AI tools in industrial settings have been hampered by the same fundamental problem: they need a cloud. Send a query, wait for a server to respond, get your answer. That model falls apart the moment you're on a platform in the North Sea, inside a chemical plant with restricted wireless access, or working a confined space entry where connectivity is patchy at best.
Google's Gemma 4 mitigates that issue. Gemma 4 is Google’s new AI tool that runs locally, entirely on-device, with no internet connection required. It’s available on iOS through the AI Edge Gallery app. Once downloaded, it works in airplane mode, underground, offshore, or anywhere else a signal can't reach.
Gemma 4's edge models (the E2B and E4B variants, meaning 2 billion and 4 billion parameters respectively) are purpose-built for mobile hardware. They run using Apple's Neural Engine on iPhones from the A16 Bionic chip onwards, which covers iPhone 14 Pro and newer.
Critically, all inference happens on the device itself. No prompts, no data, no images are transmitted anywhere. After the initial download, the model works with no network activity whatsoever. That is a meaningful distinction from every cloud-based AI tool currently marketed to industrial buyers.
For hazardous workers, this means a few things in practice. A technician in an ATEX Zone 1 or Zone 2 area can query a local AI model for procedure guidance without relying on a network connection that may not exist. A safety manager reviewing a permit to work on-site can use image recognition to cross-reference equipment states. An offshore inspection team can run document analysis on a confined space entry checklist without routing data through an external server.
Most people building AI tools for industrial use underestimate how unreliable connectivity is in real operational environments. Offshore platforms have bandwidth constraints and latency that make cloud-based AI tools sluggish or unusable. Underground utilities work and confined space operations frequently happen in areas with no signal at all. Refineries and chemical processing facilities often restrict wireless access for safety reasons, not practical ones.
On-device AI removes all of that friction. The model is downloaded once, stored locally, and functions independently of whatever network conditions exist on site. That isn't a nice-to-have feature for hazardous environments. It's the difference between a tool that works and one that doesn't.
Example applications:
The other dimension worth understanding is data handling. Cloud-based AI tools require operational data, queries, images, procedural details to pass through external servers. That creates problems for industries with strict data governance requirements, proprietary process information, or regulatory obligations around where operational data can be stored and processed.
With on-device inference, nothing leaves the device. A safety manager querying a procedure can be confident that information stays on the phone. An operator uploading an image of a pressure gauge for AI analysis isn't sending that image anywhere. For organisations operating under ISO 27001, GDPR, or sector-specific data regulations, that distinction matters.
This is where the conversation gets practical. Running Gemma 4 on an iPhone in a hazardous area is only viable if the device itself is rated for that environment. Standard iPhones are consumer electronics. They are not certified for use in ATEX or IECEx classified zones, where flammable gases, vapours, or dusts may be present.
Explosion-proof phone cases change that calculation. A case certified to ATEX Zone 1 or Zone 2 standards (and their IECEx equivalents) can make a standard iPhone a safe and compliant device for use in potentially explosive atmospheres. The intrinsically safe phone market has historically been dominated by ruggedised devices running limited software. An explosion-proof case that accommodates a current-generation iPhone capable of running on-device AI is a different category of tool entirely.
The combination is significant: a powerful, locally-running AI model on a device that can physically operate in the environments where that AI is most useful.
Consider a few scenarios that are realistic given this technology:
A technician in a Zone 1 classified area needs to check the correct isolation procedure for a piece of process equipment. With an explosion-proof iPhone running Gemma 4, they can query a locally-stored AI model, which has been loaded with relevant SOPs, without leaving the area, without needing a signal, and without that query being routed through an external server.
An inspection engineer photographs a valve or instrumentation component and uses Gemma 4's image recognition to identify the component, cross-reference it against known fault signatures, or check maintenance intervals, all on-device, in the field.
A safety officer compiling a confined space assessment uses audio input to dictate observations, with the AI model structuring and summarising them into a report format before they've left the site.
On-device AI on mobile hardware is moving fast. Gemma 4's iOS availability is recent; the ecosystem of tools built around it will expand. The practical performance on iPhones with A17 or A18 chips, which include the Neural Engine optimised for this kind of inference workload, will improve as model quantization and runtime optimisation advance.
For industrial buyers, the relevant question isn't whether on-device AI will be useful in hazardous environments. It clearly will be. The question is whether the hardware can legally and safely operate in those environments, and whether the deployment model, locally-stored models, controlled updates, managed devices, fits within existing IT and OT governance frameworks.
Can AI like Gemma 4 be used in ATEX classified zones?
Only if the device running it is housed in a certified enclosure appropriate for the zone classification. Standard consumer iPhones are not rated for ATEX Zone 1 or Zone 2 areas. An explosion-proof case, such as an Xshielder cover, that meets the relevant ATEX directive requirements is required before any mobile device, AI-capable or otherwise, can be used safely in those environments.
Does Gemma 4 work offline on iPhone?
Yes. After the initial model download, Gemma 4 runs entirely on the device with no network connection required. It has been confirmed to function in airplane mode, with no data sent to external servers during inference.
Which iPhones can run Gemma 4?
The E2B and E4B edge models run well on iPhones with the A16 Bionic chip or newer, which covers iPhone 14 Pro and later models. Devices with 8GB RAM handle both variants without significant performance constraints. The best performance you get from 16 Pro Max or 17 Pro Max versions.
Is on-device AI safe to use for safety-critical procedures in industrial environments?
AI tools, on-device or otherwise, should support and inform human decision-making in safety-critical contexts, not replace it. Any deployment in a regulated industrial environment should go through appropriate change management, validation against relevant procedures, and alignment with the site's safety management system.