This is a real recipe. We know, because we cooked it — Expo is what came out of our kitchen, and several of the burns below have our names on them. If you've ever sat in a meeting and said "we could just build this," here is what just means. Every step lists its time and its cost — split into labor and materials, the way you already read every cost in your life.

  1. Hire the data engineers. Time: 3–4 months. Labor: $0 — enjoy it, that's the last time you'll see that number. Materials: $90K in recruiter fees. Months of recruiting before anything else exists. You will read résumés full of words — Kafka, Snowflake, Airflow — and you will nod the way you nod at a sommelier. The good candidates have never heard of your restaurant group. The ones who apply ask what a comp is. Interviews where neither side understands the other, an offer, a counter-offer, a start date six weeks out. Four months gone. Nothing built.
  2. Set up the infrastructure. Time: 1 month. Labor: $51K. Materials: $4K of cloud. Before anyone can touch your numbers, the new hires need somewhere to put them. Which cloud, which database, which dozen tools with names like Norse gods. Every choice is a meeting, and every meeting breeds two more. This is the quiet stretch where the board asks how the AI project is going, and the honest answer is: we picked a database.
  3. Beg your vendors for access. Time: 6 weeks of waiting, alongside step 4. Labor: counted in step 4. Materials: ~$10K in access fees — yes, fees. Your numbers live inside other companies' software, and other companies are in no hurry. The keys take weeks. Some vendors charge rent — you will pay money to look at your own sales. Some systems have no door at all, just a nightly email with a spreadsheet attached. Franchisee? Add a formal request to the franchisor for permission to look at your own stores.
  4. Build the pipes. Time: 2 months. Labor: $102K. Materials: $8K of cloud. Now the engineers build the plumbing that pulls numbers out of the register, the schedule, the inventory counts, payroll, and the guest complaints, and pours them into that database from step 2. Kill one assumption before it costs you: everything has an API — a proper tap your engineers just connect to. No. A few of your systems have a real one; those connect like a garden hose. Several have half of one — it offers some numbers, just not the ones you actually want. The rest have no tap at all. For those, your team builds custom extraction tools — one per system, each its own little engineering project, each fragile in its own special way, each one breaking the moment the vendor changes anything. And the oldest system just emails a spreadsheet, column names drifting without warning, until your $235K engineer spends a week finding out why Tuesday is missing.
  5. Load the history. Time: 1 month. Labor: $51K. Materials: $4K. A dashboard with three weeks of data can't answer one question worth asking. Trends need years, so you load two or three of them — and discover that everything from before your last register switchover is shaped differently. Your engineers will call this a schema migration. You will call it February. You now maintain two versions of the past.
  6. Make the numbers agree. Time: 2 months. Labor: $102K. Materials: conference-room coffee. Everyone thinks the pipes were the hard part. No. The hard part is that "net sales" means three different things in three different systems, and Store 4 is also "0004" and also "Lubbock II." You will spend months in a conference room with accounting, deciding what counts as a sale. These are the most important months of the entire project, which is exactly why everyone skips them.
  7. Prove it matches the books. Time: 6–8 weeks, during which all nine weeks of PTO land. Labor: $90K. Materials: $0 — the spreadsheet is free; the trust is not. You run the new system next to the old spreadsheet, and they disagree. Every week, a little differently. Until your controller blesses the new number, the whole company keeps using the old spreadsheet — the one you are spending three-quarters of a million dollars to kill. Trust arrives one reconciled month at a time.
  8. Decide who sees what. Time: 3 weeks, alongside step 7. Labor: $35K. Materials: $0. The GM sees her store. The district manager sees his eight. The franchise partner sees their entities and nothing else, and payroll is visible to almost no one. Congratulations: your team is now building login screens and permission rules — the least rewarding work in all of software, and every bit of it load-bearing.
  9. Make it readable by AI. Time: 3 months. Labor: $153K. Materials: $2K in AI usage credits. The step nobody tells you about. The honest split: plugging an AI into a clean warehouse takes days. What takes months is making the data mean something to it — the dictionary that says what every number is, which definition of net sales is the real one, what "Store 4" refers to, what a manager is allowed to ask about. The trade calls the result agent-readable, and without it the AI answers fast and wrong. Skip this step and you've built the world's most expensive PDF. (This layer is the whole reason our product exists.)
  10. Teach the robot not to lie. Time: 2 weeks to build, forever to run. Labor: $25K to start. Materials: a few hundred dollars of wrong answers. Wiring AI to your data is one job. Making sure it doesn't confidently report the wrong store's labor number is another. So you build a test kitchen for the machine: a hundred questions you already know the answers to, run again every time the AI updates — which is every few months, forever. Skip this step, and one day your CFO quotes a hallucination to the board.
  11. Design. Time: 2 weeks. Labor: $25K. Materials: $0 — you already owned the whiteboard. Your engineers will now design the screens. This is like asking your cooks to decorate your dining room. Brace yourself.
  12. Product. Time: your nights and weekends, months 5 through 12. Labor: $0 on the books — this one costs you. Materials: $0. Someone has to decide what this thing actually does — which numbers greet you in the morning, which problems trigger a text, what the GM sees first on her phone. That someone is you, in your fifth meeting of the week, working a second job you never applied for.
  13. Launch! 🎉 Time: one afternoon. Labor: a toast. Materials: $200 of cake. Cake in the break room. Someone gives a toast. Enjoy it — because everything until now was the prep list. Service starts tomorrow, and service never ends.
  14. Watch it daily. Time: daily, forever. Labor: it's in the payroll, forever. Materials: $2–4K a month of cloud creep, rising. The overnight data runs break at 3am. Vendors change their software without telling anyone. The Tuesday numbers look wrong because they are wrong. Also: this is the month the cloud bill stops resembling the estimate from step 2 and starts resembling a car payment. For a fleet.
  15. Feed the queue. Time: weekly, forever. Labor: your roadmap. Materials: $0. Sally in accounting wants a new report. The CEO wants a new view. The CFO — who approved all of this — wants one more integration. Your two engineers' grand roadmap is now a ticket queue. Innovation is something that happens at other companies.
  16. Month 14: you're live, and you're already legacy. Time: the rest of year one. Labor: about $640K. Materials: about $110K. Total: $750K, and the meter does not stop. Now take stock of what three-quarters of a million dollars bought: your data, connected, trusted, answerable. Credit where due — that was the hard part; most builds die before reaching it. But look at what kind of project this turned out to be. You thought you were buying AI. Fourteen of the sixteen steps were data work. The thing your board asked for — an agent that does real work: watching every store overnight, drafting the Monday emails, catching the variance before anyone asks — that project starts now, on top of everything you just built. And read the check the way you'd read any P&L: this dish ran 85% labor. You'd fire a GM who brought you those numbers. Meanwhile AI moved two generations while your team was heads-down; the tools from step 2 are last year's tools, and the rebuild conversations have already started at lunch. Then one of your two engineers takes a job at an actual tech company, and everything he knew that was never written down leaves with him in a cardboard box.

Then you hire his replacement. The new engineer surveys fourteen months of work and — with the serene confidence of week two — recommends rebuilding the whole thing on a newer stack.

Return to step 1.

That's the recipe. Now go back to that meeting where somebody said "we could just build this," and notice what nobody suggested. Nobody proposed writing your own point-of-sale system. Nobody wanted to hand-roll payroll, or vibe-code an accounting package over a long weekend. That would be insane — there are entire companies for those things, and everyone in the room knew it. But a data and AI platform? That one looks easy from a distance, the way line cooking looks easy from the dining room. It is the same decision in a better disguise: a restaurant company electing to become a software company. And understand what that means. You already run one of the most painful businesses ever invented — you know it in your knees, your margins, and every walk-in compressor that died on a holiday weekend. Nobody tells you the second half: software is exactly as painful. It just hurts in a language you don't speak yet. Build it yourself and you don't trade one pain for the other — you sign up for both. At the same time. Forever.

If you read all sixteen steps and still want to build — sincerely, respect. You're probably 500+ units with a board mandate to own your stack, and steps 6 and 9 are where to overspend.

Everyone else: you already know how to make the build-vs-buy call, because you make it every week in your own walk-in. You buy garlic pre-peeled and onions pre-diced — not because your cooks can't peel garlic, but because labor spent on prep that doesn't make you distinct is labor stolen from the cooking that does. Your data platform is peeled garlic. It has to be excellent, and it will never be the reason a guest chooses your restaurant.

Stock is a different ingredient. Some kitchens are right to make their own stock, because stock can be the signature. In this kitchen, the stock is the intelligence: the questions you ask, the calls you make, the agents you put to work on your numbers. That part should be yours — nobody is suggesting you outsource your judgment. They're suggesting you stop peeling garlic.

Buy the garlic peeled. Make your own stock. The entire point of Expo is the peeling — your data connected, normalized, readable by AI, somebody else's 3am problem — so everything you build on top of it is yours. It's about a three-week recipe, and what it should cost is in the analytics buyer's guide.