Computer vision systems are proving technically capable in manufacturing, but many rarely make it all the way to the factory floor. A review published in Sensors and indexed in PubMed Central found that 77 percent of computer vision implementations in manufacturing remain stuck at the prototype or pilot stage, even though detection accuracy frequently exceeds 95 percent. The review identifies limited training data as a major barrier to production deployment, particularly when systems encounter edge cases and defect variations that controlled pilot environments fail to capture.
The challenge becomes even more pronounced at the integration layer. A National Institute of Standards and Technology symposium report, Towards Resilient Manufacturing Ecosystems Through Artificial Intelligence, found that successful AI use cases in manufacturing often remain isolated, expert-dependent efforts that do not scale easily across equipment, facilities, or companies. The report also highlights the organizational and cultural challenges involved in adapting AI software to legacy manufacturing equipment.
Taken together, the evidence suggests that the biggest obstacle to scaling computer vision is no longer whether the technology works. The larger challenge is whether organizations have the infrastructure, ownership, and operational trust necessary to turn technically successful systems into everyday manufacturing capabilities.
Emerj recently hosted a three-episode series examining what separates computer vision deployments that reach production from those that remain stuck in pilot mode. The series featured Joseph Nelson, Co-Founder and CEO of Roboflow; Jeff Witt, a manufacturing IT leader responsible for computer vision programs spanning more than 100 production facilities; and Brian Ton, Senior Laboratory Manager at Florida Crystals Corporation.
Across all three conversations, the same pattern emerged. Technology is not the primary obstacle. What determines whether a vision AI program becomes embedded in operations is how an organization prepares its technology ecosystem, assigns ownership, and builds trust among the people who ultimately depend on the system.
Ecosystem Readiness Determines Deployment Success
Turning Computer Vision Into Real-World Value at Enterprise Scale
Joseph Nelson, Co-Founder and CEO at Roboflow, approaches computer vision deployment from the perspective of what happens after a model has been developed. A system can achieve impressive accuracy in a controlled environment and still fail to create meaningful business value if the organization is not prepared to collect the right data, develop a sufficiently specific model, and connect the resulting intelligence to operational systems.
The first requirement is data readiness. Before an organization can deploy a useful computer vision system, it needs cameras or sensors positioned to capture the specific information the business wants to monitor. This is fundamentally a physical and operational question before it becomes a machine-learning problem. If cameras cannot see the relevant portion of a battery, stamping press, assembly process, or product, the model has nothing useful to learn from.
The second requirement is model specificity. Although general-purpose AI models continue to improve, manufacturing environments remain highly specialized. Companies typically need models trained around their own products, defect categories, equipment configurations, and operating conditions. A vision system designed for one production environment cannot necessarily be transferred directly to another without accounting for those differences.
The third requirement is downstream integration. Detecting an issue is only the beginning. If a vision system identifies that a component is missing or a product does not meet quality standards but the resulting signal never reaches the manufacturing execution system, quality platform, inventory system, or operator responsible for taking action, the organization has gained intelligence without gaining operational value.
Nelson points to BNSF, a Class I railroad operating across roughly 30,000 miles of U.S. track, as an example of what becomes possible when these elements come together. Visual systems can monitor wheel conditions, track infrastructure, containers, and other assets at a scale that would be impossible for human teams to match. However, the value comes from connecting that visual intelligence to the operational systems responsible for scheduling maintenance, dispatching crews, and managing assets.
As Nelson explains:
“Any sort of enterprise change requires people, processes, and technology. The technology has advanced to the point where you can build systems that understand your business. What determines whether those systems succeed is whether organizations can connect their operational teams, their engineering groups, and their decision-making infrastructure. When those pieces come together, companies see rapid acceleration — not just in deployment, but in the outcomes that matter.”
This leads to a practical deployment philosophy that Nelson describes as a “barbell strategy.” Organizations need executive-level commitment to the broader potential of computer vision while simultaneously choosing a narrowly defined, concrete use case at the production-line level.
That first use case becomes the proof point. Instead of attempting to transform an entire manufacturing network at once, organizations can demonstrate value in a controlled environment and use that success to justify broader investment.
The lesson is straightforward: computer vision deployment requires more than an accurate model. Data collection, model development, operational integration, and organizational alignment must work together. When one of those components is missing, the technology can remain trapped in pilot mode regardless of how impressive its technical performance may be.
Business-Led Ownership Accelerates Deployment Where IT-Led Programs Stall
How Vision AI Scales Across a Manufacturing Network
Jeff Witt, a Digital Transformation Leader with experience overseeing computer vision and asset-health programs across more than 100 production facilities, has seen firsthand how organizational structure can determine the speed of AI deployment.
From his perspective, successful computer vision programs are not standalone software projects. They are end-to-end operational systems that must become part of the way a plant actually works.
One of the first challenges is architectural. Many manufacturing facilities already have cameras installed, but those cameras often operate within manufacturing IT networks that are separated from enterprise data pipelines and business intelligence environments. Connecting visual information with other operational data therefore becomes an essential early step.
Once that integration challenge is solved, however, the deployment process can become much more repeatable. Instead of engineering an entirely different solution for every facility, teams can build on a common infrastructure and apply the same approach across multiple plants.
Witt’s experience also highlights the importance of starting with infrastructure that already exists. Rather than requiring every facility to install a completely new camera network before beginning, organizations can often begin by using existing cameras and demonstrate value with the infrastructure already available.
Once employees see meaningful results, the conversation changes. Plants begin requesting additional cameras, higher-resolution equipment, and broader coverage because they have already seen how visual AI can improve operations.
That visibility is particularly valuable for change management. Unlike some enterprise technologies whose benefits can remain abstract for months, computer vision produces something employees can see. Operators can watch an alert appear, review the associated footage, and understand why the system identified a particular event.
Witt describes this advantage by saying:
“The visual nature of these systems gives us a built-in advantage: People can see the alerts, see the AI processing the images, and see exactly how decisions are being made. That makes change management easier because the value is visible rather than abstract.”
This visibility also changes expectations around model perfection. Manufacturing organizations sometimes assume that an AI model must reach near-perfect accuracy before it can provide useful value. Witt’s experience suggests otherwise.
Initial models can begin producing useful outputs quickly, while human oversight manages the remaining uncertainty. The objective is not necessarily to wait until a model becomes perfect, but to create a system that can provide meaningful operational visibility while continuing to improve.
The larger organizational shift occurs when ownership moves closer to the business. Witt has observed that computer vision programs can accelerate significantly when plant and business-unit teams are empowered to define use cases, deploy models, and adapt systems without requiring IT to act as a gatekeeper for every change.
This does not eliminate IT’s role. Instead, IT becomes responsible for providing the underlying platform, integration, security, and infrastructure while operational teams take greater responsibility for determining how the technology should be used.
That distinction is critical. Manufacturing employees understand the process, the equipment, the failure modes, and the operational priorities better than a centralized technology team ever could. Giving those employees the ability to work directly with a flexible computer vision platform allows new use cases to emerge much faster.
Witt summarizes the concept by describing computer vision as a platform rather than a point solution. Once the underlying infrastructure is in place, organizations can build different applications around it instead of treating every new problem as an entirely separate technology initiative.
The result is a shift from isolated projects toward a reusable operational capability.
Operational Trust in Visual AI Is Earned Through Small Wins
Making Visual AI Standard Practice in Complex Manufacturing
Brian Ton, Senior Laboratory Manager at Florida Crystals Corporation, offers another perspective on why computer vision deployments succeed or fail. In his experience, a project can clear the technical proof-of-concept stage, generate enthusiasm within the organization, and still eventually stall.
The technology worked. The deployment did not.
The underlying problem is often that the organization has not adequately involved the people who will actually use the system. Operators, technicians, laboratory staff, and quality professionals possess detailed knowledge about how processes behave during a real production shift. They understand the edge cases, environmental conditions, workflow interruptions, and practical constraints that may not appear in technical specifications.
When these subject-matter experts are brought into the project only after development is complete, important requirements can be missed. A system may technically perform its intended task while failing to fit the reality of the production environment.
Ton emphasizes the importance of placing subject-matter expertise close to the development and deployment process. The people who understand the process should help define what the system needs to detect, how alerts should be presented, and what should happen when the system produces an incorrect result.
This involvement is also essential for building trust.
Manufacturing employees are unlikely to embrace an AI system simply because management says it is accurate. They need to see that the system understands their workflow, responds to their feedback, and improves when problems are identified.
That requires a functional feedback loop. Operators need a practical way to flag incorrect detections, identify missing scenarios, and communicate where the system does not behave as expected. More importantly, they need to see that their feedback results in actual changes.
Without that loop, employees can quickly conclude that the AI system is disconnected from the realities of their work.
Ton describes the importance of this relationship by noting that the closer subject-matter experts are to the solution, the stronger the resulting trust becomes.
The same principle influences how organizations should choose their first use case.
Manufacturers do not necessarily need to begin with their most complex quality-control challenge. In fact, starting with an enormous problem can make it difficult to determine whether the deployment is actually succeeding.
Instead, Ton recommends focusing on small, manageable victories that are easy for employees to understand and whose benefits can be clearly demonstrated.
As he explains:
“Small, manageable victories — being able to build a little bit of credibility with something that’s easy and makes sense — go a long way. And when that credibility is established, you can start thinking bigger. You can go from line level, to site level, to multiple site level, to entire enterprise level.”
This creates a progression in which trust compounds alongside deployment.
A successful line-level implementation provides evidence for a site-level rollout. A successful site deployment creates the confidence needed to expand across multiple facilities. Eventually, those individual successes can become the foundation of an enterprise-wide computer vision program.
The productivity potential can also become significant once trust is established. Automated visual inspection and data processing can operate at a scale far beyond what manual teams can reasonably achieve. But technology alone does not unlock that productivity. Organizations first need to establish the conditions under which employees are willing to rely on it.
The Path From Pilot to Production
The experiences shared by Nelson, Witt, and Ton point to the same conclusion: scaling computer vision is less about proving that AI can recognize an object or identify a defect and more about building an environment in which that intelligence can become operational.
Nelson emphasizes ecosystem readiness. Cameras must capture the right information, models must reflect the specific manufacturing environment, and visual intelligence must connect to downstream systems that can act on it.
Witt highlights ownership and repeatability. Computer vision becomes much easier to scale when manufacturing teams can take ownership of use cases while a shared technology platform provides the infrastructure needed to deploy them consistently across facilities.
Ton focuses on trust. Operators and subject-matter experts need to participate directly in the development process, and organizations need to establish credibility through focused deployments before attempting broader transformation.
Together, these perspectives reveal why so many computer vision projects remain trapped in pilot programs despite strong technical performance. A model can be accurate and still fail to become part of everyday operations. A camera can detect a defect and still provide little value if nobody receives the alert. A successful proof of concept can still disappear if the people expected to use it do not trust the system.
The organizations most likely to scale visual AI will therefore be those that treat computer vision as an operational capability rather than a standalone technology project.
The path begins with a clearly defined business problem and the infrastructure required to capture useful data. It continues by connecting AI outputs to the systems and people responsible for acting on them. It accelerates when ownership moves closer to the teams that understand the operation. And it becomes sustainable when employees can see the value, provide feedback, and develop confidence in the system.
Computer vision may already be technically ready for the factory floor. The remaining challenge is making the organization ready for computer vision.