This is why AI inspection fails long after the pilot succeeds

The computer vision model passed every test. Detection rates were high, false positives were manageable, and the accuracy figures looked exactly as promised. Six months into production deployment, the system was quietly switched off. This is not a hypothetical. Across automotive manufacturing, seventy-seven per cent of AI vision pilots never advance beyond the trial stage - not because the models fail under controlled conditions, but because controlled conditions are not what factories are.

A.J. Camber has spent the better part of three decades in enterprise data infrastructure and semiconductor storage - Intel, Numonyx, Micron, and now Solidigm, where he leads the AI Software Business Group. He arrives at manufacturing AI not from the factory floor, but from the infrastructure layer beneath it: the storage systems that now underpin everything a modern plant generates.

"The models work," he says, without preamble. "That's not the issue." The issue, as Camber sees it, is a compound problem: a chronic shortage of data expertise, an institutional tendency to build AI capability around a small number of specialists rather than distributing it across the workforce, and a fundamental misunderstanding of what it means to deploy artificial intelligence in an environment that changes constantly and without warning. "As you move from pilot to production, it becomes less about picking the technology and more about staying flexible," he explains. "The things that work are modular architectures combining cloud and edge. The things that don't work are monolithic systems that are hard to change."

The logic of a data company

Solidigm - a business whose identity is built on enterprise solid-state storage, and which holds records for some of the world's innovative highest-capacity drives - has developed a manufacturing AI platform might seem, at first glance, counterintuitive. Yet the company's Luceta AI software suite is aimed squarely at the shop floor. But the logic behind the move, Camber insists, is not a departure from Solidigm's heritage. It is a direct extension of it.

"This massive volume of new AI visual data has a large edge storage impact, and customers benefit from acting on the data where it is created because of the difficulty associated with the time, energy and costs related to moving it.” he says.

"We've always been trusted to protect customer data and help people do whatever they need to do with that data. Once you have stored the data, helping customers find the most useful aspects within their data is a logical extension of where we've been."

What Solidigm saw from its position deep within industrial data infrastructure was the early stages of a data explosion without historical precedent. Camber says that if we imagine a factory with 200 inspection points, with two cameras at every point generating approximately 40 terabytes each of visual data per day, at scale across a full operation, the factory is generating 16 petabytes of video daily. To put it in context: a decade ago, the entire ERP database of many mid-sized manufacturers might have totalled 4 terabytes. The magnitude of the shift in data density is not incremental. It is a categorically different problem.

This explosion of these data sets, combined with the vision that the emergence of many new use cases will be rooted in physical AI, shaped the Solidigm move into software. The visual dataset you build today is the asset that compounds. It is the foundation layer. Everything that follows - robot path planning, imitation learning, autonomous interaction with the physical environment - depends entirely on the quality of the visual datasets being built today.

A skills gap with no easy fix

The storage challenge is, in principle, solvable through hardware advancement. The talent problem is considerably more stubborn. There are, by most estimates, approximately one million data scientists in the world. There are far more organisations than that with ambitions to deploy production-grade AI. The stats do not work in manufacturing's favour - particularly not when those organisations are competing for the same pool of expertise against technology companies offering salaries that traditional industrial employers cannot easily match.

"Not everybody can hire enough data scientists to support their company," Camber says, flatly. His response to this is not to lower the bar on model quality - a legitimate concern, given that the output of a faulty inspection system is a defective component continuing downstream - but to reduce the fundamental dependency on specialist expertise in the first place.

Solidigm's approach is built on three principles. The first is ease of use: making AI tooling accessible to engineers and technicians whose expertise lies in manufacturing processes rather than machine learning.

The second is flexibility: building platforms capable of adapting to changing conditions, including the inevitable proliferation of product variants and the unpredictable variability of real production environments as opposed to controlled pilot conditions.

The third, and perhaps most consequential, is systematic iteration - ensuring that the platform behaves predictably and transparently enough that people beyond its original developer can understand, trust, and update it over time.

"That third point is important because it helps teams build trust that the technology is actually doing what they think it's doing," Camber says. "It also helps avoid situations where work gets thrown over the wall to a few siloed experts." The expert dependency problem is not merely an efficiency concern, but more correctly, a fragility concern. Camber highlights that when AI capability is concentrated in two or three individuals, operational continuity becomes vulnerable to factors entirely outside an engineering team's control.

The high-mix imperative

The production environment into which these tools must be deployed is itself undergoing a structural transformation. Mixed-model lines running internal combustion and battery-electric vehicles simultaneously - a configuration now common across European and North American OEMs managing platform transitions - represent operating conditions that traditional machine vision was simply not designed to handle.

Conventional optical inspection is, by architectural design, a system of rigidly defined spatial expectations. A component must present itself at a precise location, under controlled lighting conditions, within a known orientation tolerance. In high-mix production, that logic unravels quickly. Part families rotate. Lighting changes across shift patterns. The defect configuration relevant to one variant may manifest differently on a model launched mid-year,or disappear entirely and be replaced by failure modes the original system was never trained to see.

"AI is much better suited to this kind of high-mix, low-volume environment than traditional optical vision systems," Camber says. "It isn't hard-coded spatially in the way traditional systems are, so you don't have those rigid breaking points."

He describes an early Solidigm customer - operating in a notably dusty factory environment, with no fixed orientation to the components under inspection, and a workforce of highly skilled industrial and mechanical engineers with no background in data science. “The team mounted high-resolution cameras at the inspection location and began building models within weeks,” he says. 

“There was no extended programming exercise to calibrate fixed camera positions against precise three-dimensional coordinates. The system adapted to the environment, rather than requiring the environment to be reconfigured around the system.” A fine example of why the democratisation of AI systems are now key to effective production environments.

At a time, and in an industry marked by accelerating model proliferation and platform consolidation, the ability to commission a new product variant without a full inspection system re-programme is not a convenience feature. That ability is now clearly an operational requirement - and increasingly, a competitive differentiator.

The problem that appears six months after launch

Beyond all this, there is a category of AI deployment failure that rarely appears in case study presentations. And it is not the failure that occurs at launch. Tellingly, it is the one that emerges six months later, when something on the production floor shifts - a process parameter adjustment, a different material batch, a seasonal variation in ambient temperature that subtly alters the lighting profile in the inspection zone - and the model degrades quietly, without anyone noticing until defects begin appearing further down the line.

Camber identifies this as “the central design challenge of production AI”, and it is not difficult to understand why. The ability of a system to maintain its performance across changing real-world conditions is precisely where the gap between AI as a demonstration and AI as production infrastructure becomes visible and costly.

"No matter how much data diversity you have, there will always come a point where the model needs updating," he says. "That's inevitable. What matters is having a system that can systematically adapt to changing conditions - one that creates the ability for more than just the original creator of the model to update it."

This principle is embedded in how Solidigm has structured its deployment architecture. One-click model deployment at the edge, built on containerised management systems that production engineers already work with, means that updating a model does not require calling in an external specialist or waiting for an internal data science resource to become available. The broader aim is institutional: building a manufacturing workforce that is AI-literate in the same way it is literate in the statistical process control and quality management disciplines that modern production operations depend on.

"The beauty of the world is that it constantly changes," says Camber. "So you need an AI tool flexible enough to adapt to flux and change, rather than one built on the assumption that conditions will stay static. Because, inevitably, they won’t."

Data selection, not just data collection

Then there is the seemingly recalcitrant assumption around AI discussions, that more data is always better. Camber is direct about why this logic breaks down in a factory context.

“Oversized datasets,” he says, “can introduce bias and overfitting, causing a model to miss exactly the failure modes that matter. They also produce larger, slower, more computationally expensive systems - harder to run at the edge, more energy-intensive in operation, and significantly more difficult to interrogate when performance degrades.

"The goal is not maximum data volume. It is appropriate data diversity, with sufficient variety to reflect the genuine range of conditions the system will encounter, and without the noise of unnecessary duplication.”

Data selection - identifying which images and samples should actually be included in a training dataset - is among the most technically demanding steps in the model-building process. For a skilled data scientist, Camber estimates the selection work for a single model consumes approximately 20 hours. At current market rates for that level of expertise, that cost approaches ten thousand dollars per model. For manufacturers building a new model each month, that overhead accumulates rapidly - and delivers no direct production value. It is the cost of getting to the starting line, and not of running the race.

Solidigm's software contends with this issue head-on, by automating the selection process, and running it transparently during model preparation. And the benefit does not belong exclusively to organisations without in-house data science capability. It applies equally to those that have it: automating the technically intensive, routine elements of dataset construction frees specialist time for the work that genuinely requires expert judgement - and for which ten thousand dollars is not wasted. “Without tools like this,” says Camber, “building a workforce that is genuinely AI and computer vision literate becomes a very difficult challenge.”

The physical AI horizon

Computer vision for quality inspection is not, in Camber's framing, the end-state of factory AI. It is the prerequisite for something considerably more complex: physical AI - systems that must not simply interpret the real world, but act on it, in real time, with real consequences for equipment, people, and product integrity.

The distinction from large language models is fundamental. "Physical AI is much harder than LLMs because it operates in the real world rather than a digital environment," he says. "It has to deal with dust, lighting, movement, gravity - all kinds of physical complexity. Mistakes are far less forgiving. They can cause damage or create safety issues."

Two application categories are emerging as the leading edge of this development. The first is autonomous path planning: the ability of robotic or automated systems to identify collision-free routes through dynamic environments, combining pre-mapped spatial data with real-time sensor responses to obstacles and changing conditions. The second is imitation learning, in which AI systems acquire tasks by observing expert operators performing them and replicating those actions directly - a different paradigm from deriving motion patterns mathematically, and one that dramatically lowers the barrier to teaching a system new capabilities.

Both categories share a common dependency on high-quality visual datasets. And that dependency is precisely where the urgency of the present moment becomes concrete. “The manufacturers positioned to deploy capable physical AI systems in five or ten years are those building those datasets now,” says Camber. “Factory-specific visual data cannot be sourced from the internet or licensed from a third party. It must be generated on the production floor, under real operating conditions, over sustained periods of time. You can imagine companies realising too late that they haven't started building their own datasets, while competitors have been collecting data for more than a decade. At that point, there's a real challenge around how to catch up to your competitors."

Starting where you are

For manufacturers assessing where they stand, Camber's message is more practical than visionary. The barriers are real - a global shortage of data science talent, the complexity of managing data quality at industrial scale, the difficulty of maintaining AI performance across an evolving production environment - but none of them are insurmountable, and most are more addressable today than they were even two years ago.

The entry point does not require a clean data estate, a dedicated AI team, or a completed dataset. It requires cameras at inspection points, a platform flexible enough to begin building and training models under genuine production conditions, and a willingness to iterate rather than wait for perfect initial conditions. For organisations that already have visual data but have not yet structured it, Solidigm can work with what exists. For those that have not yet started, the platform is designed to begin building the dataset from the first deployment.

"The companies that succeed will be the ones that continue adding capabilities over time," Camber says. "Future-proofing means having the ability to evolve."

The pilot is not the problem. It has never been the problem. The problem is designing, from the outset, for everything that comes after it.

 ~ or ~