Learn AI /Data and Infrastructure /Edge AI

Edge AI

Intermediate
Data and Infrastructure
Programme & data teams
Policy officers
Technical readers

Also called: on-device AI, offline AI.

Detail

In five seconds

What it is

Running AI inference on a device at or near the data source, instead of sending each task to a remote server, which can reduce delay and network dependence.

In plain words

It is AI that runs on your device or near the sensor instead of sending every task to a distant server.

An everyday example

At a water station, a technician carries a small device with a checklist for pump sounds and pressure patterns. When a sensor shows an unusual reading, the device compares it with known warning patterns right there and tells the technician, within a second, whether to inspect the pump. The technician does not need to phone a central office for every reading, so the check still works during a network outage. Later, the central office can still collect selected reports to improve the checklist.

Where the story stops working

A paper-like checklist suggests simple rules, but many edge AI systems use statistical models with many parameters rather than explicit if-then steps. Also, a human technician can use judgement outside the checklist, while a model only produces outputs based on its training and design. The analogy also makes the local device seem fully independent, but many edge systems still rely on central services for updates, security patches or performance review.

Remember thisEdge AI runs inference near the data source, reducing delay, bandwidth use and dependence on connectivity. It does not guarantee privacy or accuracy, and devices, models and updates still require central oversight.

Understand it

Why it matters

For non-specialists, Edge AI matters because where a model runs changes the practical risks, costs and benefits. A local system may respond faster, keep working during weak connectivity and reduce the amount of raw data sent elsewhere. That can matter in remote sites, factories, vehicles, phones and field operations. But local deployment can also make updates, security management, quality control and oversight harder if many devices are involved. In planning or procurement, the key question is not whether edge is better in general. It is whether the use case needs low delay, offline operation, limited data transfer or local control badly enough to justify the device constraints and management effort.

Edge AI and cloud AI: where the processing happens

Show the difference between local inference on a device and remote inference in the cloud, and explain why this changes speed, connectivity needs and data movement.

01
Edge input
A local device such as a sensor box, phone or camera receives new data.
In the left column, draw a device icon at the top with a short label: 'sensor reading, image or audio'.
02
Edge processing
The model runs on that device or a nearby local processor, so the data can be analysed without sending every raw input away.
Below the device, draw a chip icon inside the same box labelled 'model runs here' with an arrow from input to chip.
03
Edge output
The device produces a local result such as a warning, classification or short summary, often with low delay.
Add an arrow to a local alert box labelled 'warning: unusual vibration'.
04
Cloud input and transfer
In a cloud setup, the device captures data but sends it over a network to a remote server for processing.
In the right column, draw the same device icon, then an arrow crossing a network cloud shape to a server icon.
05
Cloud processing and return
The remote server runs the model and sends the result back to the device or operator dashboard.
Draw the server labelled 'model runs here' and an arrow back to a result box on the device side.
06
Trade-offs
Edge can reduce delay and network dependence, while cloud can offer more compute and easier central management. Many systems use both.
At the bottom, add a split note row: left 'faster local response, less raw data sent'; right 'more central compute, easier large updates'; centre note 'hybrid systems are common'.

A two-column comparison. Left column shows Edge AI: data is captured, processed on the device, and an alert appears locally. Right column shows Cloud AI: data is captured, sent over a network to a server, processed remotely, and a result returns to the device. Read top to bottom in each column, then compare the notes at the bottom.

Hover or tap each column to highlight the full path of data and reveal short notes on delay, connectivity and data movement.

Describe this without the diagram

This diagram compares two ways to run an AI system. In the left column, labelled Edge AI, a device first receives data such as a sensor reading, image or audio. The model then runs on the same device or on a nearby local processor. The output appears locally as a warning, classification or short summary. The key idea is that the main processing happens near the data source, which can reduce delay and reduce dependence on a network. In the right column, labelled Cloud AI, the device receives the same kind of data but sends it over a network to a remote server. The server runs the model and sends the result back. This can provide more central compute and simpler management at scale, but it depends more on connectivity and usually moves more data across the network. A note at the bottom says many real systems combine both approaches.

With reduced motion: A static side-by-side comparison with arrows already visible. The left side shows data staying on the device for processing and output. The right side shows data travelling to a remote server and a result returning. Bottom notes list the trade-offs without animation.

Edge AI means running an AI model on a device close to where data is produced. That device might be a phone, a camera, a sensor box, a vehicle computer or factory equipment. The model takes local input data, such as images, sound or sensor readings, processes it on the device, and returns an output such as a label, score, alert or short summary.

What goes in depends on the use case: a photo, microphone audio, temperature readings or machine vibration data. What happens is called inference, meaning the model uses patterns learned earlier to generate a result for new input. What comes out might be “possible leak”, “speech detected”, “translate this sentence”, or “normal versus unusual reading”.

The main difference from cloud-based inference is where the inference runs. With cloud-based inference, the device sends data over a network to a remote server, and the server sends back the result. In Edge AI, much or all of that work happens locally. This can reduce delay, keep working when connectivity is weak, and limit how much raw data leaves the device.

But Edge AI does not mean no infrastructure, no risk or no internet ever. Models still need to be built, tested, updated and monitored. Edge devices also have limits: less compute, less memory, limited power and sometimes weaker security management than central systems. In many real deployments, edge and cloud are combined: the device does fast local processing, while a central system handles updates, logs or larger analysis.

How “A pump technician with a pocket checklist” maps onto the real thing

In the story	In the AI system
small device used at the water station	edge device running the model locally
checklist of warning patterns	trained AI model stored on the device
sensor reading checked immediately on site	local inference on incoming data
central office collecting reports later	cloud system for updates, monitoring or retraining

A worked example

A facilities assistant at the Lima duty station is checking an air-conditioning unit before a weekend closure. An approved phone application can identify common equipment sounds without relying on a live connection.

Starting from: A short recording of the unit, captured locally after the assistant confirms that no speech is audible.

The phone converts the recording into audio features suitable for the model.
A compact trained model runs inference on the phone within its available memory, processing capacity and battery limits.
The application returns a classification such as normal operation, possible fan imbalance or uncertain.
If inspection is needed, the application creates a maintenance ticket and queues a compressed result for synchronisation when connectivity returns.

Result: The assistant receives a prompt on site without sending a continuous audio stream to a remote service. The result supports the maintenance check but does not replace it.

In a United Nations workplace

An associate expert supporting the human resources focal point at a multilingual duty station is screening a long list of applications against a rostered profile before close of business. An approved model runs on the managed laptop and highlights passages that may show the required experience, even while the connection to Headquarters is unavailable.

The inference happens on the laptop rather than on a remote server. This allows the first-pass review to continue with less network dependence and avoids transmitting each application for processing, while leaving the actual screening decision with the responsible officer.

Before you try thisApplications contain personal data and must not be entered into any tool that is not approved for HR material. Local processing does not remove access, retention or security requirements. The model must not reject, rank or shortlist candidates on its own, and every highlighted passage must be checked against the rostered profile by an authorised officer.

Where you may hear it

A vendor demo to an ICT governance board on offline inference
A procurement meeting on sensors for a field presence
An inter-agency working group reviewing connectivity constraints before a pilot
A data protection review of processing on managed devices
A business continuity discussion after a country office network outage

Avoid confusion

Often confused with

Common misconceptions

Common misconception: Edge AI means the system never uses the internet or the cloud.
Correction: Not necessarily. Many edge systems run inference locally but still connect sometimes for model updates, logging, monitoring, fleet management or deeper analysis. 'Edge' describes where some processing happens, not a total ban on network use.
Common misconception: If processing happens on the device, privacy and security problems disappear.
Correction: Local processing can reduce data transfer, but it does not remove risk. Devices can be lost, compromised or misconfigured. Outputs can still be wrong or sensitive, and logs or updates may still move data. Good governance, access control and testing are still needed.
Common misconception: Edge AI is always faster and cheaper.
Correction: It can reduce network delay, but device hardware, optimisation work, battery use and maintenance across many devices can raise costs. Cloud processing may be cheaper or simpler for some workloads, especially if the model is large or the device is very limited.

Quick check

Which example best fits Edge AI?

Go deeper

The simple version

It is AI that runs on your device or near the sensor instead of sending every task to a distant server.

The technical version

Edge AI usually means deploying a trained model so inference runs on hardware close to the data source, such as a phone, camera, gateway, vehicle computer or industrial controller. Inputs are captured locally, processed under device constraints such as limited CPU, GPU, memory, storage and power, and converted into outputs like classifications, anomaly scores or compressed summaries. Compared with cloud inference, this can lower end-to-end latency, reduce bandwidth use and support operation during intermittent connectivity. It is often confused with cloud computing, but the distinction is about execution location for inference, not whether central services exist at all. In practice, many deployments are hybrid: local devices handle time-sensitive inference, while central systems support model hosting, rollout, telemetry, retraining and governance. Because edge devices operate in varied real-world conditions, teams also need to watch for distribution shift, version drift and uneven performance across device types.

Reviewed external resources

Primary sourceAI Risk Management Framework (opens in a new tab)NIST
The NIST AI Risk Management Framework helps readers who need a governance and risk lens for deploying AI systems, including questions about oversight, reliability and operational controls.
Intermediatelink checked 2026-07-31
Primary sourceOECD AI Principles (opens in a new tab)OECD
The OECD AI Principles suit policy and programme readers who want a high-level framework for trustworthy and accountable AI use across different deployment settings.
Intermediatelink checked 2026-07-31
EncyclopaediaWikipedia search: Edge AI (opens in a new tab)Wikipedia
Wikipedia search results for Edge AI give a broad starting point and help beginners find definitions and related topics, but the link is a search page rather than one vetted article.
search results, not one pagelink checked 2026-07-31
VideoVideo explanations: Edge AI (opens in a new tab)YouTube search
YouTube search results for Edge AI can help visual learners find demonstrations and explainer videos, but the link is a search page and video quality will vary.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Edge AI (opens in a new tab)Reddit search
Reddit search results for Edge AI show community questions and practitioner discussion that may surface practical issues and examples, but the link is a search page rather than an authoritative source.
search results, not one pagelink checked 2026-07-31

Where this came from

Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.

The one thing to remember

Edge AI runs inference near the data source, reducing delay, bandwidth use and dependence on connectivity. It does not guarantee privacy or accuracy, and devices, models and updates still require central oversight.