Data that decides: analytics for SMBs that actually works

The AI conversation at a small company almost always starts badly: somebody wants a model. The right question is which decision you are making by gut feel, and what it costs you to get it wrong.

Three decisions where analytics pays for itself

• How much to buy or produce next month. A mediocre forecast usually beats the judgement of one person.
• When a machine will fail. If you already keep maintenance history, the information is there without collecting anything new.
• Which client to call first. Sorting the list by likelihood to buy changes a salesperson’s month.

If your case looks like none of these, you probably do not need a model yet. Telling you that is part of the job.

Tidy first, model second

Seventy percent of the effort in a data project is cleaning. That is not a complaint, it is the reality: client names written five different ways, dates in three formats, empty fields that actually mean zero.

You need tidy data more than abundant data. Two years of clean history goes a long way. Ten years spread across loose spreadsheets means consolidation comes first, and that step alone delivers value before any model exists.

How to know it works

Define the metric before building, and compare it against how the decision is made today.

If your buyer hits the forecast within a 20% margin of error and the model reaches 18%, it is not worth shipping. If it drops to 8%, you now know how much inventory you stop carrying. That number justifies the project, not the technology behind it.

What to demand from whoever sells it to you

• That the success metric is written down before work starts.
• That your data is used for nothing but your project, and that it says so in the contract.
• That they hand over the model and the data, not just monthly access.
• That they tell you when it is not worth it. If they never do, be suspicious.