#00 — The whole thing in plain words
You're right that the engineer-months are the weakest number in this study. AI writes code now. Fine.
Here's the part that matters: for this product, writing the code was never the slow bit. Below is the whole thing without jargon — what you're building, what's fast, what's slow, and why.
#What you're actually building
Six things. That's it.
| The thing | What it does |
|---|---|
| Till app | Takes the order, prints to the kitchen, takes the card, prints the receipt |
| Kitchen screen | Shows the cooks what to make, in order |
| Back office | Menu, prices, staff, reports. Boring, necessary |
| Customer database | Fills itself from the till — who came, what they ate, how often |
| Marketing | Text/email those people, and prove it brought money back |
| The boxes | Tablet, receipt printer, kitchen printer, card reader, cash drawer, router |
The first three are a POS. The middle two are the CRM. The last one is a shopping list.
#The software: what's quick, what isn't
Quick. Genuinely quick — you're right about this:
- Menus, prices, back office, reports, dashboards → forms and tables. AI eats this.
- Customer database, segments, campaigns, sending emails and texts → fast.
- Ordering website, QR menu, loyalty page → fast.
- A demo that looks like a real till → a few days.
Slow. And AI barely moves the needle:
- Working when the internet dies. Not hard to type. Hard to get right. A restaurant's wifi drops mid-order, two staff touch the same bill on two tablets, the router reboots at 19:40 on a Friday. The code for that takes a week. Finding out yours is wrong takes six months, because the bug shows up as "my Saturday total is 340 CZK short" — a phone call, not a stack trace.
- Printing exactly once. Sounds trivial. It is not. Print twice and the kitchen makes two steaks and someone eats your margin. Print zero times and a table waits 40 minutes. Every reconnect, every crash, every swapped printer is a chance to do one or the other.
- Money maths. VAT, rounding, splitting a bill four ways, refunding half of it tomorrow, a discount applied before a service charge instead of after. Every one of these bugs is somebody's actual money, and you find them one angry customer at a time.
The pattern, in one line: AI makes the first 80% take a weekend and the last 20% take a year — and in a till, the last 20% is the product. Nobody churns because your dashboard is ugly. They churn because the bill was wrong once.
#The hardware: the easy part, honestly
Don't build hardware. Buy it, put your sticker on the box, resell it at cost +15%.
| Restaurant | Kit | Cost |
|---|---|---|
| Café, one till | Android till, card reader, drawer, router, UPS | ~€1,400–1,800 |
| Bistro, 2 tills + kitchen | Above ×2, plus kitchen printer, bar printer, kitchen screen | ~€4,900–5,800 |
Rules that matter more than the brand names: kitchen printers must be impact, not thermal (heat and steam turn thermal receipts blank), anything that prints gets a cable (wifi printers are a support ticket generator), and the router is part of the install, not a suggestion.
Getting one printer working: two weeks. Getting printers working in 50 restaurants you can't drive to, on their wifi, behind their landlord's router: two months and it never fully ends. "The printer is offline" will be your most common support ticket forever.
#What's actually slow — and notice that none of it is code
This is the real answer to "we can vibe code it."
| The slow thing | How long | Can AI speed it up? |
|---|---|---|
| Government permission to print a legal receipt | Czechia: none needed. Poland: 6–12 months and a certification body. Spain: don't, they fine the software maker | No. It's a queue with a civil servant at the end |
| A bank agreeing to pay you a cut of card payments | Weeks to months of meetings. This is where the actual money is — bigger than subscriptions | No. It's a negotiation |
| A lawyer confirming you can text your customers | 2–6 weeks, ~€3–5k | No |
| Putting the boxes in a restaurant | 5–8 hours, two people, on site, plus the wiring nobody budgeted | No. Someone drives there |
| Typing their menu in | 20–50 hours per restaurant. AI reading a menu photo cuts it ~40%, not 90% — menus are chaos and a wrong price is real money | A bit |
| Answering the phone at 23:00 on Saturday | Forever. Restaurants work when you don't | A bit, later |
| A restaurant trusting a 3-month-old company with their till | Months. They're not switching from a working system for a stranger | No |
Roughly two-thirds of the effort in this business is not typing. AI compresses the typing. It compresses the rest by zero.
#So how long, really?
Taking your point seriously — assume AI is as good as you say:
| Milestone | Honest estimate |
|---|---|
| A demo that looks real | 2–4 weeks |
| One friendly café runs it for a week, you standing there | 2–3 months |
| A café runs it unattended and you'd take their money without wincing | 6–9 months |
| You can sell it to strangers in one country and survive supporting them | 12–18 months |
The study says month 31 for that last one. If AI is as strong as you think, call it 12–18 months and I won't argue.
But notice what didn't change. The order of operations is identical. The certification queue is identical. The install day is identical. Saturday night is identical. AI moved the calendar. It didn't move the business.
And the study's own rule already covers this both ways: if the AI speedup is worse than assumed, redo the schedule but don't touch the revenue forecast. Same if it's better. Shipping sooner doesn't make restaurants pay more or churn less.
#The two things that actually decide this
Before any of the above matters, two questions. Four months, under €40k, zero lines of code:
- Will restaurants pay ~€89 a month? Not "would you be interested" — get 8 of 40 to put down a €200 refundable deposit. Opinions are free; deposits aren't.
- Will a bank put ≥25 bps of card revenue in writing? Not a friendly maybe. A term sheet.
If both are yes: build it, and vibe code as hard as you like. If either is no: it doesn't matter how fast you can code. A great till nobody pays for is still nothing, and subscriptions alone don't cover the cost of supporting a restaurant.
#What I'd do next week
- Book 40 coffee meetings with Prague and Brno café owners. Ask what they use now and what they hate.
- Email Viva.com, Nexi and Worldline asking for ISV partner terms. Get a number in writing.
- Buy one Sunmi till, one Epson printer and one Stripe reader — about €1,200. Have them printing by Friday.
- Ask a Czech lawyer one question: can a receipt carry a "join our club" line?
- Then, and only then, start building — and start with the till, offline, because that's the part that takes real time.
The rest of this study is the map, not the schedule. The numbers worth arguing with are in 07 — Roadmap, team and budget. The reasons things break are in 08 — Pros, cons, difficulties and risks. The actual decision is in 01 — Feasibility verdict.