the sales version, so here it is.

You can't just flip a switch and connect your data to AI, and not because the AI part is hard. The AI part is the easy part now. The hard part is everything underneath it.

Your numbers don't live in one place. They live in your POS, your labor scheduler, your inventory platform, your accounting system, your guest-feedback tool, and those systems were never built to line up with each other. One calls a store "Unit 1842," another calls it "Store 42," a third calls it "Riverside." Before an AI can answer a single question across your portfolio, somebody has to pull all of that together and make it agree. That's the unglamorous, genuinely difficult work, and it has nothing to do with AI at all. It's plumbing — and it's exactly the cleanup the thumb-drive analogy skips over. The plug is standard; your data has to be brought up to code before it fits.

That's the work Expo does. We connect your systems, normalize the data so every store means the same thing across every source, and stand up the MCP on top of it. Then you get the easy part — an AI that knows your business — without ever touching the hard part. Step one for you isn't a download. It's a conversation.

We won't quote a price in a blog post, because it depends on how many systems and stores you're running. But the comparison that matters is the timeline: weeks to get live with a partner doing the plumbing, versus the do-it-yourself build we timed and priced step by step in how to build your own restaurant AI platform in 14 months. The connecting is the project; the AI is what you get once it's done.

What is an AI agent, and how is it different from just connecting data?

Connecting your data is the foundation. An agent is what you build on top of it, and it's where this gets specific to your business.

Once your numbers are connected, the first agent lives inside Claude or ChatGPT and answers questions and builds reports. But the more interesting ones are built for a single job. An agent that drafts Monday's prep order off last week's sales and what's on hand. An agent that watches how each LTO is performing across stores and flags the ones under-attaching it. An agent that assembles the franchise-meeting recap so nobody rebuilds it by hand. You tell Expo what the job is; we build and deploy the agent for it — that's the full Expo picture.

The order matters: connect the data first, then put agents to work on it. Skip the connecting and an agent has nothing real to act on.