In 2012, I joined Direct Line Group’s motor team to help build a telematics insurance product. The idea was simple: use real driving data to move beyond crude proxies—age, postcode, occupation—and price risk based on how people actually drive.

I believed in it. Most people working in insurance technology at the time did.

Fourteen years on, I think we owe ourselves an honest look at what telematics delivered and what it didn’t, because the gap between what we promised and what we built has real consequences—especially for carriers writing commercial auto.

We said we were measuring behavior, but we were mostly measuring exposure.

The headline metric was harsh braking. Sounds intuitive; drivers who brake sharply are driving less well, right?

Not quite. Harsh braking frequency is dominated by factors that have nothing to do with individual skill. Drive through a city at rush hour and you’ll log far more harsh braking events than the same driver on an empty road at 6 a.m.

These are variables that do predict claim frequency: how much the vehicle is driven, where it is driven, and when it is driven. Commercial underwriters already capture much of this through proxies at bind, including business location, operating radius, industry class, client base, cargo type and hours of operation. A 2025 Risks study of around 100,000 Italian policyholders is a useful warning: Harsh braking looked significant in some univariate analysis, but once the authors adjusted for actual kilometers driven it was no longer significant. So, telematics confirms what underwriters already know, continuously and at finer resolution. That has genuine value, but that value is incremental.

Some will push back, saying continuous data at vehicle level is more granular than anything reconstructed from a submission. Fair point. But does the additional precision actually change underwriting decisions, or does it mostly validate what the underwriter already thought? Many carriers are asking whether this incremental refinement justifies the investment.

As a behavioral measure, telematics does identify the extreme tail—drivers whose behavior is so far outside the norm that the signal cuts through the noise. However, for the broad middle of any fleet, where most claims actually occur, the data is underpowered. That’s the central limitation that the industry has been slow to name.

The Compromises We Settled For

When raw telematics data didn’t deliver the behavioral insight we’d promised, the industry adapted. The focus shifted to compliance (often with dashcam video): phone usage, gaze detection, seatbelt wearing, speeding. All real risk signals worth capturing. But they are a proxy for driving quality, not a measure of it. A driver who never touches their phone but consistently follows at unsafe distances is a worse risk than one who looks at a notification but reads the road well.

Compliance does not mean competence. That distinction matters.

In commercial fleets, the pattern is familiar. Safety teams review dashcam event clips, coach on harsh braking and monitor phone usage. Compliance metrics improve as a result. But I have not found independent published evidence showing that telematics adoption, by itself, has produced a robust causal improvement in commercial auto loss ratios. There is evidence that telematics can improve risk selection, influence driver behavior and reduce crashes in some fleet settings. The harder question is whether those effects reliably translate into carrier-level underwriting profitability.

The language shifted, too, which is worth noticing. “Telematics will transform how we price risk” became “telematics is useful for FNOL, fraud, claims management.” Those are real benefits. But they’re not what was originally promised, and right now they’re not what commercial auto carriers most need.

Crash Detection: Where the Gap Is Most Visible

In 2015, my team at Direct Line Group tried to build automatic crash detection. We had claims data, telematics data, accelerometers. The plan was straightforward: train a model to recognize what a real collision looks like in the G-force signal.

We couldn’t get past the first step. Claims data didn’t match telematics records cleanly enough to build reliable training labels. The accelerometer couldn’t separate collisions from potholes and speed bumps. So, we fell back on a crude heuristic—G-force spike plus vehicle dwell—and set the threshold high enough to make the noise manageable. (Editor’s Note: Vehicle dwell is the time a vehicle is not moving.)

The industry has invested heavily since then. Providers now layer AI on top of raw accelerometer data, applying proprietary collision rules and signal-processing filters. Some employ large teams of analysts reviewing flagged clips before anything reaches a fleet manager. Serious efforts, genuinely.

But the underlying constraint hasn’t changed. An accelerometer records deceleration force. It doesn’t record what caused it. No amount of filtering changes what the sensor can see. And on the dashcam side, crash detection still scales with reviewer headcount, not technology.

The trade-off is structural. Raise the detection threshold and you reduce false positives from potholes, curbs and speed bumps, but you also filter out lower-energy collisions. Lower the threshold and you drown in noise. The danger is that current systems become best at detecting the crashes that would have been reported anyway while missing lower-speed, disputed-liability or vulnerable-road-user incidents where immediate evidence is most valuable.

What the Loss Ratio Is Telling Us

In 2024, S&P Global Market Intelligence reported a commercial auto combined ratio of 107, only modestly better than 109 in 2023. The liability component remained deeply unprofitable at 113. Separately, Triple-I and CAS found that commercial auto liability claim severity rose 94% between 2015 and 2024, a 7.6% compound annual growth rate. Depending on definition, recent North American fleet surveys put telematics adoption at roughly 80-90%. The technology is succeeding. It just isn’t making a big enough difference to underwriters.

Related article: Why Insurance Telematics Integrations Fail

The carriers that outperform use telematics too, but that’s not why they outperform. Progressive reported a commercial lines combined ratio of 87.0 in 2025. While Progressive has been an early mover on telematics, it attributed this profitability to rate management, underwriting capability investment and expense management. Telematics is one input. The edge is in how that input gets used within a disciplined underwriting operation.

Where and when a fleet operates are largely fixed—determined by industry, clients, routes. An underwriter captures that at new business and renewal. Telematics confirms it continuously but rarely changes the picture. How drivers actually drive is different. That’s something fleets are actively trying to change, and something carriers can underwrite more precisely with the right data.

So, the question for any carrier considering another telematics investment cycle isn’t whether telematics has value. It’s which specific value lever you’re trying to pull. Faster FNOL—telematics does that reasonably well for high-severity impacts. Compliance monitoring—dashcams deliver. But genuine risk selection, identifying the drivers whose behavior actually predicts future claims—that requires asking whether the data was ever designed to answer that question.

What a Different Data Source Makes Possible

This isn’t an argument against telematics. The selection benefit is real; fleets and drivers who voluntarily enroll tend to be lower risk (all other things being equal), regardless of what the driving data shows. Mileage, time-of-day, route—useful exposure measures, and telematics does them well.

But the industry has reached a point where the compromises need naming. We measured exposure and called it behavior. We measured compliance and called it quality. We built crash detection that catches the crashes we’d have heard about through a phone call anyway.

The data gap those compromises created is addressable now through video, not because cameras are new—dashcams have been in commercial vehicles for years—but because the AI processing the footage has changed fundamentally in the last three years. Modern camera-based neural-network approaches used in systems such as Tesla’s FSD can analyze footage contextually rather than simply triggering object-detection rules. These new models understand how a situation develops: the child near the curb, the gap in traffic a driver does or doesn’t anticipate, the following distance that tells you something real about risk before anything goes wrong.

The compromises were rational at the time. The technology has moved on. Carriers that find a genuinely different signal, rather than a more refined version of what’s been available for 15 years, will have an edge that compounds.