Ask a category manager about their last serious supplier failure and you will hear a story with a sudden ending. The letter arrived, the insolvency was announced, the shipment simply did not come. Then ask what the data looked like for the six months before. Deliveries had been slipping by growing margins. Quality rejects were trending up. Invoices were being chased harder. Two named contacts had left in a quarter. The failure was announced suddenly; it happened slowly, in public, in the buyer's own data.
That gap between observable and official is the entire economics of supplier continuity, and the standard tooling is not built to price it.
Scores answer the wrong question
The supplier risk market sells scores. Each supplier gets a composite number, refreshed periodically, that says how risky. Scores answer a static question, and continuity is a timing problem. You do not need to hear that a supplier is amber. You need to hear that at the current rate of deterioration you have about ten weeks, and what that estimate is based on.
Interventions are priced in time. With twelve weeks of warning, qualifying a second source is a project. With two weeks, it is air freight and overtime. With none, it is a stopped line and a commercial apology. The same failure, caught at different horizons, differs in cost by an order of magnitude. The detection horizon is the product.
To our knowledge, no published study quantifies supplier-failure lead time across signal types. That silence says something about a market that has sold detection for a decade; the vendors hold the data and have not shown the horizon.
Buy lead time, not scores.
Reading the telegraph
The signals mostly sit in your own systems already. Delivery performance decays against its own history, quality drifts by part family, order acknowledgements stretch, and payment behaviour changes where you can see it. Ownership and management churn gets its own compliance urgency from the fifty percent rule. And the failure often starts one tier down, among your supplier's own suppliers.
Each signal alone is noisy. Together, on one timeline per supplier, they turn the vague question of risk into a concrete one. How long until this becomes your problem, with what confidence, and which signals drive the estimate?
What changes when timing is the product
A stated detection horizon can be backtested against the failures that actually happened, so "we would have seen it eight weeks out" becomes checkable rather than atmospheric. Precision matters, because false alarms burn the category team's attention and real misses burn the line. And every warning needs an owner and a decision attached, or the countdown is just another dashboard widget.
Prophesee treats supplier failure as exactly this problem. Timing models run on the supplier graph, backtested against history, and each warning routes as an exception with an owner, a horizon and the intervention economics attached. The backtested horizon is the moat, because it cannot be marketed into existence. Start here.