Five signs the software AI built for you is quietly failing

Software built with AI tools rarely fails with a bang. It fails quietly, in ways that each look like a one-off annoyance, until the day they don’t. If you own a system that was built quickly and cheaply in the last few years, these are the five signs worth taking seriously.

1. The same bugs keep coming back

You paid to have a problem fixed. It came back. You paid again. It came back wearing a different hat.

In well-built software, a fixed bug stays fixed, because the person fixing it understood why it happened. AI tools are good at making a symptom go away and poor at addressing causes. Each “fix” adds another patch on top of the last, and the pile itself becomes the problem. If your bug reports feel like a subscription, the underlying structure needs looking at, not the latest symptom.

2. Small changes take longer every month

The first version arrived in a fortnight. Two years later, changing the wording on an invoice takes three weeks and breaks the reports page.

This is the clearest measurable symptom there is. Healthy systems get easier to change as the team learns them. AI-built systems tend to get harder, because the code was generated without an overall design, and every change fights the tangle a little more. Owners often blame the developer. It’s usually not the developer.

3. Your hosting bill creeps up and nobody can say why

AI tools write code that works in a demo with ten records and struggles with fifty thousand. The classic pattern is a system that asks the database the same question hundreds of times instead of once, and the standard remedy applied by whoever is maintaining it is to pay for a bigger server.

If your monthly infrastructure costs rise faster than your customer numbers, you are very likely paying rent on inefficiency. An afternoon of targeted fixes can make a serious dent in a bill like that; a bigger server never does.

4. Everyone has quietly agreed not to touch it

Listen for the phrases. “Best not to change that bit.” “It works, leave it.” “We’ll do that after the busy period.” When the people around a system start treating it like an unexploded bomb, they are telling you something true: the last few changes broke things nobody expected, and confidence is gone.

A system nobody dares change isn’t stable. It’s frozen, and your business keeps moving while it stands still.

5. It handles money or personal data, and nobody has ever checked the locks

This is the quiet one, because there are no symptoms at all until there is an incident. AI tools cut security corners with complete confidence: passwords stored badly, payment logic trusting the browser, customer records readable by anyone who knows where to look. The system works perfectly the whole time.

If your software takes payments or holds personal information and it has never had an independent security review, put this at the top of the list. Under UK GDPR the exposure is legal as well as technical, and “our supplier said it was fine” is not a defence the ICO recognises.

What to do if you recognised your business

Don’t panic, and don’t commission a rewrite. Almost everything above is fixable, usually for far less than starting over, and the earlier it’s caught the cheaper it gets. The sensible first step costs nothing: get a clear picture of where you actually stand, then decide with real information instead of dread.