Ask one question of any predictive system in your company. Which decision consumes this? Put it to every demand forecast, churn score, risk rating and cash projection you can find, and the usual answer is silence. That silence is expensive, because a prediction no decision consumes is just a report, and reports change nothing.
The difference between the two is a contract, and you can write it down before anyone trains a model.
The four clauses
A prediction owes the decision it feeds four things.
- What, precisely. State a quantity, over a horizon, at a granularity someone can act on, not "demand will soften". Vague predictions escape accountability.
- By when. Every decision has a deadline; after it, the best prediction in the world is trivia. A system that cannot meet the deadline should say so.
- Who acts. Name one owner, not a distribution list. If the answer is "the dashboard's viewers", the answer is nobody.
- How it is scored. When the outcome lands, someone compares the call to what happened and keeps the comparison. Skip this clause and the other three decay, because nothing separates the predictions worth trusting from the ones worth retiring.
Dashboards break all four
Most enterprise predictions arrive on a dashboard, and a dashboard breaks every clause at once. It advises, so nobody has to act. It stands open forever, so no deadline applies. It broadcasts, so nobody owns it. And it refreshes rather than scores, so last month's wrong call disappears under this month's confident one.
Watch what happens after a bad quarter. Nobody asks the dashboard to explain itself; they ask the person standing nearest the number, by name.
None of this is the dashboard's fault. It was built to distribute information, and it does that well. The mistake is asking a distribution surface to do a decision system's job, then blaming the models when nothing improves. The models were often fine. Nobody wrote the contract.
The decision side is only half
Palantir has argued for connecting agents to decision structure (Connecting Agents to Decisions), and the framing is right as far as it goes. But a well-staged, well-owned decision built on an unscored prediction just delivers confident error faster. You need both ends. The decision holds the prediction to account; the score tells the decision how much trust the prediction has earned.
A scored prediction gets better or gets retired. An unscored one just gets older.
Prophesee writes the contract in as engineering. Every model is bound to the named decision it serves, every call carries its deadline and owner, and every outcome is scored against the call that preceded it, with the record kept. Competitors find that hardest to copy, because a scored claim can be checked and theirs cannot. To see the contract on one of your own decisions, start here.