Stuart Hubbard of Zebra Technologies discusses how edge AI, small language models and AI-ready devices are helping enterprises improve security, cost control and frontline productivity
Enterprises are moving from AI experimentation to scaled adoption, with growing attention on architectures that can deliver intelligence securely, efficiently and closer to where work happens. For frontline-intensive sectors such as retail, logistics and manufacturing, on-device AI is emerging as a practical way to reduce latency, manage costs, strengthen data security and support workers in high-volume operational environments.
In this interview, Stuart Hubbard, Senior Director, AI and Advanced Development at Zebra Technologies, discusses how small language models, edge devices, AI-ready processors and real-time data capture are making AI more accessible for the frontline. He also shares his perspective on the evolving balance between cloud and on-device AI, and what enterprise leaders should prioritise as they shape their AI strategies.
Interview Excerpts
Why is on-device AI emerging as a more sustainable approach to scaling AI adoption?
Businesses in sectors like retail, manufacturing, and logistics realise that small language models (SLMs) running on device are very well-suited for a range of high-volume tasks carried out by frontline workers on a daily basis. AI helps solve labour training issues, and fill gaps as hiring and retaining labour remains a challenge.
May 2026 research by the S&P found that organisations continue to identify major challenges around generative AI. The most commonly identified challenges relate to data privacy (38%), security risks (38%) and costs (37%). On-device AI goes some way towards overcoming these challenges.
SLMs can be highly tailored to specific jobs and sectors, so data is accessed quickly and intelligence is reliable and up to date. On-device AI is also particularly suited to high-volume, fast-paced working environments and locations with limited or no connection to the cloud, such as rural settings, retail stores, and certain areas within factories and warehouses. On device eliminates latency as there is no data going to and from the cloud, which is important in environments where speed matters.
Edge AI means its capabilities can be scaled across fleets of devices, cameras, and sensors in physical environments, and existing IoT data capture and device management software can provide new use cases and sources of data for AI.
What are the biggest financial, operational and security advantages organisations can expect by shifting AI workloads from cloud to the edge?
Financially, businesses can save on token APIs, as on-device AI means no data needs to leave the device. Data capture and AI inference happen on-device in the moment, whether in a retail aisle, doorstep delivery, or warehouse picking zone. We’ve seen headlines about ‘tokenmaxxing’ as a proxy for adoption and ROI from AI, and the spiralling costs it has caused, which budget holders want to avoid.
Relatedly, as no data leaves the device or IT network, businesses gain an additional layer of security, with concerns around end-to-end data encryption, API security, identity and access management removed or reduced in relation to the cloud. And where cloud is needed, such as more complex AI agent chain of actions, an edge server or hybrid architectures can provide the security and compute needed.
However, businesses need to invest in workforce devices that are AI-ready, with the on-device architecture needed for AI capabilities, and appropriate endpoint cybersecurity with things like Secure Element, Android Strongbox and timely OS updates. IT and operational leaders should be thinking about AI readiness, security, fleet management and total cost of ownership rather than cheap per device unit costs at the start.
AI is becoming a new type of infrastructure underpinning revenue-generating work on the frontline. Elevating operational work and customer engagement using AI makes workers more autonomous and efficient at getting more done per shift. Customers get reliable service and personalised invitations for additional purchases. Inventory management visibility means shrink can be detected and fixed and over or understocking avoided.
How are advances in AI models, processors and edge devices making real-time, on-device AI practical for industries such as retail, logistics and manufacturing?
Today’s mobile computers come fitted with a robust system-on-Chip (SOC) with a dedicated neural processing unit (NPU), GPUs and CPUs, providing the power to make advanced AI capabilities a reality for on-the-go workers. Users can run SLMs for translation or transcription, utilise AI computer vision models to examine shelves, products, labels, and text, 3D measurement tools and enable generative AI assistants.
Zebra is an industry leader in enterprise mobile computing and partnership with Qualcomm means our mobile computers come fitted with the AI-enabled chips like the Qualcomm® Dragonwing™ Q-6690 processor (up to 2.9 GHz).
AI models and agents on-device are surrounded by exceptional data capture sensors for multimodal AI. The physical environments are digitised across workers, machines, interactions, assets, inventory and working conditions. A living digital twin of the frontline is becoming a reality, what we call ‘ambient intelligence’ where agents act upon the data and within the digitised environment.
These features could include a microphone for natural language interface or an enterprise AI camera app that prompts users to clean the lens, blur personal data and faces, and add a precise location with a timestamp to verify on time tasks. Scanners, OCR, 3D time-of-flight sensors, and RFID capture barcode, tag, inventory, and text data, while voice-enabled AI and noise cancellation and an easy to use interface add up to a powerful solution in the hands of a frontline worker.
How can organisations leverage their existing devices and data to unlock greater business value through AI?
Organisations need to invest in the hardware needed to run AI models from a partner who can also provide the endpoint security needed. Once the hardware is in place, businesses can work with providers on use cases and developing AI applications.
That might be a swarm of AI agents for tasks around merchandising, sales and customer questions, or employee onboarding and training, or working within in-house developers who need AI components and templates like AI data capture SDKs, sample applications, and APIs.
Zebra’s AI blueprints are ready-made, real‑world AI frameworks that modernise manual work, increase accuracy, and accelerate productivity across high‑pressure environments.
AI enablers allow users to optimise workflows, streamline processes, and simplify complex tasks, including industry specific, trained AI vision models, ready-to-use sample applications, and Application Programming Interfaces (APIs) to integrate AI-enabled apps with their enterprise applications. It means buying and creating AI applications are both options.
In terms of data, businesses are finding new value from their current IoT, software stack, and data capture technologies, with data management and quality remaining important. Machine performance and operations, defects in goods, labels and barcodes, RFID tags, inventory and sales lists, SOPs, worker-to-worker and worker-to-customer interactions are now valuable sources of data for agents, on-device AI, computer vision and deep learning for training, workforce and workflow improvement, and predictive operations. Digitised physical operations, software applications and frontline IoT data is environment for agents to act upon and within.
Looking ahead, how do you see the balance between cloud-based AI and on-device AI evolving and what should enterprise leaders prioritise as they shape their AI strategies?
Both are needed. On-device has definitely grown in importance, with 2026 seeing a range of product launches aimed at leveraging on-device AI and there have been some widely publicised lessons learnt about cost control, the need for a culture that drives adoption, and meaningful measures of success.
Cloud, whether private, edge servers, or hybrid architectures are also here to stay, giving organisations the choices they need. Not all data requires the same level of security, and more complex AI inference requires cloud compute. We can see how much capital is currently being invested into hyperscaler data centres, while some are already talking about data centres in space.
Enterprise leaders need to prioritise developing businesses cases with clarity around use cases, metrics of success and ROI, realistic timelines, culture change. A report from MIT’s NADA found businesses that partner with AI providers had a success rate of 66% compared to those who preferred internal development alone (33%). While Standford’s Digital Lab research concludes that the top factor (43%) for AI success was executive sponsorship and a culture of experimentation and permission to fail.
Source: Tahawul Tech