Samsara's New Risk Model Says Half of Crashes Trace to One-Tenth of Drivers — and Named Which Habit Combinations Matter Most

Samsara published its Compounding Risk Report on Sept. 8, 2026, drawing on telematics data collected across its fleet customer base from July 1 through Dec. 15, 2025. The company's core finding: the top 10% of drivers by its proprietary risk score account for 47% of crashes, and the top 30% account for 76%. Arpan Podduturi, Samsara's head of safety product, said the practical implication is concentration of effort — "by concentrating their most intensive coaching on just 10% of drivers, managers can reach the group associated with nearly half of crashes" — while cautioning in the report itself that "crashes are rare and depend on many factors outside any model's view."

The model's actual contribution isn't the concentration finding by itself — fleet safety programs have long known that a minority of drivers drive most of the risk. It's the shift from scoring single behaviors to scoring combinations of them. Evaluated alone, none of Samsara's roughly 50 tracked risk factors is a strong predictor: harsh braking carries a 1.8x crash-likelihood multiplier, mobile phone use 1.7x, distracted driving 1.7x, speeding 1.5x. Stacked together, the multipliers compound sharply rather than simply adding — mobile use plus harsh braking rises to 4.5x, and mobile use plus distraction plus harsh braking reaches 5.4x. Samsara built the model using double machine learning and causal forests, methods intended to separate genuine causal risk factors from behaviors that merely correlate with risk, and reports that it correctly prioritizes crash-involved drivers over non-crash-involved drivers three times out of four across both next-day and seven-day prediction windows.

No independent trade coverage found so far — including Commercial Carrier Journal's writeup of the same report — includes outside validation of the methodology; both CCJ's piece and Samsara's own materials rely entirely on Podduturi's statements and the report's self-reported figures. That's worth flagging plainly: this is vendor-published research, not third-party audited data, and Samsara's own disclaimer states the report is "for general informational purposes only" with no guarantee of results. The underlying claim — that combining exposure, behavior and context data outperforms single-behavior scorecards — is consistent with peer-reviewed driving-risk research generally, but the specific multipliers and the 47%/76% concentration figures are Samsara's numbers, drawn from Samsara's own customer base, using a model Samsara has not published for outside replication.

What this means for brokers and fleet safety budgets

Three places this actually changes a decision, not just a scorecard:

Broker carrier-vetting standards just got a data point to point to. Since the Supreme Court's May 2026 ruling in Montgomery v. Caribe Transport II, LLC raised the stakes on negligent-carrier-selection claims, brokers have faced pressure to show they did more than check a safety rating before dispatching a load. A telematics vendor publishing a model that claims to flag high-risk drivers before a crash — rather than after one shows up in a CSA score — gives brokers and their insurers a new category of pre-incident evidence to ask carriers for, whether or not they use Samsara specifically.

Single-metric coaching triggers are the wrong target. A fleet that coaches every driver flagged for one harsh-braking event or one phone-use alert is, per this model, spending attention on a weak signal; the report's own numbers say the combination of behaviors — not any one of them — is what concentrates real risk. Safety budgets built around single-event triggers should be re-examined against combination-based flagging regardless of which telematics vendor a fleet uses.

Treat the specific multipliers as vendor-reported, not settled science. A forwarder or broker citing this report to a customer or an insurer should attribute the 47%, 76% and 5.4x figures to Samsara by name and note they come from Samsara's own customer telematics data — not from an independent or peer-reviewed study — since no third party has yet reproduced them.