Can I build my MVP with AI tools?
By Bonaventure Ogeto, Founder and CEO, Mctaba Labs
Published
Yes, for a simple first version. AI tools can now produce a working web app from plain English, enough to put in front of customers and test demand. They do not make it safe or correct for you. Payments, private data and anything that breaks in ways you cannot explain still need a person who can read the code.
On a Saturday afternoon you type a paragraph into an AI app builder: a page where customers order a cake, choose a collection date and leave a phone number. A few minutes later there is a working page on your screen, with a form, a list of orders and a pleasant colour scheme. It feels like the developer you could not afford just turned up for free.
Then you ask it to add M-Pesa, and things get strange. The button appears but nothing arrives. You ask it to fix that, and the order list stops loading. Three hours later you are pasting error messages you do not understand into a chat box and hoping.
Both halves of that afternoon are true, and together they are the answer. You can build an MVP with AI tools. You need to know which parts they do well, which parts they fake, and what to check before a stranger trusts the result.
What kinds of AI tools can build software?
Three families, from least to most technical.
Chat assistants such as ChatGPT, Claude and Gemini. You describe what you want and they write code and explain it. You have to put that code somewhere yourself, so these are most useful for planning, for explaining errors and for learning.
AI app builders such as Lovable, Bolt, v0 and Replit. You describe the product in a chat box and the tool builds a running app in your browser, which you can keep changing by talking to it. Many can connect a database and publish the result to a web address. This is where most non-technical founders should begin.
AI coding tools for developers such as Cursor, GitHub Copilot and Claude Code. These work inside a real code project on a computer. They are the most capable and assume you can find your way around files, a terminal and error messages.
These products change quickly, so treat any description of one as a snapshot and try the current version yourself.
What can I realistically build this way?
A first version with one flow. Good candidates:
- A page that explains an offer and takes orders, bookings or sign-ups.
- A simple tool with login, where each user sees their own records.
- A catalogue or directory with search.
- An internal tool for your own business: orders, stock, members, deliveries.
- A clickable version of a bigger idea, to show customers before you pay for the real thing.
Poor candidates for a first attempt: anything that holds a balance or pays people out, anything with several kinds of user who must not see each other's data, USSD services, and apps for the Android and iPhone stores. The tools can attempt all of these. You will not be able to tell whether the result is right.
Where do AI-built products go wrong?
In the places you cannot see from the screen.
Private data left open. The app looks fine, but the database behind it lets anyone read everyone's records. This is a known trap. Supabase, a database many AI builders connect to, says in its Row Level Security documentation that a table in an exposed schema without that protection is readable and writable by any role with a grant on it, and tells you to enable it on every such table. An AI tool may or may not do that for you. You have to check.
Secret keys in the wrong place. Payment and messaging providers give you secret keys. If the AI pastes one into code that runs in the customer's browser, anyone can copy it. Keys belong in the hosting platform's protected settings. Vercel's documentation, for example, describes a Secret type for passwords, API keys and tokens.
Payments that only look finished. An M-Pesa payment is not complete when the customer taps "pay". The provider reports the result to your app afterwards, and your app must record it correctly even if the customer has closed the page. AI tools often build the button and skip or fumble that second half.
The fix loop. You report a bug, the tool changes five things, two of them break something else. Without the ability to read the code you cannot tell it where to look, so each round makes the project a little more tangled.
Confident wrongness. The tool will tell you something is done and secure when it is not. It is not lying. It has no way to know, and neither do you unless you test.
None of this is a reason to stay away. It is a reason to keep the first version small and to test like a suspicious customer.
How do I build an MVP with AI, step by step?
Work in short sessions, one piece at a time. Here is the plan for the cake order page from the opening, which is a made-up example. Change the words to fit your product.
Session 1. Describe the product before you build it. In a chat assistant, not the builder:
I am a home baker in Nakuru. I want a simple web page where a customer chooses one of five cakes, picks a collection date at least three days ahead, enters their name and phone number, and pays a deposit by M-Pesa. I need a private page where I see all orders and mark them as collected. I am not a developer. Ask me questions until the requirements are clear, then write them as a short numbered list I can give to an AI app builder.
Answer its questions. You now have a specification, which is the thing most AI projects lack.
Session 2. Build the order flow only. Paste the list into an app builder with this instruction on top:
Build only items 1 to 3 for now: the cake list, the date picker and the order form. Do not add payments or login yet. Keep the design plain and make it work well on a small Android phone.
Open the link on your own phone. Place three orders. Try to break it: a date in the past, an empty phone number.
Session 3. Add your private page.
Add a login for the owner only. After login, show all orders, newest first, with a button to mark each as collected. Customers must not be able to open this page or see other customers' orders. Explain to me how you have protected the order data.
Then test the claim. Open the site in a private browser window, without logging in, and try to reach the orders page.
Session 4. Add payment the simple way. Do not ask the AI to write an M-Pesa connection from nothing. Open an account with a payment provider that supports M-Pesa (Paystack, for one, lists it for businesses in Kenya) and ask:
Use this provider's hosted checkout to collect a deposit of KES 500 after the order form. Mark the order as paid only when the provider confirms the payment to the server, not when the customer returns to the page. Tell me exactly where to put my secret key so it is never visible in the browser.
Pay the deposit yourself with real money. Then do it again and close the page straight after paying. Does the order still show as paid?
Session 5. Save, publish and stop. Export the code or connect the project to your own GitHub account if the tool allows it. Publish to a web address, ideally your own domain. Then stop adding features and send the link to ten people who have never seen it.
Two habits make all five sessions go better. Save a version before every change, so you can go back. And when something breaks, paste the exact error and say "explain what this means in plain English before you change anything".
What must I check before real customers use it?
Go through this list yourself. If you cannot say yes to a line, do not launch that part.
- I placed an order from a second phone, and it could not see the first phone's order.
- Without logging in, I cannot open the owner's page, even by typing its address.
- I asked the tool to list every table in the database and confirm which protections are on, and I looked at the database settings myself.
- I searched the code for my secret key and it appears only in the hosting settings.
- A real M-Pesa payment was recorded once, not twice and not zero times, including when I closed the page early.
- A failed or cancelled payment leaves the order marked unpaid.
- The page loads on a cheap Android phone on mobile data.
- I collect only the personal details I need, and the page says what I do with them.
- The code is exported or backed up somewhere I control.
- I know how to take the site offline quickly if something goes wrong.
When should I bring in a developer?
Earlier than pride suggests, and for less time than you fear.
Pay for a short review before launch if you take payments or store personal data. For a small app this is a short job for a competent developer: ask them to check that the data is protected and the payment handling is sound. That is a far smaller job than a build.
Bring one in properly when customers are paying and you hit any of these: bugs you have asked the tool to fix three times, a feature involving payouts or balances, more than one kind of user, or a need for an app in the stores.
There is also a middle path, which is to learn enough yourself. You do not need to become an engineer. If you can read code well enough to see what the AI changed and why, you get far more out of these tools and far fewer lost Saturdays. The guide to learning to code with AI covers how to go about that.
Use AI to get a small, true first version in front of customers this month. Check it like someone whose money is at stake, because someone's will be. And treat what you build as proof of demand, which is what an MVP is for, not as the finished company.
Key takeaways
- AI tools are good enough to build a one-flow web product that real customers can try.
- Build one small piece per session and test it on your phone before asking for the next.
- Use a payment provider's hosted checkout for M-Pesa instead of asking the AI to invent payment code.
- Keep the code and every account in your own name, and export the code where the tool allows it.
- Before launch, check that one user cannot see another user's data and that no secret keys sit in the code.
- Pay a developer for a short review once money or personal data is involved.
Frequently asked questions
- What is vibe coding?
- It is a casual name for building software by describing what you want to an AI tool and accepting what it writes, without reading the code closely. It is a fast way to get a first version on screen. It is a risky way to run anything that holds money or personal information.
- Do I need to know how to code to use AI coding tools?
- Not to begin. The app builders that work from a chat box need only clear instructions. You will go further, and get stuck less, if you learn enough to read what the tool produced and to describe errors precisely.
- Who owns the code an AI tool writes for me?
- Check the terms of the tool you use, because they differ. In practice, make sure you can export or download the code, that the project sits in an account you control, and that hosting, domain and payment accounts are in your name.
- Can an AI tool connect my app to M-Pesa?
- It can write code that looks right, and sometimes is. It cannot open your payment account, pass the provider's checks or test with real money. The safer route for a first version is a hosted checkout from a payment provider that supports M-Pesa.
- Will a developer be willing to take over an AI-built app?
- Most will look at it, and some will advise rebuilding parts. That is a normal outcome, not a failure. You will hand over a working example, real customer feedback and a clear list of what matters, which makes their job smaller.
Sources
- Row Level Security, Supabase Docs. Checked on 19 November 2026.
- Config and Secret environment variables, Vercel Documentation. Checked on 19 November 2026.
- Payment Channels, Paystack Developer Documentation. Checked on 19 November 2026.
