Why factories need an AI orchestration layer, not more apps

Ask most plant managers what has changed on their factory floor in the past five years and the answer is, in one sense, everything. Machine vision watches every weld, digital twins simulate every changeover, and industrial IoT sensors report on every motor bearing before it fails. And yet, in another sense, very little has changed at all.

Each of those systems still speaks its own language, to its own screen, for its own narrow purpose. A plant manager wanting to know why output on Line 3 slumped this morning is still, more often than not, stitching the answer together from five different applications rather than receiving it from a single source.

This is the gap that Rae Jeong, co-founder and chief executive of Maneva, has built his company to close. Rather than adding a further point solution to an already crowded stack, Maneva has positioned itself as an orchestration layer, a system that sits above the cameras, sensors, MES and SCADA platforms already installed on a line and coordinates the AI agents working across them.

It is a modest-sounding claim with an immodest implication: that the real bottleneck in factory AI is no longer the intelligence of any single tool, but the absence of anything capable of making all of them talk to one another.

A young research engineer's scepticism of the desk-bound model

Jeong's route to this conclusion runs through territory unusual for a Silicon Valley chief executive. He began his career welding in Alberta and working as a semi-truck mechanic before becoming one of the youngest research engineers at Google DeepMind, and he is candid about how that combination shapes the way Maneva operates. "We walk onto a factory floor to learn first, not to hand someone a dashboard and explain their own operation back to them," he says.

His diagnosis of where AI developers go wrong in manufacturing is without equivocation. Too many, he argues, treat the factory as a data problem that can be solved remotely, when the more valuable expertise sits with the person who has run a line for two decades and can sense a fault before any sensor registers it.

BMW's Woodruff plant readies AI-led battery assembly

Ahead of December's series launch, BMW Group's new Woodruff facility in South Carolina pairs artificial intelligence with virtual reality training to deliver zero-defect Gen6 batteries for the fully-electric BMW iX5.

"The real expertise lives with the person who's run a line for twenty years and can tell you something's off before a sensor ever flags it," Jeong says. "We build around that knowledge instead of assuming a model can replace it, which is also why our teams have full-time manufacturing veterans working alongside the engineers, not just engineers guessing at the floor."

That insistence on shop-floor grounding is not merely a matter of company culture. It informs Maneva's central technical argument, that automotive manufacturers have already spent heavily on the individual pieces of a smart factory, from machine vision and next-generation robotics to digital twins and industrial IoT, without gaining the coherent operational picture those investments were meant to deliver.

The case against standalone applications

Jeong's argument for orchestration begins with what each existing system fails to do alone. "Every one of those systems gives you a piece of the picture, and none of them talk to each other well enough for a plant manager to act fast," he says.

"A standalone AI application can tell you a machine is jammed, but it can't tell you why output on the whole line is down or how that jam connects to what happened upstream an hour earlier."

Maneva's own path to that conclusion was incremental rather than designed from the outset. The company started by deploying autonomous agents to solve single problems on a line, and only once enough of those agents were live did the case for tying them together become apparent. The result is Orchestration, Maneva's platform, alongside a natural-language interface called Kaizen AI that lets a plant manager pose a plain question, such as what slowed Line 3 down that morning, and receive an answer grounded in live data rather than five open screens.

Crucially, Jeong stresses, the system plugs into the MES and SCADA infrastructure a plant already has rather than requiring it to be ripped out. "That's the difference between one more point solution and something that acts like a single unified operation," he says.

The pitch lands at an opportune moment. AMS has reported that other manufacturers, including Mercedes-Benz, are experimenting with their own multi-agent systems designed to trace root causes and propose actions across quality, assembly and logistics functions within minutes, evidence that the appetite for coordination across AI tools, rather than merely more of them, is spreading well beyond one vendor's pitch deck.

Where autonomy stops and judgment begins

If orchestration is the architectural argument, the harder question for any automotive customer is where to draw the line between autonomous decision-making and human oversight, particularly for safety-critical production. Jeong's answer distinguishes sharply between speed and ambiguity. Fast, repeatable, well-defined calls, he argues, belong entirely to the machine.

"We build for full autonomy on fast, repeatable calls, like a quality agent rejecting a defective part thirty times a second at a junction point on the line," he says. "If that decision still needed a human in the loop, it would be too slow to matter and the ROI wouldn't be there." Safety decisions are treated with more caution. Maneva's approach is to alert the worker first, through an audible cue akin to a lane-departure warning in a car, and to escalate to a supervisor only once a pattern has repeated several times.

Toyota's AI push to unify a fractured production language

Toyota is deploying artificial intelligence to reconcile 45,000 internal terms into 5,000 common ones across planning, production and sales, aiming to cut translation steps by 30% and speed vehicle delivery by 2028.

Jeong is explicit that automotive production raises the stakes here, given the traceability and quality obligations attached to parts destined for the road, and that the thresholds for escalation are agreed jointly with each plant team rather than imposed from outside. "Judgment, accountability, and anything with real ambiguity stay with people," he says. "The split-second, well-defined call is where the agent takes over."

That same emphasis on staged trust extends to how Maneva's agents learn. Because they improve continuously through reinforcement and active learning, a technique that unsettles an industry accustomed to deterministic production systems, the company sets safety and quality thresholds with each plant before an agent runs live, and streams video alongside every alert so a supervisor can verify what the agent saw rather than take its judgement on faith.

Trust, in Jeong's telling, is earned the way a new operator earns it. "We also start from the simplest, most provable unit of value, a single quality check at one junction point, before expanding scope, so the system earns trust the same way a new operator would, by being right, being visible, and being consistent over time."

Why pilots stall and scaled deployments do not

AMS has long tracked a persistent pattern in factory AI, in which vision and analytics initiatives launch with fanfare and then stall at the pilot stage, never reaching a second line, let alone a second plant. Jeong's explanation for why some manufacturers break out of that cycle while others remain trapped in it has little to do with the sophistication of the underlying model. "It comes down to people more than the model," he says.

Sponsored

This is why AI inspection fails long after the pilot succeeds

A recent peer-reviewed meta survey found that 77% of AI vision automotive production pilots never reach full deployment. A.J. Camber, VP at Solidigm, argues the problem is rarely the tech - and what it really takes to go from controlled trial to plant-wide reality.

Maneva staffs a continuous improvement team of manufacturing veterans, people who have spent twenty or thirty years as plant managers, working full-time at the company rather than as occasional consultants. Their role is to roll deployment out one line and one facility at a time instead of attempting to convert an entire plant at once, and their credibility with customers rests on having lived the work themselves.

"They speak the customer's language because they've lived it, and that's what earns the trust to expand from one deployed agent to ten or fifteen," Jeong says. Manufacturers who remain stuck in proof-of-concept mode, in his view, tend to treat the pilot as an endpoint rather than the first stage of an ongoing partnership, and have not built the internal capability to prove out return on investment at each subsequent stage before asking to expand further.

AMS's own reporting on the industry's AI adoption curve has noted a comparably steep failure rate among pilots more broadly, underscoring that the organisational discipline Jeong describes is scarce rather than standard practice.

Humanoid robots as one endpoint among many

No conversation about physical AI in 2026 avoids the subject of humanoid robots, and Jeong's answer is notable chiefly for how little urgency he attaches to it. He has "a lot of respect" for the teams working on humanoid platforms, he says, but regards a genuine fleet operating in manufacturing at scale as still five or more years away, and sees no problem in that timeline.

He draws a comparison with the years of hype that preceded Waymo's actual scaling of self-driving vehicles, arguing that early demonstrations of a technology can be healthy for an industry even before that technology is ready for volume deployment.

For now, Jeong locates most of the value elsewhere, in autonomous agents running on the fixed cameras and sensors already watching product and machinery in real time, a view that aligns with AMS's coverage of manufacturers such as Shanghai Electric, which has likewise framed embodied robots as one expression of a broader industrial AI architecture rather than its centrepiece. Robots, smart glasses and autonomous yard vehicles, in Jeong's framing, are simply further endpoints on the same orchestration layer, rather than the transformation itself.

"The layer that understands what's happening across a facility is the constant," he says. "The form factor acting on that understanding, whether it's a fixed camera, a robot arm, or a pair of glasses, is what will keep expanding."

Shanghai Electric expands AI robotics for industrial manufacturing

Shanghai Electric has unveiled a new generation of 'embodied' AI robots, industrial AI agents and an AI-native smart factory architecture at WAIC 2026, signalling a broader push to embed autonomous intelligence across manufacturing operations.

From automation to something that learns

Asked to look five years ahead, Jeong resists framing the future purely in terms of hardware. The technology, he argues, is still in its first wave, one in which AI mostly performs tasks a human already knew how to do, only faster and more consistently. The distinction he draws between today's highly automated plants and a genuinely autonomous factory of the future is subtler than robot headcount.

It rests on whether enough agents are tied together that an operation begins to accumulate what he calls a living memory, in which what happened on a given line on a particular shift informs what the system watches for afterwards, across every subsequent shift, without anyone having to notice the pattern manually.

The capability he believes manufacturers most consistently underestimate is not robots displacing workers on the line, but this compounding effect, root cause analysis that sharpens the longer a facility runs, paired with a system able to act on a goal a plant manager sets rather than simply alert on a threshold someone configured once. "That's the shift from automation to something that's actually learning," he says.

It is a claim that will be tested, plant by plant, over the coming years, and one that puts Maneva's founder in the somewhat unusual position of arguing that the industry's AI problem is, for now, less about building smarter individual tools than about persuading the tools already installed to finally speak to one another.