Interview with Cognex President & CEO Matt Moschner

When Matt Moschner, President and CEO of Cognex, talks about machine vision, he isn't describing a niche inspection tool. He's describing the nervous system of the modern factory.

And with the launch of OneVision, Cognex is making a deliberate bet that the bottleneck was never the camera or the algorithm, it was everything else.

"Where customers were feeling pain wasn't on-device performance," Moschner told AIM Media House. "It was everything around it."

Solving the Deployment Problem, Not the Detection Problem

Cognex has spent more than four decades building the tools manufacturers use to guide, inspect, and identify products at line speed. Its InSight devices are embedded in factories across the world, handling the kind of high-speed, high-stakes decisions that leave no room for error.

But as industrial AI matured, a new problem emerged. Not whether a model could do the job, but how long it took to get that model into production.

Deploying AI vision at scale, Moschner explained, requires collecting and labeling images, validating models across different lines and SKUs, managing version histories, and coordinating rollouts across multiple sites. Each step added time, cost, and friction. OneVision is Cognex's answer.

The platform keeps inference at the edge, while moving training, data curation, model management, and fleet-wide updates to the cloud. It also unifies Edge Learning and deep learning in a single workflow, so manufacturers can start with simpler applications and scale without switching platforms, swapping hardware, or redesigning jobs.

"It's a pressure-relief valve," Moschner said. "Teams get the simplicity of the edge, and the scalability of the cloud, all in one unified workflow."

The practical result was demonstrated at Paldo, one of Korea's largest noodle manufacturers and an early OneVision adopter. The company needed to tighten its seal-width inspection and reduce false rejects without adding infrastructure or moving to a PC-based solution.

Using OneVision, Paldo's team captured images directly from the production line, trained a more capable model in the cloud, and deployed it back to their existing edge devices, all without touching their production setup. Higher accuracy and fewer false rejects followed, at the same line speed.

Latency, Moschner noted, is effectively a non-issue once deployment is complete. "Runtime models no longer require a connection to the cloud," he said. Even if connectivity drops, inspection continues uninterrupted.

A Unified AI Platform

Cognex's move toward deep learning is often framed as a transition. Moschner pushes back on that characterization.

"Rule-based vision isn't disappearing," he said. "It's still the optimal tool for a huge percentage of factory-floor applications where repeatability, explainability, and microsecond-level speed matter most."

What deep learning enables, in his telling, is an expansion of what machine vision can solve, reading fonts a system has never seen, detecting subtle defects that don't follow a rule, inspecting surfaces with a high degree of natural variation. Problems that were previously out of reach.

Cognex invested early and intentionally to get there. Acquisitions including ViDi and SUALAB brought deep learning capabilities into the company, and those were integrated into a unified platform rather than bolted onto legacy products.

The result is a shared hardware and software environment where customers can blend rule-based and deep learning methods where each makes sense, and move between them without rework or workflow disruption.

"We didn't bolt a neural network onto a legacy product and call it AI," Moschner said. "We invested early and integrated those capabilities into a unified platform."

Culturally, he describes the shift as energizing rather than disruptive. Cognex engineers, he said, are motivated by difficult technical problems. Deep learning gives them new tools to push boundaries without requiring them to abandon the discipline that built the company.

"It's an expansion, broadening the range of problems Cognex can solve and giving customers more flexibility in how they solve them," he said.

From Specialists to Shop Floors

One of Cognex's more deliberate design choices has been building products that don't require machine vision engineers to operate them. The SLX portfolio, which combines advanced barcode reading with AI-powered item detection in a single unit, is designed for non-technical staff. According to the company, the SLX290 reduces hardware complexity and simplifies maintenance while improving overall equipment effectiveness.

Skeptics of this approach typically raise the same concern that simplicity comes at the cost of accuracy. Moschner rejects the premise.

"Removing the need for machine vision specialists does not mean sacrificing quality," he said. "We've simply packaged that power in an intuitive, web-based interface."

Guided setup, built-in guardrails, and validated templates ensure that users configure applications correctly and deploy reliably across lines and sites. Under the hood, the same advanced algorithms and AI-driven training workflows used by seasoned vision engineers are doing the work. The expertise has been embedded into the product itself.

This design philosophy is reshaping who buys Cognex. Historically, the company served roughly 30,000 customers, primarily large, technically sophisticated manufacturers.

Over the last several years, Cognex expanded its direct sales coverage specifically to reach regional and mid-market accounts. The addressable market, Moschner said, is five to ten times larger than the traditional base.

"We're seeing this especially in packaging and consumer goods, where many producers are regional, mid-market manufacturers," he said.

The results are measurable. In 2025, Cognex added 9,000 new customers. The company attributes the figure directly to the combination of simplified tooling and expanded go-to-market reach. These are manufacturers that get a clear on-ramp to advanced AI without disrupting operations or rebuilding workflows.

Trust, Autonomy, and What Comes Next

Moschner is careful about how he frames the future of agentic AI in industrial environments. The technology, he says, is moving in that direction, but the path has to be deliberate.

Cognex systems already feed closed-loop process control like detecting drift, triggering upstream parameter adjustments, and preventing defects before they occur. In his framing, that's a meaningful early step toward autonomous decision-making.

Full agentic autonomy, however, is a different matter. "Our customers don't just need AI that sees," Moschner said. "They need AI they can trust, audit, and defend to a quality engineer or regulator."

He argues that what manufacturers are ready for today is greater automation with guardrails. On-device inference for immediate response, cloud tools for governance and traceability, and auditable outputs that integrate cleanly into existing quality systems. The same Cognex hardware and software can grow with them as trust develops, without disruptive overhauls.

Looking three to five years out, Moschner sees the real frontier as something few are actively discussing. Vision becoming a continuous intelligence layer across entire production flows. Once every cell, line, and asset shares visual context, vision stops being a point tool and becomes the real-time fabric coordinating quality and maintenance across the plant.

Predictive quality, detecting subtle process drift before it produces scrap, is one major horizon. Autonomous setup, where systems self-calibrate and select models without deep technical expertise, is another. And transferability, where models maintain accuracy when deployed across different lines, machines, or facilities, could turn isolated wins into global standards.

"The next wave isn't just better defect detection," Moschner said. "Its vision is to become the continuous intelligence layer of the factory."

But for all the talk of platforms, agentic loops, and AI-native workflows, Moschner returns to a more grounded observation when asked what actually holds adoption back. It isn't the algorithms or the infrastructure. It's the people.

"The biggest barrier is organizational confidence," he said. "The technology is ready, but adoption slows when teams are asked to trust a system they don't fully understand with decisions that carry real cost, quality, and compliance consequences."

Confidence is built by keeping operators in the loop during setup, building on validated pilots, and letting familiarity grow alongside results. The manufacturers that succeed aren't the ones who deploy fastest. They're the ones who earn trust incrementally, and expand from there.

"Trust and autonomy grow together on one continuous platform, not through disruptive overhauls," said Moschner.