A roughly 30-minute panel at Advantech’s 2026 Edge AI Conference argued that the gap between promising pilots and full production rollouts is where the real engineering work begins. Panelists Ed Doran, PhD (VP of Strategy at the Edge AI Foundation), Umang Garg (Managing Director at Nagarro) and Richard Huang (Chief Software Architect at Advantech) cut through the hype to map practical challenges and architectural choices for real-world deployments.
Edge AI in constrained environments
Doran framed edge AI not simply as “AI running somewhere other than a data center” but as the extension of centralized intelligence into environments defined by hard constraints: limited bandwidth, tight power budgets, strict latency requirements and security boundaries that can make cloud connectivity impractical or unacceptable. He offered concrete examples — the shipping container packed with inference hardware outside a mine in Western Australia, a vehicle making safety-critical decisions with a degraded radio link, and a surgical suite where patient data cannot leave the room — to show how each use case demands different trade-offs. “It’s really hard to do edge AI in that shipping container,” he said. “It’s really vital to do edge AI well in the surgical suite.” The panel summed this up: edge AI is not one problem. It’s dozens of constrained optimization problems sharing a name.
The conversation moved from passive inference to physical AI — the convergence of edge AI with robotics, digital twins and autonomous systems that close the loop between perception, decision-making and real-world action. Physical AI raises stakes where low latency and low tolerance for error are essential, and it shifts architecture discussions toward embodied intelligence rather than research demos.
From pilots to production: failure points and architectural fixes
Garg drew on Nagarro’s industrial deployments across more than 20 countries and said the barrier to scale is empirical. His team has catalogued more than 60 distinct failure points between proof-of-concept and production. Common challenges include ROI modeling, feasibility assessment and scalability planning, but deployments most often fail because of solution strategy and enterprise integration: the choice between custom software, off-the-shelf tooling or configurable accelerators is an architecture decision that determines whether a solution can scale from one site to 50.
Garg illustrated the gap with a machine manufacturer case study. The customer, in business for 400 years and operating roughly 250 machine types with field lives of 25–30 years, often lacked modern sensors on machines deployed in the last decade, forcing engineers to travel for diagnostics. The deployed solution layered edge-based predictive analytics with a parallel stream of real-time telemetry and used AI throughout development, compressing a six-month schedule to about six weeks. The rollout now covers 11 machine types across 200 customers, with predictive maintenance running entirely on the local edge device.
On the architecture and platform side, Huang detailed the cross-silicon challenge: edge deployments target Intel, NVIDIA, NXP and Rockchip among others, and moving an inference application between them is difficult without an abstraction layer. Advantech’s WISE platform (Wireless IoT Sensing Embedded) addresses that need with containerized, pre-validated AI workloads that absorb platform-migration overhead. WISE also extends traditional OT data handling beyond numeric streams to support image data, binary payloads and tagged datasets via digital twin protocols, so the same data can be consumed on the edge or passed to the cloud without re-engineering pipelines. Model lifecycle management and zero-touch device onboarding further reduce infrastructure friction that often consumes development cycles before application logic is written.
The panel described the AI agent layer sitting atop that stack: Advantech’s approach maps edge AI development architecture to vertical-specific AI agents for factory, energy, healthcare and retail environments. The agent layer runs on NVIDIA NeMo, an agent-first open suite of libraries with built-in skills to accelerate agent specialization, optimization and governance. NeMo integrates with existing AI tools and agent frameworks to optimize specialized agents across cloud, on-premises or hybrid environments, matching the deployment diversity Doran outlined.
The session closed with Doran’s blunt framing: “intelligence without action has no value, and action without intelligence is a liability.” Edge AI enables agents to act with low latency and high reliability where cloud roundtrips are not viable, and the outcomes those agents produce will feed the next generation of edge models. “The loop is already running,” the panel said. The remaining challenge is building agents and platforms stable enough to be trusted in the field; the panelists agreed that no tooling yet fully solves every problem, but the pieces appear closer to assembly than they were two years ago.
