Data is a raw material, not a strategy

A drug does not fail in the boardroom. It fails on the line, in a batch, because something went into it that should not have. AI is the same, and almost nobody is treating it that way.

A drug does not fail in the boardroom. It fails on the line, in a batch, because something went into it that should not have. Every pharma operator knows this in their bones. You can have the best process in the world, and a tainted active ingredient will still ruin the batch, and no amount of process brilliance downstream will save it.

AI is the same, and almost nobody is treating it that way.

Walk into most large organisations and ask about data, and you will get a strategy. There is a data strategy, a chief data officer, a lake, a governance committee, a slide with arrows. What you will not get is an answer to a simpler question: is this data, the data going into this model, fit to go into the line. That gap, between the strategy on the slide and the raw material in the pipe, is where AI projects quietly fail.

In manufacturing nobody has a raw material strategy. They have specifications. They have qualified suppliers. They have incoming inspection and a batch record that says where every input came from and whether it met spec. Raw material is not a theme to have a point of view on. It is a thing that is either fit or not fit, today, for this batch.

Data in an AI production system is raw material. It has a supplier. It has a provenance. It has a quality that is either adequate for the use or it is not. And like any raw material, if it is wrong, everything built on top of it is wrong, confidently and invisibly.

The trouble with data strategy is that it stays abstract. It lives at the altitude of slideware and committees, where it can be discussed forever without anyone having to qualify a single input. Meanwhile, underneath it, models are being fed whatever was to hand: the export someone pulled last quarter, the scraped corpus nobody checked the rights on, the spreadsheet with the lineage of a rumour. The strategy is immaculate. The inputs are unqualified. The output is a coin toss with good grammar.

Treating data as raw material is not a bigger strategy. It is a smaller, harder discipline, and pharma already owns every piece of it.

It means specifications. What does good enough mean for this input, stated as acceptance criteria, not as a feeling.

It means provenance. Where did this come from, through whose hands, with what right to use it. A batch record, not a shrug.

It means supplier qualification. If you are buying or licensing data, who is the supplier, what is their process, and would you let them into your line if it were a physical ingredient.

And it means incoming inspection. Before the data goes into the model, something checks that this batch is fit, the way goods-in checks a delivery against the spec before it ever reaches the floor.

None of this is exotic. It is the most ordinary thing in regulated manufacturing. The shift is only in where you point it. Stop asking whether you have enough data. Start asking whether this data is fit to go into the line. The first question has kept consultants busy for a decade. The second one is the question that decides whether your AI works.

The winners here will not be the ones with the best data strategy. They will be the ones who treated their inputs like a regulated raw material while everyone else was still admiring the lake. You cannot strategise your way out of a contaminated input. You can only qualify it, or ship the consequences.