AI can now write most of the code in a business application. That part is real. We recently redesigned our own website with AI coding agents, and the new version went live the same evening the idea came up. A person approved every step, and a few things the AI got wrong were caught before anyone outside the company saw them.

The speed is not the lesson. What decides whether AI-built software helps a business or hurts it is everything around the code. These are the principles we hold to, and the ones we'd suggest to anyone using AI to build, whether they hire us or not.

1. Decide what to build before the AI writes anything

AI will happily build the wrong thing very quickly. The expensive mistakes come from a vague request, not from bad code. Write down who will use it, what problem it solves and what "done" looks like before any building starts. The time AI saves on typing is best spent here.

2. Never let the builder grade its own work

Asking an AI whether its own work is good tends to get a yes. Have something separate check it: a second reviewer that only sees the finished result, a person, or both. On our own site, an independent review sent the first version back with several fixes, one of them important. None of them would have been caught by asking the builder.

3. Keep people at the points that need judgment

AI is good at doing. It is not accountable. A person should decide what the product should feel like, what is safe to show publicly, how customer data is handled and whether something is ready to launch. If nobody can explain how a system works end to end, it isn't ready, no matter how well it seems to run.

4. Test it the way a real customer would

Code that looks right can still fail in the real world. Click through it on a test copy, on a phone, with a slow connection and with the wrong input. Some of the most useful fixes on our own site came from simply using a preview copy the way a visitor would, not from reading the code.

5. Put a ceiling on anything that costs money

AI features and text messages are billed per use. Without limits, one bug or one bot can run up a large bill overnight. Set spending caps and usage limits before launch, not after the first surprise invoice.

6. Plan for the AI being wrong or unavailable

AI services slow down, go offline and sometimes give confident wrong answers. Any feature that depends on one needs a sensible fallback, so the customer still gets something useful, and a clear line on what the AI is never allowed to decide on its own.

7. Check what search engines and customers actually see

Modern websites can look perfect in a browser and still appear blank to search engines. Before calling a public page done, check it the way Google reads it, not just the way it looks on your screen.

8. Collect less, and delete on a schedule

The safest data is the data you never stored. Ask only for what you need, keep visitor input out of analytics tools, and set an automatic deletion date for anything temporary.

Questions to ask before anyone builds with AI for you

  • Who decides what gets built, and is it written down?
  • Who or what checks the work, other than the thing that built it?
  • Which decisions always go to a person?
  • What stops this from running up costs if something goes wrong?
  • What happens when the AI is down or wrong?
  • What data does it keep, and for how long?

Good answers to those questions matter more than which AI tool is used. If you're weighing a custom build, we're happy to talk through how they'd apply to your project.