\n<\/p>\n
Rajesh Subramaniam is Founder and CEO of <\/em>embedUR systems<\/em><\/a>.<\/em><\/p>\n Taking AI models out of the cloud and running them on devices at the edge may sound simple, but the reality is anything but. When you move AI closer to where decisions actually happen, the stakes get higher. A cloud model can afford to be occasionally imperfect, but a model inside a medical device, an industrial robot or a smart door lock has little room for error. That shift from a centralized environment to thousands or millions of devices in the field changes the entire development and validation mindset.<\/p>\n So, when people talk about intelligence at the edge and how much computing power these devices now have, I always say the same thing: Be careful. The moment AI leaves the protected environment of the cloud, accuracy becomes the most critical thing to get right. <\/p>\n Validation is the hurdle that will decide who succeeds with edge AI and who learns painful lessons. At the same time, people tend to underestimate the amount of testing required to achieve meaningful accuracy on edge devices. You are no longer just proving the model in a lab; you are proving it in the wild.<\/p>\n Consider a door lock camera system that is trained to perform facial recognition. A 97% accuracy rate in testing may sound good, but in the real world, with an armful of heavy groceries, that 3% gap becomes a big deal\u2014let alone when it\u2019s a burglar trying to get in.<\/p>\n Accuracy starts with credible data. If the training data does not reflect the real environment, the model will behave differently the moment it is deployed. Edge AI demands longer validation cycles because you have to confirm the model works not just on historical data but also on real hardware, in real conditions, with noise and variation that never show up in clean cloud training environments.<\/p>\n Once a model is in production, a new challenge appears: model drift. The risk starts when the device starts seeing data it has never been trained on. It grows stale, and that leads to wrong predictions or missed alerts. The only solution is a constant feedback loop. Data collected at the edge still must move to the cloud so the model can be retrained with the new patterns. Then the updated model needs to be pushed back to the device so it grows smarter over time.<\/p>\n We talk a lot about decreasing cloud dependency, and that is definitely happening because edge devices now have enough compute power to perform local inference. But an effective model life cycle still requires a back-and-forth connection, and that loop is the only way to keep edge AI accurate.<\/p>\n<\/section>\n Organizations should adopt a software environment that keeps them flexible. Hardware evolves faster than most product teams can keep up with. New processors and chipsets will continue to arrive, and if everything is tied to one type of hardware, updates become slow and difficult.<\/p>\n I always suggest building and deploying with a consistent software layer that remains platform independent. If you lock yourself into the wrong hardware stack, you will struggle to scale, while if you decouple your software, you can swap hardware when you need to without breaking the whole system.<\/p>\n That independence supports three priorities for edge AI. The first is security. Software deployed at the edge cannot be manipulated or compromised. These models are in the field, not inside a guarded data center, meaning threats are closer and more unpredictable.<\/p>\n The second is supporting the reality of model drift and updates. If the platform is rigid, it becomes easy for the model to age faster than developers can keep up.<\/p>\n The third is remaining efficient with compute resources. You will make smarter decisions about where to run workloads if your software and hardware are not tightly bound together. Sometimes the cloud is best; sometimes the device is best. The goal is to choose each time without friction.<\/p>\n<\/section>\n There is a tremendous world of hype surrounding AI. Some of it inspires innovation, some encourages shortcuts. My biggest caution to leaders is to resist the urge to add AI everywhere simply because you want to check a box. Not everything needs AI embedded in it. A smart door lock? Sure. A washing machine? Maybe not\u2014or not yet, at least.<\/p>\n If you put a flawed model into the field, you risk losing customers and revenue and damaging your brand credibility. Once trust is gone, it is very hard to win it back. If you deploy edge intelligence, take the time to validate every piece. Ensure the data is real, the model behaves as expected and there is a plan to keep it learning. In edge AI, patience and discipline in both design and deployment are what will ultimately protect your business and customers.<\/p>\n Forbes Technology Council<\/u><\/a> is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?<\/u><\/em><\/a><\/p>\nGood Data Is Key<\/h2>\n
How To Decide Where AI Should Run<\/h2>\n
Do It Right Or Don\u2019t Do It<\/h2>\n
\n
<\/section>\n<\/div>\n