Turning Visual AI into Enterprise Business Impact

​This article is sponsored by Roboflow and was written, edited, and published in alignment with our Emerj sponsored content guidelines. Learn more about our thought leadership and content creation services on our Emerj Media Services page.

Computer vision systems are proving technically capable in manufacturing but rarely reach the factory floor. A review published in the journal Sensors and indexed in PubMed Central finds that 77 percent of computer vision implementations in manufacturing remain stuck at the prototype or pilot stage despite detection accuracy frequently exceeding 95 percent. The review points to limited training data as a primary constraint on moving these systems into full production, particularly for the edge cases and defect variability that controlled pilot environments don’t capture.

The barrier compounds 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 remain isolated, expert-dependent efforts that do not scale to other equipment, facilities, or companies, and that adapting software to legacy equipment demands top-down leadership to overcome organizational and cultural barriers.

The evidence shows that the real barrier to scaling computer vision is organizational and infrastructural, not technical.

Emerj recently hosted a three-episode series examining what separates computer vision deployments that reach production from those that stall. The series features 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 Tan, Senior Laboratory Manager at Florida Crystals Corporation.

Across all three conversations, the same pattern surfaces: technology is not the primary obstacle. What determines whether a vision AI program becomes embedded in operations — or stays in pilot indefinitely — is how an organization handles three things: ecosystem readiness, program ownership, and the accumulation of operational trust on the floor.

This article examines insights drawn from each episode:

• Ecosystem readiness determines deployment success: The three most common bottlenecks to computer vision reaching production — data readiness, model specificity, and downstream system integration — each require distinct investments, and organizations that skip any one of them remain in pilot.
• Business-led ownership accelerates deployment where IT-led programs stall: Transferring operational control of vision AI programs to plant and business-unit leaders — with the right platform in place — consistently shortens deployment cycles and expands use-case coverage.
• Operational trust in visual AI is earned through small wins: Embedding subject matter experts in the design and feedback loop of a visual AI system, and targeting a narrow, high-visibility first use case, is what separates deployments that become standard practice from those that get quietly archived.

Ecosystem Readiness Determines Deployment Success

Episode: Turning Computer Vision Into Real‑World Value at Enterprise Scale – with Joseph Nelson of Roboflow

Guest: Joseph Nelson, Co-Founder and CEO at Roboflow

Expertise: Computer Vision, Physical AI, Enterprise Platform Strategy, AI Deployment​

Brief Recognition: Joseph Nelson is co-founder and CEO of Roboflow. Previously, he founded ROC AUC, a data science consulting firm, and co-founded Represently, which was acquired by Fireside21. He also taught machine learning and data science at General Assembly, where he developed curriculum and enterprise AI training programs.

Computer vision systems often demonstrate strong technical performance in controlled environments, but they fail to deliver value when organizations are not structured to receive, integrate, and act on what the model produces. Nelson frames the deployment challenge as three sequential requirements that must be in place before a system can operate reliably in production:

1. Data readiness.  

Does the organization have cameras or sensors positioned to capture what it actually wants to monitor? This is a physical question before it is a technical one. Do you have eyes on the cross-section of the battery? Do you have eyes on the installation at each step of the process? Do you have eyes on your stamping presses?

Without visual data of the right quality and position, there is nothing for a model to learn from.

2. Model specificity.  

Even as general‑purpose AI models improve, organizations deploying computer vision in manufacturing typically need to train models against their own products, their own defect types, and their own operating conditions. A vehicle manufacturer’s assembly line looks different from any other manufacturer’s, and the model has to reflect that.

3. Downstream integration.  

A visual AI system that correctly identifies four screws where eight are required produces no business value if that signal cannot reach the manufacturing execution system, the quality management platform, or the operator who needs to respond. Nelson describes this as connecting visual intelligence to the “downstream systems that allow you to run your business better” — whether that is a manufacturing execution system, a transportation operating platform, or an inventory and returns layer.

Nelson points to BNSF, a class one railroad operating across 30,000 miles of U.S. track that moves millions of containers annually, as an example of the value that becomes available when all three components are in place. Wheel inspection, track condition monitoring, container tracking, and preventive maintenance scheduling are all visually intensive problems that no human team can match for coverage or consistency. The value is only realized when visual intelligence connects downstream to the systems that schedule maintenance and dispatch crews.

According to Joseph:

“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.”

– Joseph Nelson, Co-Founder and CEO at Roboflow

Nelson recommends pairing executive‑level commitment to the long‑term potential with a concrete, bounded first use case at line level — what he calls a “barbell strategy.” That first use case becomes the proof point from which broader deployment can be justified and scaled.

​

Business-Led Ownership Accelerates Deployment Where IT-Led Programs Stall

Episode: How Vision AI Scales Across a Manufacturing Network – with Jeff Witt

​Guest: Jeff Witt, Digital Transformation Leader

​Expertise: Computer Vision, Manufacturing IT/OT Integration, Asset Health Management, Enterprise-Scale Deployment

​Brief Recognition: Jeff Witt leads computer vision and asset health programs for a large-scale manufacturing organization with more than 100 production facilities. He has built and scaled visual AI infrastructure from early pilot stages through full production deployment, navigating the architecture, change management, and organizational challenges that define enterprise-scale rollouts.

Jeff Witt’s experience leading computer vision and asset‑health projects has exposed a pattern: computer vision projects stall because organizations struggle to integrate and operationalize them. In his experience, the projects that progress are end‑to‑end solutions embedded in daily operations rather than being treated as standalone software projects.

A central barrier Witt identifies is architectural. Many facilities already have process cameras installed, but these systems typically sit on manufacturing IT networks that are isolated from enterprise data pipelines and business intelligence systems. According to Witt, combining manufacturing‑level vision data with other process data requires early integration work. Once his team solved this integration challenge, deployment became repeatable and scalable across sites, eliminating the need for plant‑specific engineering.

Jeff describes the practical starting point as building on existing infrastructure. Most facilities had some level of camera maturity, and his initial goal was to layer computer vision on top of what was already in place. As teams saw value, demand grew for additional cameras, higher‑resolution systems, and expanded coverage. This shift — from leveraging existing cameras to justifying new infrastructure — occurred only after early deployments demonstrated clear operational benefit.

He summarizes it as:

“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. I can pull video from an event, embed it in a project update, and everyone immediately understands what happened and why it matters. It’s one of the few technologies where the story tells itself.”  

— Jeff Witt, Digital Transformation Leader

On model readiness, Witt challenges the assumption that organizations must refine models for months before expecting value. In his experience, value arrives almost immediately. Forward engineers can produce initial model outputs within hours, and human‑in‑the‑loop oversight manages residual risk while the system improves. “None of our models are perfect,” Witt says, but they still deliver meaningful operational visibility and anomaly detection without requiring full optimization.

He emphasizes that successful deployment depends on organizational ownership. His team saw significant acceleration when responsibility moved from IT to plant and business‑unit teams. When operators and supervisors could define use cases, deploy models, and adapt systems directly — with vendor support but without IT gatekeeping — adoption expanded rapidly across similar lines in multiple facilities. As Witt explains, “Taking it out of IT hands and just letting the business run with the platform… has really allowed our plants to accelerate and expand use cases.”

The broader lesson, from Jeff, is that computer vision becomes a scalable operational capability when integration, ownership, and repeatability are addressed. As he summarizes, “It’s a platform, not a point solution… we can pretty much build any vision system or application we want.”

​

Operational Trust in Visual AI Is Earned Through Small Wins, Not System Rollouts

Episode: Making Visual AI Standard Practice in Complex Manufacturing – with Brian Ton of Florida Crystals Corporation

Guest: Brian Ton , Senior Laboratory Manager at Florida Crystals Corporation

Expertise: Quality Management, AI Integration, Manufacturing Operations, Change Management

Brief Recognition: Brian Ton is Senior Laboratory Manager at Florida Crystals, where he leads laboratory operations supporting manufacturing processes. Previously, he held a series of engineering and operations leadership roles at U.S. Sugar, including Chemical Engineer, Laboratory Supervisor, Manager of Laboratory & Water/Wastewater Operations, and Shift Manager. Ton holds a bachelor’s degree in Chemical Engineering from Florida State University.

The pattern Ton describes is familiar to most manufacturing organizations: a visual AI project clears the proof-of-concept phase, generates internal momentum, and then quietly stalls. The technology worked. The deployment did not.

In Ton’s experience, the cause is not technical failure. It is that the organization was not structured to absorb the system. The people most affected by the new tool — the operators, technicians, and quality staff working directly with the process — were not sufficiently involved in designing it. Their understanding of what the system needed to catch, and the practical realities of how it would be used during a shift, did not make it into the design. Critical features were missing. Edge cases that any experienced operator could have anticipated were never discussed in the planning phase.

Tan distills these early‑stage failures into two structural conditions that determine whether visual AI earns lasting operational trust:

• Proximity of subject‑matter expertise: The people who understand the process must be embedded in development and deployment, not consulted afterward. As Tan explains, “The closer the subject matter experts are to the solution… it builds the trust better.” This mirrors Jeff Witt’s observation that the people best positioned to specify what the system needs to do are almost never in IT.
• A functional feedback loop: Operators need a continuous mechanism to flag issues, suggest improvements, and see their input reflected in system behavior. Tan notes that earlier deployments failed when feedback surfaced too late or not at all, describing the loop as a calibration process rather than a user‑satisfaction channel.

Ton’s guidance on where to begin is clear:

“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.”  

— Brian Ton, Senior Laboratory Manager at Florida Crystals Corporation

The goal of the first deployment is not to solve the organization’s largest quality problem. It is to establish a track record that earns the credibility to tackle larger problems later.

Ton describes the productivity upside once trust is established as a multiplier of throughput — with data processing and measurement volumes increasing by orders of magnitude compared to what manual processes could achieve. The constraint is not the technology’s ceiling. It is the organization’s willingness to define a sufficiently specific starting point so that winning is clearly visible when it happens.