Edge AI: when should you run AI directly on a device?

For many businesses, the challenge is not a lack of data. It is getting the information they need in time to act. A camera captures an image, a sensor takes measurements, or a medical device takes a photograph, but the analysis happens in the cloud. The data has to be transmitted and processed before the result reaches the user or system. When every second counts, connectivity is unreliable or the data is sensitive, this approach can limit the effectiveness of the entire solution.

Can a device make decisions on its own?

Yes. AI can analyse data directly on a machine, camera, terminal, vehicle, sensor or medical device. This approach is known as Edge AI. A trained AI model processes data locally, where it is generated, and immediately identifies an event, pattern or anomaly. This does not mean giving up the cloud. The two approaches can complement each other.

When should you consider Edge AI?

Local data analysis can be a good fit when:

• a process requires a real-time response, without waiting for cloud analysis;
• a device needs to work with limited connectivity or offline;
• images, audio or user data should not leave the device unnecessarily;
• data transmission and storage costs increase as the deployment grows;
• large volumes of data are sent to the cloud even though the business only needs selected events and results.

What should happen on the device, and what belongs in the cloud?

Fully local processing works well when autonomy, minimal delay and reduced data transfer are priorities. A hybrid approach is useful when a device needs to respond immediately, while the business also needs centralised reporting, comparisons across devices, remote management or model improvements based on data from the entire device fleet. In this setup, the device handles the immediate response and initial data filtering. The cloud handles reporting, brings together information from multiple devices, manages the fleet, and supports model development and updates.

Applications include automated quality inspection, predicting equipment failures, intelligent medical devices, IoT systems, logistics and monitoring. In each case, the goal is the same: to shorten the time between an event and the right decision, rather than simply generate more data.

Example 1: a machine that does not wait for cloud analysis

Consider a manufacturer that needs to inspect every component leaving a production line. Manual inspection is costly and can produce inconsistent results, while sending the entire camera stream to the cloud adds delay and expense. If a defect is detected too late, the result could be a whole batch of faulty products or a production stoppage.

An AI model running locally at the production line analyses images in real time and detects defects. The system then routes the affected component for further inspection. It also sends the results and selected images to a central system. The device handles the immediate decision, while the central system provides reporting, tracks quality trends and supports ongoing improvements.

The result is inspection of every product without slowing the line, faster responses to deviations, fewer defects and less data transfer. The value of the deployment can be measured through response times, defect rates, false alarms and inspection cost per unit.

Example 2: faster support for clinical staff while protecting patient data

A portable dermatoscope presents a similar challenge. Clinical staff need a fast, consistent assessment of image quality, including during mobile examinations or in locations with limited connectivity. Sending every image to the cloud increases waiting times and means sensitive patient data is processed more widely.

An AI model running on the dermatoscope or a connected terminal analyses the image during the examination, assesses its quality and flags cases that need further consultation. The result is available within seconds, and the image can remain on the device. Only the information needed for documentation or further assessment is sent to the healthcare provider’s system. The solution assists clinical staff without replacing a doctor’s diagnosis.

This division of responsibilities combines speed and offline operation with central oversight. It reduces repeat examinations and the volume of data transferred and stored outside the device. Key measures include analysis time, image quality, the accuracy of referrals for further consultation and the amount of data leaving the device.

How do you move from an idea to a working solution?

Bringing AI onto a device should not start with choosing a model. First, define which decision needs to happen faster, what data is available, what must work offline and which information should stay local. You can then assess the hardware and decide whether processing should take place entirely on the device or use a hybrid architecture.

Delivering this kind of project requires expertise in AI, device software, electronics, infrastructure, integration and cybersecurity. At Sii, our Edge AI projects bring together AI specialists and embedded systems engineers. We help assess the use case, prepare data and build a prototype, optimise the model for the target device, and integrate it with the client’s systems. We test the solution on the hardware it will run on and support scaling, remote updates and ongoing maintenance after deployment.

How can you assess the potential of AI in your product?

If response times, unreliable connectivity, transmission costs or data protection requirements limit the use of the cloud in your product or process, it is worth exploring whether some analysis could happen directly on the device. Start with a specific business problem, the available data and your product requirements. A prototype can then help establish which benefits are achievable and how to measure them before a wider rollout.