At Geotab Join 2026 in Las Vegas final week (10-12 Feb), greater than 4,000 transportation professionals gathered to listen to how AI is reshaping business fleet administration.
On stage, Geotab CEO Neil Cawse delivered a keynote charting the corporate’s 26-year development trajectory to three,500 workers throughout 22 international workplaces. He pointed to greater than $200m in annual R&D funding and described how AI now underpins car utilisation, predictive upkeep and security optimisation. AI methods, he famous, already outperform people in sure coding duties, a sign of how quickly the know-how is advancing.
Entry deeper business intelligence
Expertise unmatched readability with a single platform that mixes distinctive information, AI, and human experience.
Because the occasion shifted to breakout periods and the exhibition corridor, dialogue turned to linked fleets, product design, vibe coding and digital twins. But behind the product launches and accomplice bulletins lay a extra elementary query, one which goes to the center of the Geotab Security Middle and its broader security proposition: what sort of AI structure ought to energy collision-risk evaluation?
That query surfaced most clearly in a non-public dialog on the sidelines of the convention. Mark Miller, CEO of InsureVision, argued that a lot of the business is constructing on the unsuitable technical foundations.
This isn’t a distinct segment debate between engineer-CEOs. Europe’s dashboard digital camera market is valued at roughly $1.41bn in 2026—a few third of worldwide income—and hundreds of thousands of autos now stream video into telematics platforms. The structure chosen to interpret that footage carries vital operational and business penalties.
On one facet, are the world’s prime telematics suppliers similar to Geotab, Samsara, Powerfleet and Verizon Join. Their methods mix sensor fusion, video analytics and convolutional neural networks (CNNs), refined via years of large-scale deployment.
In Geotab’s case, the stack additionally consists of XGBoost, an open-source machine-learning mannequin optimised for structured information. Whereas video methods detect and classify occasions, XGBoost analyses pace, braking, acceleration and driver historical past to estimate collision danger. The mixture permits Geotab to rank drivers by danger degree and prioritise critical occasions with out relying solely on giant, compute-intensive fashions.
On the opposite facet is InsureVision, an AI start-up based in 2022, contending that incumbents are fixing the unsuitable AI downside.
An AI firm, not an IoT firm
Miller speaks shortly, intensely and passionately, framing collision-risk modelling as a essentially completely different self-discipline from conventional telematics. “Understanding danger is refined and requires sophisticated machine studying,” he says, including: “Doing so requires an AI firm, not an IoT firm.”
In his view, most platforms nonetheless centre on accelerometer and GPS information, mixed with CNN-based video triggers. Even superior CNN variants, he argues, are inclined to classify objects or occasions fairly than interpret the complete driving context. InsureVision’s various is an end-to-end transformer structure educated on tens of 1000’s of crashes and near-misses. The system analyses forward-facing video and learns temporal patterns instantly from information.
“We by no means educated our fashions about crimson lights,” Miller says. “However they perceive that driving via crimson lights is dangerous.” What’s attracting consideration to its providing, says Miller, is its automated fleet security evaluations which may upscale dramatically with deeper automation. One fleet buyer, he says, beforehand reviewed round 200 dashcam clips per day. After deploying InsureVision’s system, that determine rose to roughly 5,000 clips day by day, with AI scoring and categorising occasions so managers can give attention to genuinely high-risk behaviour and training.
Having launched the product solely just lately, Miller is now targeted on constructing traction, providing free pilot programmes to fleet operators. The system, he says, “integrates with a number of completely different dashcam corporations. It doesn’t substitute the dashcam, it’s only a layer that sits on prime,” including an AI-driven analytics layer to current digital camera infrastructure fairly than requiring new {hardware}.
For Miller, the benefit of transformers lies in contextual reasoning. For instance the declare, he flips open his laptop computer and hundreds a video clip shot by an unnamed telematics firm. “I downloaded it and reprocessed it via our system,” he says. “We utilized our AI danger evaluation to the very same footage.”
The forward-facing dashcam view reveals a straight stretch of highway. A car within the adjoining lane begins edging in direction of the host driver. A number of seconds in, Miller pauses the clip. On his overlay, a big crimson “87” dominates the display screen, his system’s danger rating.
“At this level,” he says, tapping the display screen, “their system is silent. Ours is already screaming at you you’ve received an enormous downside.”
Based on Miller, his mannequin has detected the early indicators of a cut-off manoeuvre, delicate lateral drift and shutting distance earlier than any abrupt motion happens. “We’ve received a sensor that understands danger,” he says. “Not simply occasions.”
The video resumes. The neighbouring automobile completes the cut-in, forcing the host driver to brake sharply. The opposite car accelerates away. No collision happens.
Because the braking occasion unfolds, Miller factors once more to his overlay. The danger rating drops quickly, ultimately returning to zero because the menace dissipates.
“Now look,” he says. “In the meanwhile the opposite car speeds away, our rating is again to zero as a result of the chance has handed. That’s when their system flags an issue.”
In his telling, the competing platform reacts solely after the accelerometer registers harsh braking, not when the visible cues first sign hazard.
“So the one cause this turned an occasion clip,” Miller says, “is as a result of there was a harsh brake. The accelerometer detected it, not the pc imaginative and prescient.”
For Miller, the demonstration encapsulates his broader argument: conventional telematics platforms are event-led and sensor-triggered, whereas AI-native methods goal to mannequin danger constantly, earlier than and after a measurable incident happens.
Additionally it is price noting that InsureVision’s comparability was carried out on archived footage that had been downloaded and reprocessed, fairly than in a reside, real-time deployment.
Edge versus information centre
The divide just isn’t solely architectural however geographical, in computational phrases.
Edge processing means working AI fashions instantly on or close to the car. These methods function with restricted energy and reminiscence and should ship choices in milliseconds, for instance, issuing a real-time warning if a driver is tailgating.
Knowledge-centre processing, in contrast, entails transmitting footage or telemetry to distant servers outfitted with highly effective GPU clusters. These methods can run bigger fashions however eat extra power and introduce latency.
Miller’s strategy leans towards giant, end-to-end fashions able to deep contextual inference. Cawse questions whether or not such methods are sensible at fleet scale.
Later on the convention, Cawse addressed the suitability of transformers for real-time security with a measured, deliberate tone and quiet confidence. “Transformer structure is successfully a text-token output mannequin,” he says.
Whereas highly effective, it’s not robotically one of the best match for analysing telemetry or issuing split-second alerts. “You’ll be able to’t flip measures like laborious braking and pace turn into textual content tokens. It’s not the fitting device for the job. The older fashions are nonetheless one of the best,” he provides.
His issues centre on latency, price and energy consumption. A big Transformer mannequin, he urged, might require eight to 10 high-end GPUs, draw round 30 kilowatts of energy and take a number of seconds to generate a solution. “You’ll be able to’t run that on the edge,” Cawse says. “When anyone’s getting too near the car in entrance, you may’t wait 5 seconds.”
Such infrastructure might strategy $200,000 in {hardware} prices, he added, troublesome to justify for in-vehicle deployment.
The sustainability dimension was taken up by Mike Department, Geotab’s vice chairman of knowledge & analytics. “And from a sustainability perspective, you wish to compute what that takes,” Department says, pointing to the power depth of enormous AI methods working in information centres.
“In the end, we’ll decide one of the best mannequin for the job to attain these outcomes,” he added. “The fashions that we’ve created up to now, you may take a look at the outcomes we’ve already created for purchasers. That’s what we actually search for. Is the know-how and the fashions we’re selecting proper now going to suit the invoice for reaching that discount of collisions? And it already has, and it’ll proceed to evolve.”
Join our day by day information round-up!
Give your corporation an edge with our main business insights.
