Recently, a small business owner walked me through a system they had built with ChatGPT. It had a public website, a portal for their customers and a portal for their own staff. The screens were polished, the buttons worked and every table was full of data.
Then came the question: what would it take to put this live?
My first instinct as an engineer was to estimate a proper build, screen by screen. The number came out far beyond what a small business could pay. That gap told me I was answering the wrong question.
§1The demo had already done the hardest part
Clicking through their prototype taught me more about their business in an hour than a week of requirement meetings usually does. I could see how an order moves, which fields matter to them, and where the work passes from one person to the next.
The design was done. The workflow was clear. The scope was visible on screen.
So my job was not to build that again. My job was to make it real.
§2What "go live" actually means
The prototype was built to be shown, not to be run. Behind the screens, the things a real business depends on were missing:
- logins for each customer company and roles for staff, instead of demo accounts printed on the login page;
- emails that are actually sent, from the company's own domain;
- a real database with backups, instead of sample data;
- somewhere safe to keep invoices, shipping documents and payment slips;
- calculations someone has checked, such as balances, stock allocation and overdue amounts;
- hosting that the company owns, instead of the AI tool's own.
None of this shows on screen, which is exactly why it is so easy to underestimate, for the client and for me.
§3Don't price what you haven't seen
How much of the prototype can be reused decides almost everything about the cost. And that cannot be judged from the screens.
There are roughly three outcomes once you see the code:
- It exports cleanly. Harden it, connect real services and deploy it.
- It exports partly. Keep the screens and rebuild what sits behind them.
- It barely exports. Use the prototype as a detailed specification and build it properly.
Each outcome is a different price. Quote before seeing the code, and you have to assume the worst. So I asked for a full export, including a list of anything that only works on the AI tool's own hosting. That list tells me more than the code itself.
§4Ask before you assume
Before talking about any number, I sent a list of questions. What must work on day one? Who will use it, and how many orders do they expect? How is the work done today? When do they want to go live? Would they want a monthly support plan afterwards?
Their answers changed my plan. A part I had quietly moved to a later phase turned out to sit right in the middle of their daily work. Another part I thought was essential could wait.
I would have quoted the wrong project if I had not asked.
§5The scope is the real product
A prototype tends to contain everything the owner has ever imagined. Building all of it at once is how these projects go over budget.
What I am really delivering is an agreed scope:
- Phase 1 is the core loop that customers, money or compliance depend on.
- Phase 2 is everything that can wait, listed and acknowledged.
- The scope is frozen to what the prototype does. New ideas go on the Phase 2 list.
- Calculations are written down before I build them. A formula that is still being "fine tuned" is the most expensive kind of change.
§6Small decisions that protect the client later
Some of the most useful advice cost nothing to give:
- Open the hosting, database, email and domain accounts in the company's name, and invite me as a member. No passwords are shared, and if they change developer one day, everything stays with them.
- Check the plan terms. Some free hosting plans are for personal use only, so a business system needs a paid plan.
- Tell them the monthly platform bill up front, separate from my fees.
- Keep two environments: one for testing and acceptance, one for production. They can share a database service with separate databases, which keeps the cost down without mixing test data into live data.
§7Working alone, with AI
I plan to build this on my own, with AI as my pair. That changes the economics. Reading an exported codebase, hardening it, wiring up services and writing tests are exactly the jobs AI speeds up most.
But AI does not shorten everything. Client calls, acceptance testing, loading their real data and support after launch still take my own hours. And when one person writes the code with AI, a second pair of eyes on security matters more, not less.
§8Closing thoughts
A client with an AI prototype is not asking me to build their idea. They have already built the idea. They are asking me to make it something they can depend on.
That is a different job, and it needs a different estimate.
§9Before I quote the next AI prototype, ask myself
- Have I clicked through the whole prototype and listed everything it does?
- Have I seen the code export, including what cannot be exported?
- Have I asked what must work on day one, instead of assuming?
- Is the scope split into phases and frozen to what the prototype does?
- Are the key calculations written down and agreed?
- Will every account belong to the client, with me invited as a member?
- Is there a testing environment and a plan for acceptance testing?
- Have I counted the hours AI cannot save me?