By Luis Lopez, AI transportation consultant, CEO of Go Hub.io Holdings Corp and subsidiaries, and host of the Freight Guru Podcast
Most fleets already collect more driving data than they can read. The truck reports speed, braking, cornering, idle time and location. The camera, if there is one, flags following distance and phone use. Software then rolls all of it into a single number per driver and calls it a safety score.
That number is useful, and it is also easy to misuse. As an AI transportation consultant, the first thing I look at in a scoring program is not the model. It is what the fleet does with the score once it exists. A fair program makes drivers safer and keeps them on the payroll. An unfair one produces resentment, turnover and a pile of documented warnings nobody acted on.
How a Driver Safety Score Is Built
The mechanics are simpler than the marketing suggests. A scoring system does three things.
- Detects events. Sensors and cameras identify things like hard braking, harsh cornering, rapid acceleration, speeding against the posted limit, close following, lane departure, rolling stops and distraction.
- Weights them. Each event type gets a severity. A brief speed overage counts less than tailgating at highway speed.
- Normalizes them. Events are divided by exposure, usually miles or hours driven, so a driver who runs long weeks is not penalized for simply being on the road more.
The AI part sits mostly in step one. Computer vision decides whether a driver was looking at a phone or a mirror. Machine learning decides whether a deceleration was a panic stop or a normal stop on a steep grade. Newer systems also add context, such as weather, traffic density and road type, before deciding whether an event should count at all. If you are still evaluating the camera side, I covered the practical questions in the top questions to ask about AI dash cameras.
Where Scores Become Unfair
A score is only as fair as the comparison behind it. These are the problems I see most often.
Comparing drivers who do different jobs
A port drayage driver working city streets and terminal queues will brake hard far more often than a linehaul driver on open interstate. A driver hauling a loaded tanker corners differently than one pulling an empty dry van. Ranking them on one leaderboard punishes people for their route assignment, which they usually did not choose.
Counting defensive driving as a violation
A hard brake caused by a car cutting in front of the truck is a good outcome. The driver avoided a collision. A system that cannot tell the difference between that and inattentive driving will lower the score of your most alert people.
Bad map and sensor data
Speeding alerts depend on posted speed limit data that is sometimes wrong, especially in construction zones, on frontage roads next to a highway, and where truck limits differ from car limits. Accelerometers that were mounted loosely or calibrated poorly generate false harsh events. Cameras misread sunglasses, hands near the face, or a driver glancing at a mirror.
Small samples
A driver who worked two short days this week can swing from the top of the list to the bottom on a single event. Weekly rankings built on thin mileage are noise.
Rules for Using Telematics Data Fairly
None of this requires exotic technology. It requires policy decisions made before the first report goes out.
- Compare like with like. Group drivers by operation type, equipment and region. Rank within the group, not across the fleet.
- Normalize by exposure. Use events per set distance or per driving hour, and set a minimum mileage before a period counts toward any ranking.
- Review video before acting. No score should lead to discipline until a person has watched the footage or examined the underlying event data. The model flags; a human decides.
- Give drivers a way to contest events. A simple process where a driver can flag an event as wrong, and get an answer within a few days, does more for trust than any incentive.
- Show drivers their own data. If the score is only visible to management, it feels like surveillance. If the driver sees the same events, clips and trends the safety manager sees, it becomes a tool.
- Track trends, not single weeks. A rolling multi-week view smooths out the noise and shows whether coaching is working.
- Write the policy down. State what is recorded, when the inward camera is active, who can view footage, how long it is kept and what it can be used for. Drivers should sign off on it at hire, and it should not change quietly.
Coaching That Actually Changes Behavior
The score exists to start a conversation. The quality of that conversation determines whether anything improves.
Pick one behavior at a time
A driver handed a list of six problems fixes none of them. Find the one behavior that carries the most risk, usually following distance or distraction, and work on that until the trend moves.
Use the driver’s own clips
Generic training videos are easy to ignore. A thirty-second clip of the driver’s own truck closing on a vehicle ahead is not. Keep sessions short, specific and private.
Recognize the good driving
The same system that catches risk also catches skill: the avoided collision, the long stretch with clean following distance, the driver who improves steadily. Fleets that only call drivers in for bad news teach them that the camera is the enemy. Recognition programs tied to improvement, not just to the top rank, keep average drivers engaged too. This matters for staffing as much as safety, a point I expand on in AI in driver recruiting and retention.
Let AI handle the routine, not the relationship
Automated in-cab alerts and self-coaching clips sent to a driver’s phone work well for minor, first-time events. Repeated or severe events need a person. A driver who gets ten automated messages and never talks to a human will tune them all out.
The Liability Question
There is a legal side to this that owners should think through with their attorney and insurer. Safety data can help a fleet after a crash by showing what actually happened. It can also hurt. If your records show a driver was flagged repeatedly for the same behavior and nobody coached them, that history may be requested in litigation and used to argue the company knew about the risk and ignored it.
The practical lesson is plain: do not collect data you are not prepared to act on. Every flagged event above your severity threshold should have a documented response, even if the response is “reviewed, not at fault.” I go deeper on how small fleets protect themselves in nuclear verdicts and small fleet protection.
Match the Review Load to Your Staff
Before turning scoring on, ask who will review flagged events and how many that person can realistically handle each week. A scoring system that generates more flagged events than your safety team can review is worse than no system, because it creates a record of risks you saw and did not address. Set thresholds so the weekly review load matches the people you have.
The Bottom Line
AI scoring is good at watching every mile, which no safety manager can do. It is not good at knowing why something happened. Fair programs keep those two jobs separate: the software finds the moments worth looking at, and people decide what they mean. Build the grouping, review and dispute steps first, share the data with the drivers it describes, and the score becomes something drivers use instead of something they fear.
If this was useful, subscribe to the Freight Guru Podcast for more practical conversations about trucking, logistics and AI.
About the author: Luis Lopez is a Miami-based AI transportation consultant and logistics entrepreneur, the CEO of Go Hub.io Holdings Corp and subsidiaries, and host of the Freight Guru Podcast.