It looked perfect. It wasn't safe.
Weeks before launch, the platform looked finished. Sellers could sign up and list their homes. Agents had accounts and subscriptions. Every page worked.
Then I asked a second AI to find what was wrong with it.
Its first security review found that anyone could make themselves a verified agent and see private listings. An email function had no login check, so anyone could have used it to send mail. An endpoint with no login could suspend agents and email every admin.
None of it showed on screen.
From a printed sheet
The platform is the idea of my co-founder, now our CEO. In his years as an estate agent, he sold a surprising number of homes off-market, for sellers who didn't want to appear on Rightmove or Zoopla. He handed me a printed sheet with the idea sketched out.
I took it to ChatGPT for the business plan, spec and PRD, then into Lovable for a first draft. Within weeks it looked like a finished product. That's exactly the trap.
Past projects had taught me that AI writes code that runs, then tells you, with total confidence, that it's done. What it doesn't do is ask what happens when someone uses it in a way you didn't expect. So I gave a second AI, Claude Code, a different job. It wasn't there to build. It was there to act as the senior engineer.
Fixed isn't the same as shipped
We fixed every flaw. Then came the deployment surprises. One fix sat unshipped for three days because everyone assumed pushing the code was enough. A single bad import quietly took eleven server functions offline for two days, and nothing alerted us. Then we found seven functions that had never reached production at all.
The rules that made it safe
Every mistake became a rule, written into a file the AI reads at the start of every session:
- Pushing isn't deploying. Check what's actually live.
- Test before production. A smoke test checks that every function starts up, and nothing reaches customers until it has passed on a separate staging site.
- A human approves every release. The AI never ships to production itself, and every release needs a real description of what's changing.
- Write everything down. AI forgets between sessions. If it isn't committed, it's gone.
- Ask "how do you know?" For anything touching money, personal data or production, the answer must point to code, a test or a check, not confidence.
- Argue with me. If an instruction looks wrong or risky, the AI says so before acting.
That last rule earns its keep. When I asked for posts promising "fixed agent fees", the AI checked our contract and pointed out that the fee is a cap, not a fixed fee. I made the final call, but I made it with the facts.
None of these rules are new. They're what good engineering teams have always done. What's new is writing them down for an AI, and covering its blind spots: it only sees the files it opens, it starts every session from scratch, and it can be confidently wrong.
What we built
The platform is now live, and AI runs right through it.
Picture a seller listing their home. Smart forms help them fill in their details as they go. AI gives them a sense of what the home is worth, then writes a description with nearby amenities, schools and transport links. It reads their photos to paint a picture of the property for agents.
Agents are checked too. Remember the flaw that let anyone become a verified agent? Now AI looks each one up on Companies House, review sites and industry bodies, and a person makes the final call.
Behind the scenes, everything connects: payments, email, Meta, Microsoft Teams, GitHub and Vercel. Our marketing draws on our own intelligence built from public data, and feeds a dashboard that shows how customers interact with us, from the first ad they see to the moment they sign up. The platform then adjusts our marketing based on what it learns.
What this means for your business
- You can test an idea without hiring a team first. With AI, you can go from idea to a working product in front of real users in 8 to 12 weeks.
- "Looks finished" isn't finished. Have something, or someone, review the work the way a senior engineer would.
- Write your rules down. An AI follows the process you give it, and only that.
- Make it argue with you. The pushback is where much of the value is.
- Experience still counts. My years leading tech teams, alongside business, marketing and operations colleagues, shaped every decision. AI amplifies what you bring.
- Hire engineers when you can. A good engineer will keep improving the platform and make sure it can be supported by more than one person, or one AI.
So, can AI help build a start-up?
Yes. And the biggest change isn't speed. It's freedom. When building a feature no longer means hiring for it, the limit stops being budget and becomes imagination.
My co-founder's printed sheet is now a live platform. If you have an idea on a sheet of paper of your own, or a team wondering how to build with AI safely, let's talk.