AI data analytics

Why the answers disagree, and it is not the AI

• The same word means three things. A sale on the day it is ordered, on the day it ships, on the day the money clears. All three are defensible. Nobody chose one, so every report is right and none of them match.
• Returns are counted in different places. One system nets them off, another does not. Two honest reports, two revenue lines, an afternoon lost to reconciling them.
• There is no single identity for a product or a customer. The same item is three records across three systems. Without one true record tying them together, any total is an approximation with a confident face on it.
• Nobody knows when the data last arrived. An import that failed on Tuesday looks like a quiet Tuesday. Freshness has to be on the answer, not in a log nobody reads.

What to build, in the order that pays

None of this is exotic, and the sequence matters more than the tools. Business intelligence and AI on top of it fail in the same place: the layer below.

• Agree the definitions. Write them down. This is a meeting, not a purchase, and it is the single highest-return hour in the project.
• One place where every system lands — raw kept as it arrived, so a disagreement can always be traced back.
• One true record per product and per customer, tying the same thing across systems.
• The dozen numbers the business actually runs on, computed once, in one place, with a timestamp.
• Then the AI layer, answering in plain words from those numbers — and saying so when it is not sure.

What it is good at, and what it should refuse

Once the layer underneath is honest, the AI part is genuinely useful, and the useful parts are narrower than the marketing.

• Good: the question you would not have asked an analyst. Too small to be worth someone's afternoon, so it never got asked. Those questions are most of the value, and they are the ones a queue kills.
• Good: noticing. Reading everything daily and surfacing only what changed. No person reads a whole catalogue every morning; a machine does it before anyone is awake.
• Good: the plain explanation. Not just that a number moved, but which segment moved it. That is a query a person could write and usually does not.
• Should refuse: causes. It can say margin fell and where. It cannot know that a competitor changed price or that a shipment was stuck in customs. An AI analytics layer that offers a confident why is telling you a story.
• Should refuse: silence about uncertainty. When the data is stale, partial or contradictory, the answer must say so. An honest 'these two systems disagree' beats a clean number every time.

Too early if

• You have one system and it is right. Its own reports are enough; buy nothing.
• Nobody can say what a sale is. Fix that first and you may find you did not need the rest.
• You want a forecast before you have clean history. Any forecast should be measured against a plain moving average on your own data before it is trusted, and that comparison is free.
• The real ask is a dashboard for someone senior. Dashboards get built, admired for a fortnight, and abandoned. A short daily message with the exceptions survives.