Most businesses do not need a model yet

Every few weeks someone asks whether machine learning could help their business. Usually it could, eventually. But the honest answer is that most of them are two steps away from that conversation, and skipping those steps is how projects fail expensively.

Here is what those steps are, and how to tell where you actually stand.

The question comes before the data

• A model predicts something. Before anything else, you need to know what that something is and what you would do differently if you knew it.

"We want to use AI on our sales data" is not a question. "We want to know which customers are likely to stop ordering, so we can call them before they do" is. The second one tells you what to predict, who acts on the prediction, and what happens next. Without that, you get a model that produces a number nobody uses.

The test is simple: if the prediction were perfect, what would change on Monday morning? If you cannot answer that, the model is not the problem to solve first.

Then comes the data you already have

Most businesses have more data than they think and less usable data than they hope. The gap is usually in one of three places.

It is spread across systems that do not talk to each other. Sales in one place, inventory in another, customer records in a third, and nothing linking them.

It is inconsistent. The same customer entered four different ways, dates in three formats, a product renamed halfway through last year with no record of the change.

It is too short. Machine learning finds patterns in history. Six months of data covering one season tells you very little about what happens next season.

None of these are exotic problems. They are the normal state of a business that has been running for a while and recording things as it went. But they are the actual work, and they come before any model.

What to do instead, for now

If the question is clear and the data is messy, fix the data. Consolidating your records into one clean, consistent place is unglamorous work that pays off regardless of whether a model ever gets built. You will find things in the process. Businesses routinely discover their best customers are not who they assumed.

If the data is clean and the question is clear, start with something simpler than a model. A moving average forecast, a rule based flag, a well built report. These take days rather than months, and they set the bar that any model has to clear. Plenty of times nothing clears it, which is useful to know before spending on the alternative.

If both are in place and the simple approach is not good enough, now the conversation about models is worth having. And it will go faster, because the hard parts are already done.

Why this gets skipped

Partly because machine learning is more interesting to talk about than data cleanup. Partly because the industry rewards ambitious projects over sensible ones.

But mostly because the simple version feels like admitting the problem was not that hard. It usually was not. That is good news, not a disappointment.

The businesses that get real value out of this stuff are rarely the ones who started with the most sophisticated approach. They are the ones who knew what question they were asking, cleaned up enough to answer it, and only reached for something more complicated when the simple thing genuinely ran out.