Your model works. It's still going to fail. Welcome to the 85%.
Only ~11% of ML models ever reach production. After a decade building them, the reason has almost nothing to do with the model — and everything to do with the other 90% of the work.
⚙️ Your model works. It's still going to fail. Welcome to the 85%.
A poll doing the rounds this month put a number on something everyone in this field feels: barely one in ten machine learning models ever reaches production. Gartner's version is blunter — around 85% of ML projects fail to deliver. After a decade building these systems, my honest reaction is not surprise. It is recognition.
Because here is the uncomfortable truth: the model is the easy part. A capable data scientist can hit strong accuracy on a clean dataset in a fortnight. That is where the demos live, the notebooks impress, and the leadership deck gets its chart. And then the real work begins — the 90% nobody photographs:
The data pipeline that has to feed the model reliably, at 3am, when the source system changes format without telling anyone.
The drift — the model was trained on a world that keeps moving, and quietly gets worse every month it isn't watched.
The integration — a prediction is worthless until it lands inside a workflow a real person already uses, at the moment they need it.
The trust — a business user who cannot see why the model said what it said will not act on it, and is right not to.
The ownership — who retrains it, who monitors it, who is accountable when it is wrong at 2pm on a Friday.
None of that is modelling. All of it decides whether the model ever matters.
What I have learned leading ML work: the projects that succeed are not the ones with the cleverest algorithm. They are the ones that treated the model as one component of a product — with an owner, a pipeline, a monitoring plan and an adoption strategy — from day one. The failures almost always fell in love with the accuracy score and forgot the other five things.
So the question I put to every ML project I review, before anyone touches a model: who owns this in production, and how will we know when it starts to fail? If those two questions do not have names attached, the model does not need building yet — the plan does.