On-Device AI Compliance Requirements: What a CE/FCC/UKCA Dossier Must Now Disclose

Editorial Team August 12, 2026

What on-device AI compliance actually covers

On-device AI compliance requirements are the disclosure obligations a CE, FCC, or UKCA conformity dossier must now carry when a device runs its AI inference locally rather than in the cloud. In short: the standard electronics certificate alone is no longer a complete conformity file. On-device AI refers to the capability to perform AI tasks such as processing and learning directly on the unit, ^[1] explains, because the work shifts from a server to the shipped hardware. That relocation changes exactly which details an importer must document — the AI accelerator, its memory, its RF/EMC behaviour, and its always-on power draw.

For product details and project planning, see What IP65 actually means for outdoor kiosks · Wintouch.

On-device AI vs edge AI: which declaration applies to your device

The two terms describe different proof points for a conformity file.

On-device AI	Edge AI
Inference runs on the same unit that hosts the sensor and use case, typically the display’s own NPU	Inference runs on a nearby local endpoint such as an edge box, gateway, or local server
Declaration attaches to the shipped SKU	Declaration may attach to a separate endpoint device

The decision rule is straightforward: if the on-device AI inference runs on the touchscreen’s built-in SoC, the on-device AI CE marking sits in that display’s dossier; if it runs on a separate edge box, certify that endpoint on its own. Either way, the on-device AI compliance requirements still oblige you to state where inference executes.

What a CE/FCC/UKCA dossier must now disclose for an on-device AI device

The core declared section should itemise four disclosure lines:

NPU/SoC architecture — the specific AI-capable system-on-chip and its neural processing unit, since suitability is bounded by the processing and memory budget of the hardware (^[1]).
Memory and storage — declared RAM and the storage reserved for model weight files and updates.
RF/EMC behaviour of the AI accelerator — the electromagnetic and radio behaviour of the added silicon within the whole unit.
Energy consumption for always-on inference — the power draw declaration for a device that infers continuously.

Note that an NPU does not exempt a unit from standard electronics and EMC certification; ^[3] states that CE, FCC, and outdoor IP/IK ratings still apply to the whole unit. These four on-device AI compliance requirements sit alongside, not instead of, that base certificate.

Memory and SoC declarations when models run on-device

Memory and storage play a central role in enabling edge AI because they determine how quickly and efficiently data can be processed on-device (^[2]). In practice, ^[3] recommends specifying minimum RAM of 8 GB for light analytics and 16 GB or more for video AI inference, plus storage for model updates and over-the-air firmware. A vision model can need 2–8 GB just for inference buffers, so the declared memory figure must match the shipped configuration rather than a marketing baseline.

The related memory-and-soc-declaration guidance in this series covers how to document RAM, storage, and the accelerator layer as separate conformity lines, which is where these on-device AI compliance requirements belong.

High-risk AI technical documentation: when more than CE/FCC applies

Beyond basic electronics marking, on-device AI used in a high-risk category triggers a conformity assessment and formal technical documentation. The EU AI Act treats Annex III uses — employment decisions, education and vocational training, credit scoring, and similar — as high-risk (^[4]). Those systems require risk management, data governance, human oversight, and accuracy testing, culminating in a CE marking applied to the AI system itself before market placement.

The decision rule for your edge AI certification requirements: if the device only infers locally for display or analytics, CE/FCC electronics certification suffices; if its model informs a high-risk decision, you add the AI Act technical documentation on top. These on-device AI compliance requirements then apply to both the product and the system it supports.

Energy-consumption declarations for always-on AI devices

An always-on AI device changes the power-draw profile a declaration must capture. An energy consumption declaration for a touchscreen that idles between user interactions assumes a low standby load; a kiosk running continuous inference draws power every second, not only on interaction. The declared figures must reflect that sustained load rather than the display’s nominal idle rating. For importers, the practical route is to pull the measured always-on figure from the power-consumption-declarations guidance in this series and fold it directly into the CE/UKCA file.

Your on-device AI compliance checklist for 2026

Confirm the NPU/SoC architecture is declared for the shipped SKU.
Verify minimum memory and storage match the model size actually deployed.
Include RF/EMC behaviour of the AI accelerator in the electronics file.
Add an always-on energy declaration for continuous inference.
Prepare high-risk AI technical documentation if the use falls under Annex III.
Map the destination market — CE for the EU, UKCA for Great Britain, FCC for the US — for each SKU.

Use the regional regulatory certification mapping in this series to confirm which mark applies, then align the other five declarations to it. Working through the list in order turns these on-device AI compliance requirements into a one-pass review of the conformity file.

Frequently asked questions

How is on-device AI different from edge AI for certification? On-device AI runs inference on the delivered unit itself, so the declaration belongs to that shipped SKU. Edge AI runs on a nearby box or gateway, which may need its own certification. If the model executes on the display’s own NPU/SoC, the display dossier carries the on-device AI CE marking.

For a practical vendor example, readers can review Outdoor LED Displays for Transit & Smart City Projects · Wintouch.

When does high-risk AI technical documentation apply beyond CE/FCC? When the on-device AI model informs a high-risk decision such as employment, education, or credit scoring, the EU AI Act requires a conformity assessment and technical documentation in addition to electronics certification (^[4]). Inference for local display or analytics generally stays on the basic CE/FCC file.

Does an NPU exempt a device from EMC or energy certificates? No. CE, FCC, and outdoor IP/IK ratings still apply to the whole unit, and the accelerator’s RF/EMC behaviour must be declared. The NPU adds disclosure lines for architecture, memory, and power draw; it does not replace the base certificates.

Content reviewed: 2026-08-12.

Evidence confidence

Confidence: Medium. This rating reflects cross-checking 4 sources across 4 independent domains. It measures evidence coverage, not certainty; verify safety-critical work against manufacturer instructions and local requirements.

References

APA 7th edition

↑Cited 2 timesSilextechnology. (2025). What kind of device is suitable for your on-device AI. https://www.silextechnology.com/unwired/what-kind-of-device-is-suitable-for-your-on-device-ai-inference-1.
↑Micron. (n.d.). Edge AI. Retrieved August 12, 2026, from https://www.micron.com/markets-industries/ai/edge-ai.
↑Cited 2 timesWintouchtech. (n.d.). Edge AI in Commercial Displays & Kiosks - Wintouch. Retrieved August 12, 2026, from https://wintouchtech.com/en/blog/edge-ai-commercial-displays-kiosks-beyond-soc.
↑Cited 2 timesModulos. (n.d.). AI Compliance Guide 2026: Global Regulations | Modulos. Retrieved August 12, 2026, from https://www.modulos.ai/ai-compliance-guide.