Every enterprise now owns AI (e.g. copilots in the office suite and a machine learning team in the data function). Ask the same enterprises to name one decision that improved because of it, and the room goes quiet.
The numbers say the quiet is deserved. The MIT NANDA initiative's report, The GenAI Divide: State of AI in Business 2025, found that 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 models were never the constraint. The distance between a model's output and a decision someone commits to is.
Where the value leaks out
Watch a typical AI output travel through a large company. A model produces a forecast or a risk score, and it lands in a dashboard or an email. A person, if they see it at all, weighs it against their own judgement and decides the way they always have.
The value leaks three times on that short journey:
- Nobody has to act on the output. It informs a decision at best. It does not carry one.
- Nobody owns it. No named person is accountable for acting, so acting is optional, and optional actions lose to the day job.
- Nobody scores it. No one records what was predicted, what was decided and what happened next, so the pilot that saved nothing and the pilot that saved millions produce the same slide.
The MIT researchers place the stall at the workflow boundary, not the model boundary. People were handed fluent output that did not fit how work got decided, and they quietly routed around it.
What the 5% do differently
The minority getting value do not deploy AI at the organisation. They wire it into specific, named decisions.
A decision, concretely, is four things. It is 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 inside that loop, not next to it.
Their questions are worth borrowing. 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? Where is the evidence, six months later, that acting beat not acting? None of these are data science questions. They are operating questions, and most AI teams are empowered to build models, not to change who decides what.
The debt that builds while you wait
Buying more intelligence does not close a decisional gap. The upgraded model lands in the same dashboard its predecessor did, which is why abandonment climbed in the very year model capability jumped. What piles up in the meantime deserves a name. Call it decision debt. Every recurring decision made late, blind or unowned adds a little interest, and the balance compounds quietly while the pilots multiply.
Paying it down is unglamorous work. Pick the decisions. Name the owners. Set the thresholds, route the prediction to the person rather than the portal, and record the outcome so the record settles which predictions deserve trust.
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. It is a decision layer that turns enterprise data into predictions, exceptions, plans and answers, wired to named owners with outcomes recorded. And it is built as a dozen shared problem classes rather than thirty modules, so every decision one team wires up makes the next one cheaper to wire. Built with and proven inside global enterprises. To see it on your own data, start here.