The Corrections Are the Leak

Every time one of our consultants corrects a model's output, on a client project, in a proposal draft, in a code review, that correction is a lesson. The question nobody asks is who's actually getting taught.

Satya Nadella put a name on this last week: the reverse information paradox. The pitch on enterprise AI has always been that the model gets smarter the more you use it, and that's true. What doesn't get said out loud is whose model gets smarter. The prompts your team writes, the fixes they make when the output is wrong, the evaluation sets that define what "good" looks like for your business specifically, all of that becomes institutional knowledge somewhere. It just might not be your institution.

I've watched teams spend months building the thing that actually makes an AI tool useful for a client: not the model, the layer around it. The eval set that encodes what "correct" means for this business. The library of corrections that turns a generic assistant into one that thinks like this company. That layer is the actual differentiator - it's the material that makes the outputs into business value. It's also the part that gets fed straight back into the provider's loop, one correction at a time, invisibly, in a format nobody signed off on sharing.

Here's the part that stings a little. We tell clients AI adoption is about competitive advantage. And it can be. But if every company in an industry is running the same frontier model, refining it with the same kind of corrections, the advantage doesn't accumulate to any one of them. It accumulates in the model. Everyone gets a slightly smarter assistant. Nobody gets a moat.

The enterprise contract tiers exist for exactly this reason: the "we won't train on your data" clauses. Worth reading closely before you assume they cover what you think they cover, and worth asking what "your data" even means when the leak isn't the data, it's the pattern of how you correct it. Start thinking of THAT as your moat, and you're heading in the right direction. And one that might actually be a competitive advantage.