What AI can’t do for a mid-market business, and where it goes wrong

The most expensive AI mistakes are not technical failures - they are buying something AI cannot actually do in your situation. Understanding the genuine limits of the technology is more useful, right now, than any capability list on a vendor slide.

Every conference, vendor and headline now tells you artificial intelligence changes everything. Some of it is true. Most of it is sold without the other half of the sentence. The more useful question, and the one almost nobody answers plainly, is what AI cannot do, and where it quietly goes wrong once it is in front of real people making real decisions. Knowing the edges is what lets you use the middle well.

AI is a pattern machine, not a mind.

At its heart, a model learns patterns from what has happened before and uses them to say something about what happens next. That is powerful, and it is also the whole of what it does. It does not understand your business. It does not reason about a future that looks nothing like the past. It cannot weigh a decision the way a person who carries the consequences can. When you hear a machine described as thinking, read it as predicting from history, and you will judge it far more accurately.

It cannot fix data it was never given.

A model can only work with what it can see. If the history is missing, inconsistent, or scattered across systems that do not agree, the model does not tell you so. It answers anyway, confidently, and the answer is wrong in a way that reads exactly like an answer that is right. This is the single most common way AI goes wrong in a mid-market business. The tool did not break. It was asked to reason over a mess, and it dressed the mess in a clean sentence.

It cannot tell you why, only what is likely.

A churn model can tell you a customer is likely to leave. It cannot tell you why that customer is leaving, and it certainly cannot tell you whether the reason is something you caused. Models trade in correlation, not cause. They are very good at spotting that two things move together and silent on whether one drives the other. Treat the output as a prompt for a human to investigate, not as an explanation, and you stay on the right side of the line.

It cannot own a decision.

When a number turns out to be wrong, a board does not want to hear that the model said so. Accountability stays with people, and that does not change because the analysis got cleverer. A regulator, an auditor or a chair will ask who decided and on what basis, and a person has to be able to answer. AI can inform a decision, sharpen it, speed it up. It cannot take responsibility for it, and any business that lets it try is storing up a problem.

Where it earns its place.

None of this is an argument against using it. It is an argument for using it where it is strong. Predicting from your own history, who is likely to lapse, what is likely to sell. Working at a scale no person can, scoring every customer rather than the few someone has time to think about. Turning plain-language questions into answers over a trusted model. Taking the drudgery out of summarising and drafting. Point it at those and it pays. Point it at judgement, cause, or accountability and it will let you down with great confidence.

So this week, take one thing someone has promised AI will do for you, and ask two questions of it. What data would it learn from, and is that data clean and complete enough to trust. And who owns the decision once the answer comes back. If both have good answers, it is worth doing. If either does not, you have found the work that comes first, and it is usually getting the data and the decision clear rather than buying a tool. That groundwork is the Establish phase of our Analytics Acceleration Programme, and building the models that pay is our machine learning work. AI is a strong tool with clear edges. The businesses that win with it are the ones that respect both.