The obvious move almost killed our project

We were deploying AI models to hospitals.

The plan was straightforward: get the data into our environment, run the models, ship the results back.

We de-identified everything. Built SQL extraction pipelines. Stood up the infrastructure.

And watched the accuracy numbers fall off a cliff.

So we stopped pulling data toward us and asked a different question: what if the model goes to the data instead?

The hospital keeps everything on-premises.

Nothing sensitive leaves their walls, so there's no de-identification stripping out the signal the model needs.

They control access, they control security, and if they want to fine-tune on their own proprietary data, they can do that without sending us a single record.

Accuracy recovered. Compliance got straightforward. The architecture that felt counterintuitive turned out to be the one that actually worked.

Now it's the industry standard.

What assumption is your team treating as a constraint that's actually just a habit?

Originally posted on LinkedIn.