When AI software fixes create more problems

You've used AI to build a useful app for your business. It saves time, replaces a spreadsheet or makes an awkward task easier. Then you start adding to it.

Someone needs a different view. A number looks wrong. A new workflow doesn't quite fit. You describe each problem, AI makes a change, and you carry on.

Eventually, you notice that fixing one thing seems to disturb something else. The app has more settings and exceptions. Explaining how it works takes longer. You're spending more time correcting it than improving it.

That's a useful moment to pause and ask whether you understand what needs fixing.

A requested fix can miss the cause

A person can accurately describe a frustrating experience without knowing its cause. "This number looks wrong" is a valuable observation. "Change the calculation" is a proposed solution. There's investigation to do between the two.

Morgan Philo describes this in an account of a customer directing software changes through an AI agent. Some reported problems involved correct data presented confusingly. Other issues shared a deeper cause in how the system represented the business. Working through the requests individually hadn't resolved that cause.

The frustration was real. The changes being requested weren't necessarily the changes needed.

Consider an equipment register. Staff say returned equipment still appears overdue, so they ask for a button to dismiss the warning.

That button might be useful. But perhaps the return was never recorded, the equipment came back to a different location, or the screen is showing an old assignment. Each explanation calls for a different response.

Adding a dismiss button before establishing the cause could make the warning disappear while leaving the records unreliable.

The same thing can happen with a developer writing every line by hand. AI makes it easier to move quickly from a plausible explanation to an implemented change, so checking that explanation becomes especially useful.

Signs your AI-built app needs a closer look

There's another question to ask, too: how does this change affect the rest of the work?

An extra approval step might resolve one complaint while slowing everyone else down. A second place to edit information might make one task convenient while creating disagreement between records. A new setting might solve an unusual case but leave every user with another decision to make.

Each change can work as requested and still leave the overall application harder to use.

There are a few signs that deserve a closer look:

• The same issue keeps returning in slightly different forms.
• A change for one group causes trouble for another.
• Information appears in several places and doesn't agree.
• Staff rely on increasingly complicated workarounds.
• Nobody can clearly explain why a rule or exception exists.

These are reasons to investigate. They don't automatically mean the app needs rebuilding.

Check the problem before changing the software

Before asking for another change, write down what happened and what you expected. Be specific about who was doing the work and what they needed to achieve. Then follow the relevant information through the process. Where was it entered? Who changed it? What should the next person see?

AI can help with that investigation. Ask it to explain its assumptions, consider alternative causes and identify what evidence would distinguish them. Check those explanations against how the business actually operates.

You can also use AI to try an idea before committing to it. A small prototype shown to the people doing the work may reveal that a proposed screen is confusing or an approval step is unnecessary.

That early version doesn't need to become the finished system to be valuable. If it exposes a bad assumption, you've learned something while changing direction is still relatively easy.

The difficulty comes when every version adds more software while the original assumptions remain untested.

Business knowledge matters throughout this process. The person using the tool understands the work and the consequences when something goes wrong. Technical assessment helps connect that experience to the software's behaviour and the choices available. AI can help investigate and implement those choices.

If you're weighing conflicting recommendations, I've also written about how to assess AI advice in "When ChatGPT disagrees with your software developer".

Choosing the next step for your app

At Calo, that's the conversation we want to have when a useful AI-built tool becomes difficult to change. What has it proved? Where is it causing friction? What evidence would help us choose the next step?

Sometimes the answer is a clearer screen. Sometimes it's a better workflow or a change to how information is stored. Establishing which one you need can save a long sequence of otherwise successful fixes.

Our AI prototype-to-production service helps businesses assess what they've built and choose a practical next step.

Further reading

This article was prompted by Morgan Philo's account of AI-directed development, alongside writing from Sean Goedecke and commentary from Theo. These explore the software development perspective in more detail:

• Morgan Philo: AI can fix the symptom and miss the problem
• Sean Goedecke: If you are good at code review, you will be good at using AI agents
• Theo: You're using AI coding tools wrong. A detailed video for developers about using prototypes and feedback to test ideas earlier.