The practice of running AI model inference directly on local devices, such as smartphones, in-store displays, cameras, and IoT hardware, rather than sending data to a central cloud server for processing. For agencies, edge AI is the architecture behind personalization and analytics capabilities that must operate in real time, offline, or under data privacy constraints.
Also known as on-device AI, edge inference, on-device machine learning
Edge AI moves the inference step of AI from a remote server to the device where the data originates. Instead of capturing an image, sending it to the cloud, receiving a classification back, and acting on it, an edge AI system classifies the image on the device itself within milliseconds. The same applies to text, audio, sensor readings, and any other data type: the model runs locally, and only the decision or a summary of the result needs to be transmitted.
Running models on edge devices requires smaller, more efficient model architectures than cloud deployments allow. Techniques like model quantization, which reduces the numerical precision of model weights, and model pruning, which removes less important parameters, are used to compress models enough to fit on hardware with limited memory and compute. Dedicated chips including Apple’s Neural Engine, Google’s Edge TPU, and Qualcomm’s AI chips are designed specifically for efficient on-device inference.
Edge AI intersects with data privacy requirements because sensitive data never has to leave the device. A model that analyzes a customer’s face for in-store personalization without transmitting biometric data to a server satisfies a different regulatory profile than one that sends that data to the cloud for processing. This distinction is increasingly meaningful as privacy regulations expand their reach to behavioral and biometric data.
Agency campaigns increasingly extend into physical spaces: retail environments, out-of-home displays, connected vehicles, and wearables. Each of these contexts has real-time response requirements, connectivity constraints, or privacy obligations that make cloud-dependent AI inadequate. Edge AI is what makes intelligent experiences in these contexts actually work.
In-store personalization requires it. A working ad agency building a personalized in-store display system cannot tolerate the 200-400 millisecond round-trip latency of a cloud inference call for every customer who walks past. An edge model that classifies audience segment from camera input in 15 milliseconds on local hardware can update the display in real time. The difference in latency determines whether the experience is seamless or perceptibly delayed.
Mobile ad personalization increasingly depends on it. As third-party data sources are restricted and on-device signals become more important, the personalization that happens inside mobile apps is moving toward on-device models that infer user context from local data without transmitting it. Agencies building mobile-first campaign strategies need to understand what this architecture enables and what it cannot do.
It changes the privacy conversation with clients. Clients in healthcare, finance, and retail who have been hesitant to use AI in customer-facing contexts because of data transmission concerns often have a different reaction to edge AI. The data stays on the device. The privacy profile changes substantially. Understanding this distinction lets agencies reopen conversations that were previously closed.
An agency is designing a digital signage personalization system for a grocery retail client. The brief calls for displays that adapt product recommendations in real time based on the demographic profile of the person in front of the screen. A cloud architecture is initially proposed, but the client’s legal team flags the biometric data transmission requirement as a potential GDPR issue. The agency redesigns using an edge AI model that runs on the display hardware itself, classifying general audience segments from camera input locally without transmitting any image data. The edge model updates the display content within 20 milliseconds. The legal approval takes two weeks instead of three months.
The automations and agents module of the workshop covers how to build AI workflows that operate at campaign speed, including the architectural decisions that determine where inference happens and what data leaves the device.