David Julian is the CTO and cofounder of Netradyne.gettyPicture this: At a construction site, two excavators are digging into the ground. A bulldozer idles nearby, its operator waiting to move in once the area is cleared for grading. Meanwhile, workers are moving around between the machines. The combination of people and heavy equipment may create a ​dangerous situation. Several miles away, an 18-wheeler carrying hazardous chemicals is en route to deliver its load to local businesses. Passenger vehicles are whizzing by, some even cutting off the truck. If the truck driver reacts a second too late, a tragedy could occur. ​These are just two of many examples of situations where high-stakes decisions happen—and where seconds, even milliseconds, matter. In some of those environments, uploading data to the cloud may be inefficient due to slow internet speeds, making it challenging for cloud solutions to process and deliver critical information in real time. If there's no internet connectivity, it's impossible for cloud solutions to process and deliver critical information in real time. That’s where edge intelligence can step in. The Difference Between Edge Intelligence And Cloud-Based AI Edge intelligence is a form of AI that runs locally on devices (such as heavy equipment, vehicles, cameras and sensors) rather than in the cloud. Edge intelligence can power physical AI systems, enabling real-time perception and action in the physical world. ​In the earlier examples, if the excavators, bulldozer and 18-wheeler were equipped with cameras and sensors running edge AI models, the operators would be able to receive real-time alerts if workers got too close. Because edge AI systems don’t rely on the cloud, they have lower latency and reduced dependency on connectivity and bandwidth. They can gather data, process it and act on it in real time. If a distracted worker walks in front of a bulldozer, how quickly the operator receives an alert is crucial—a few seconds can be the difference between life and death. Similarly, if a passenger vehicle cuts off an 18-wheeler, the sooner the truck driver gets alerted, the less likely a collision becomes. Best Practices For Designing Edge Intelligence Solutions That Can Power Physical AI SystemsBased on my team’s work implementing edge intelligence capabilities into our solutions, which are focused on enabling safer driving and more effective fleet management in the logistics industry, I recommend several best practices for determining whether an edge intelligence approach is appropriate for a given problem and, if so, how to design an effective edge intelligence solution. ​​First, it’s important to know where the loop closes. If a decision must occur within hundreds of milliseconds to seconds, in a zone where there may be no internet connectivity or when the volume of sensor and video data is too great to upload to the cloud efficiently, then I advise moving forward with the edge intelligence route. Without such time and connectivity constraints, a solution can be cloud-based. ​​​I also advise designing edge intelligence systems to be searchable rather than merely reactive. Traditional approaches rely on predefined event triggers, but when edge intelligence solutions can capture rich, structured representations of the physical world, it becomes possible to search across a fleet of devices in real time for conditions that were never pre-programmed—and without pulling raw video to the cloud. For instance, a single 30-second scene at an intersection could help users answer safety, operational and risk questions long after the moment has passed. This turns edge devices from narrow event detectors into a distributed, searchable intelligence layer.​Governing at the edge is also vital—thinking through privacy, prompt screening and policy enforcement at the device level rather than tied to the cloud, as well as implementing guardrails. Any solution that runs edge intelligence should have appropriate guardrails in place to minimize the risk of unintended behavior or harm. ​It's also important to measure all behavior, not just flagged activity. Dedicated, on-device processing enables holistic analysis with the same rigor and consistency, rather than just analyzing isolated moments that trigger alerts. For example, measuring what percentage of stop signs a driver observes, how consistently they maintain a safe following distance and how they operate across an entire shift, gives a comprehensive view of their driving. That comprehensive view of the driver's good, compliant and risky behaviors can help managers better understand how that behavior relates to real-world outcomes like accidents.​ Cloud-based approaches are generally not well suited to cost-effectively transmitting and processing large volumes of continuous data, which is why they generally tend to rely on sampling or triggered events.Additionally, edge intelligence systems should be built on a shared foundation. When a common model is trained centrally and deployed uniformly across devices via over-the-air updates, it enables the model to incorporate a larger breadth of data points. In turn, this enables stakeholders to gain a better understanding of whichever factor they're evaluating, such as driving behavior. ​Challenges Of Edge Intelligence—And How To Mitigate Them From my observations, there are a few key challenges of edge intelligence, but they're mitigable.​First, edge environments have highly specialized performance, power, thermal, latency and cost trade-offs. To navigate that, it’s essential to consider factors such as model distillation, quantization, hardware and software codesign in-house. It's also vital to pursue ongoing over-the-air improvements to work around those limits. Additionally, the ability to continually update edge intelligence models is paramount to ensuring that systems remain effective as new data and capabilities emerge.Edge intelligence also requires higher precision, because incorrect alerts erode trust. I advise running and training edge intelligence systems, especially vision-based ones, on as much real-world data as possible. That way, the system can understand the broader operational context rather than relying solely on predefined trigger events. ​The Future Of Edge Intelligence Edge intelligence is evolving beyond safety alerts into a general-purpose intelligence layer for the physical world, one that can continuously perceive, reason and learn from its environment. Proactive insights, such as an edge system that warns a driver they are about to run a stop sign, help prevent bad outcomes. As models grow more capable and hardware becomes more efficient, I believe the question will shift further from "What events can we detect?" to "What questions can we answer?" In my view, organizations that invest in this architecture now are building a compounding advantage because every scene observed and every edge case encountered makes their systems smarter.​Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?