AI is precise. Original thinking, less so.
An honest assessment of where AI helps in data projects, from the practice of someone who works with it every day.
I work with AI every day in my data projects. I have even built my own tooling for it, so I can work directly with my data model instead of adjusting everything by hand.
So this is not an article by someone judging it from the outside. It is my honest state of affairs after a year of intensive use.
Where it excels
Precision and stamina.
Checking a calculation, and then again, and then again. Reviewing the same data three times without getting sloppier on the third pass. That is exactly where a human gives up and where the difference is made in this trade, because most reporting errors are not thinking errors but attention errors.
Automating is the second one. Repetitive work you can describe, you can hand off. That saves not minutes but hours, and above all it removes the resistance to starting that work at all.
And as a sparring partner it works better than I expected. Not because it has better ideas, but because you have to formulate something before you can put it forward. Half the time I already know the answer before I have finished typing the question.
Where it falls short
Thinking outside the box.
That is the honest limitation and it is structural. You get excellent execution of what already exists, and rarely the sideways step. The solution that is not obvious, the question nobody has asked, the connection that is only visible if you know that sector: that comes from you.
In a data project that is precisely the valuable part. Building a model is execution. Deciding which question that model has to answer is not.
I notice it most clearly with definitions. Ask what an active customer is and you get a defensible answer. Whether that answer fits how your business works is something only you know, and that is immediately the most important decision in the whole project.
Why I built my own tooling anyway
Because the combination works.
I make the decisions about what has to happen and why. The execution, the checking and the repeating I hand over. That is a division of labor that matches what each side is good at, and it gives me more time for the part where I make the difference.
What I deliberately do not do is let it decide on the content. A model that looks correct and answers the wrong question is more dangerous than a model that is visibly broken.
What that means for an SME
If you are considering starting on your reporting with AI: do it, but expect the right things.
It is going to help you build faster, check more than you otherwise would, and automate work you would otherwise never start.
It is not going to tell you which numbers matter for your business. That comes out of your head, or the head of someone who knows your sector. And anyone who tells you a model will do that for you is selling you something other than what they deliver.
The claim
AI raises your speed and your accuracy, and it changes nothing about the quality of your questions. In a data project the expensive mistakes sit on the question side, and there is still no tool for that.