Every enterprise now owns AI. Copilots in the office suite, a chatbot on the intranet, a machine learning team somewhere in the data function, a slide in the annual report. The adoption question is settled.
The value question is not. The MIT NANDA initiative's report, The GenAI Divide: State of AI in Business 2025, put a number on it that should be uncomfortable: about 95% of enterprise generative AI pilots deliver no measurable impact on profit and loss. S&P Global Market Intelligence found that 42% of companies abandoned most of their AI initiatives during 2025, up from 17% a year earlier. Spending rose. Results did not.
The constraint was never intelligence. It was the distance between the model's output and a decision someone actually commits to.
Where the value leaks out
Watch what happens to a typical AI output inside a large company. A model produces a forecast, a summary, a risk score. It lands in a dashboard, a report, an email. Then a person, if they see it at all, weighs it against their own judgement, their spreadsheet, and the way things are usually done, and makes the decision the way they always have.
Three leaks, in sequence:
- The output is advisory. It informs a decision at best; it does not carry one. Nobody's Monday changes because a dashboard exists.
- The output is unowned. No named person is accountable for acting on it, so acting on it is optional, and optional actions lose to the day job.
- The output is unaccountable. Nobody records what was predicted, what was decided, and what happened next, so the organisation cannot tell its working AI from its decorative AI. The pilot that saved nothing and the pilot that saved millions produce the same slide.
The MIT research itself locates the stall at the workflow boundary, not the model boundary: fluent output that did not fit how work got decided, quietly routed around.
What the 5% do differently
The minority that gets value does something structurally different, and it is visible from outside. They do not deploy AI at the organisation. They wire it into specific, named decisions.
A decision, concretely, has four parts: a prediction of what happens if nothing changes, a threshold at which someone must act, a named owner who acts, and a record of whether the action worked. The companies extracting value put AI into that loop, not next to it.
That changes the questions. Not "which model?" but:
- Which recurring decisions in this function are made late, blind, or not at all?
- What would a prediction have to say, and how far ahead, for the owner to act on it?
- Who is the owner, and what does their runbook say when the threshold trips?
- Where is the evidence, six months later, that acting beat not acting?
None of these are data science questions. They are operating questions, and they are the reason value concentrates in so few hands: most AI programmes are run by teams empowered to build models, not to change who decides what, when.
The uncomfortable implication
If the gap is decisional, buying more intelligence does not close it; the upgraded model lands in the same dashboard its predecessor did. This is why the abandonment numbers climbed in the same year model capability jumped: the bottleneck moved, and budgets did not move with it.
The work that pays is less glamorous. Pick the decisions. Name the owners. Set the thresholds. Route the prediction to the person, not the portal. Record the outcome and let the record settle which predictions deserve trust. Do that for one function's ten most expensive recurring decisions and the P&L impact stops being unmeasurable, because measurement was built in from the start.
AI creates value at exactly one moment: when a person decides something differently because of it. Everything else is cost.
This is the problem Prophesee was built for. A decision layer that turns enterprise data into predictions, exceptions, plans and answers, wired to named owners with the outcome recorded. Built with and proven inside global enterprises. If you want to see it on your own data, start here.