The Upstream Data Discipline Every AI Product Depends On

I have shipped enough products to know where they break, and it is almost never the part everyone worries about. The model is rarely the problem. The data feeding it is, specifically whether the meaning of that data was pinned down before anyone trained on it. Here are five principles I hold to before I let a model touch health data. They all live upstream of the algorithm, which is exactly why they get skipped.

A model trained on inconsistent definitions learns your mess, then scales it. The OMOP Common Data Model exists for exactly this reason. OHDSI describes it as an open community standard that puts observational data into one structure and one set of vocabularies, so the same concept means the same thing across every source. Before I trust any AI product on health data, I want that layer in place. Otherwise diabetes in one feed and diabetes in another are not the same variable, and the model quietly averages two different things into one confident number.

The codes you standardize on decide what your model can and cannot see, so I treat vocabulary mapping as a product call. Choosing how source codes map to standardized concepts sets the ceiling on every downstream feature. OHDSI's standardized vocabularies do this work in the open, which is why one institution can run another's analysis without rewriting it. If you hand that decision to a script nobody reviews, you have outsourced your product's semantics to whoever wrote the mapping, and you will not like the result.

Here is the map I keep in my head. The dangerous corner is not low ambition, it is high ambition sitting on weak data.

Findable, accessible, interoperable, reusable. Skip one and the model inherits the gap. FAIR is not academic hand-waving. OHDSI's own work on FAIRness for common data models shows why it matters in practice: if data is not interoperable and reusable, every new model starts from zero. I want the second AI product my team ships to reuse the first one's data plumbing, not rebuild it from scratch. FAIR is what makes that reuse possible.

This is the counterintuitive one. Every hour spent on clean semantics returns days you would have lost debugging the model. When meaning is fixed, you can query across organizations, reproduce a result, and swap one model for a better one without re-plumbing everything underneath. OHDSI's network runs standardized analytics across dozens of institutions precisely because the data speaks one language. That is not slower. That is how you go fast without breaking things.

This is the newest shift, and I take it seriously. The next consumer of your data may not be a person. OHDSI is now running a survey on OHDSI in the era of medical AI and has put out a journal call for agentic AI in real-world evidence. Agents querying health data will expose every semantic shortcut you took. A human analyst guesses what a messy field probably means and usually guesses right. An agent does not guess. It takes the field at face value and hands you a confident, wrong answer at scale. Clean semantics used to be good hygiene. With agents in the loop, it is the difference between an asset and a liability.

None of this is about slowing down. It is about putting the work where it actually pays, upstream, before the model, before the agent, before the demo that convinces everyone the hard part is done. The hard part is the data. It always was.