You cannot manage on the number you care about.
A 200-vehicle fleet has about 2 serious incidents a year. Going from 2 to 1 is not evidence of anything. That is not a measurement problem you can solve with better reporting — it is arithmetic.
How long before you would know
Pick a level and an improvement. The wait is computed, not asserted — and it is the reason safety programmes are managed from the bottom.
One year, 200 vehicles
select a levelHow large an improvement you are trying to prove
of watching serious injury or write-off before a 30% improvement could be told apart from ordinary variation.
- Events a year
- 2
- Events needed
- 148
- Costs a year
- 840,000
Longer than anyone will wait. Manage this level and you are reading noise — which is what most fleet safety reporting is doing.
Two Poisson counts over equal periods, 5% significance and 80% power. The normal approximation is optimistic at very small counts, so the top of the pyramid is worse than this, not better.
Proving a 30% cut in serious incidents would take 74 years — a working lifetime. Proving the same cut in risky behaviour takes 4 days. Both describe the same fleet getting safer; only one of them tells you inside a quarter.
So you manage the base and let the apex follow
Every level in that pyramid is the same driving. The difference is how often it produces something you can count, and whether it has already cost you money by the time you count it.
A near miss and a collision are frequently the same event with a different amount of luck attached. One of them you get 15 times more often, and it is free.
| Level | Per year | Costs | Time to prove −30% |
|---|---|---|---|
| Serious injury or write-off | 2 | 420,000 | 74 years — a working lifetime |
| Reportable collision | 22 | 38,000 | 6.7 years |
| Minor damage | 96 | 4,200 | 19 months |
| Near miss | 340 | — | 23 weeks |
| Risky behaviour | 14,200 | — | 4 days |
Four things change an outcome
Ordered by how fast they act, because that is the only axis on which the in-cab warning is unbeatable — it is the sole intervention that happens inside the event.
The slowest one is usually the largest. Shift length and night driving change how much exposure a fleet has at all, and no amount of coaching competes with simply not sending someone out at 03:00.
- 01 Under a second
In the cab
An audible warning while there is still room to stop. Nothing else in this list can act inside the event itself.
Near misses that would have become contact - 02 Same day
In the driver's app
Their own events and their own clip, before a manager raises it. Most events end here.
Repeat behaviour by the same driver - 03 Weeks
A coaching session
Only for a sustained pattern. Coaching everything trains people to ignore coaching.
The tail of the score distribution - 04 Months
Policy and rostering
Shift length, night driving, route choice. The slowest lever and often the largest.
Exposure itself, rather than behaviour
"Average claim cost" describes almost no claim
122 claims, 2,276,200 in total. The mean is 18,657 and the median is 4,200 — the average is 4.4 times the claim you actually have most often.
One claim carries 27% of the cost. Any business case built on reducing the average is really a bet on not having that one, which is a different and much less controllable thing.
- Under 5,000 96 · 403,200
- 5,000 – 25,000 14 · 168,000
- 25,000 – 100,000 7 · 385,000
- 100,000 – 500,000 4 · 700,000
- Over 500,000 1 · 620,000
What we are not claiming
This category sells on a number nobody can honestly promise. Since the page has just spent a section explaining why that number is unmeasurable, it would be strange to then quote one.
- We cannot promise fewer serious incidents
- Nobody can, on a fleet this size — and the arithmetic above is why. What we can show is movement in the base of the pyramid, which is the only part that produces a signal inside a budget cycle.
- The relationship is assumed, not proved
- That fewer near misses means fewer collisions is a well-supported assumption in safety research. It is not something your own data will demonstrate, because your own data will never have enough serious events in it.
- Cameras change behaviour partly by being cameras
- Some of the early improvement is people knowing they are recorded rather than anything we detected. That effect is real, and it fades. The coaching is what makes it stick.