Artificial intelligence has become one of the most widely discussed technologies in manufacturing, but much of the conversation remains focused on software – dashboards, analytics, predictive models, and digital decision-making tools.
Increasingly, however, manufacturers are asking a more practical question: how can AI move beyond identifying problems and begin helping to solve them on the factory floor?
That challenge is particularly relevant in the automotive industry, where production lines operate at high speed and even small quality issues can have significant downstream consequences.
While AI-powered visual inspection systems have become more common, many still stop at defect detection, leaving human operators responsible for deciding what action to take and when to take it.
Brandon Speweik, head of manufacturing at GFT Technologies, believes the next stage of industrial AI will be defined by its ability to interact directly with physical production processes.
GFT recently demonstrated a manufacturing system that combines machine vision, robotics, cloud infrastructure, and AI-driven root-cause analysis to not only identify defective components but also reposition, remove, or escalate them automatically on live assembly lines.
In this interview, Speweik discusses why manufacturers are becoming impatient with AI systems that merely generate insights, and why the industry is increasingly focused on connecting AI to real-world action.
He explains the challenges of integrating machine vision, robotics, cloud systems, and operational data in demanding automotive environments, and outlines why trust remains one of the most important barriers to wider adoption of autonomous decision-making systems.
The conversation also explores the realities of deploying AI in factories filled with legacy equipment, the growing role of AI-enabled quality control and predictive maintenance, and why fully autonomous factories may still be further away than some industry forecasts suggest.
Rather than replacing human judgment entirely, Speweik argues that the most successful manufacturing AI systems will be those that earn trust gradually by delivering measurable results in specific, high-value applications.
Interview with Brandon Speweik
Robotics & Automation News: Many manufacturers already use AI-powered visual inspection systems, but most still rely on humans to respond when defects are detected. Why has the industry struggled to close the gap between detection and physical intervention?
Brandon Speweik: Detection has matured faster than intervention. Cameras and AI models can now identify anomalies consistently, but the harder part is what happens after a defect is detected.
On a modern assembly line, the gap between a defect being identified and that part becoming embedded in a larger subassembly is measured in seconds. Once a software system flags an issue, someone still has to walk over, look at it, decide what to do and then physically act.
Even when that handoff works smoothly, manufacturers lose time and introduce the risk of human error. For example, if a staff member incorrectly clears a flagged defect, that part moves down the line, compounding the issue.
That is why this cannot be solved by a better AI model alone. Closing the gap requires computer vision, robotics, operational data, workflow design, and human escalation paths to work together.
The real opportunity is moving from AI as a detection layer to AI as part of a broader execution orchestration system, where detection, intervention, evidence capture, and learning are connected in one operating flow.
R&AN: GFT’s new system combines machine vision, robotic manipulation, cloud infrastructure, and AI-driven root-cause analysis. From an engineering perspective, which part of that integration was the most difficult to solve reliably on a live automotive assembly line?
BS: The hardest part is making the full system operate reliably in the full context of a live production environment. Machine vision, robotic manipulation, cloud infrastructure, and AI-driven root-cause analysis each bring their own complexity.
But the bigger challenge is synchronization, which includes detecting an issue, triggering the correct physical action, preserving the right evidence, and feeding that event back into the broader operational data environment without slowing or destabilizing the line.
In a live factory, the system has to deal with lighting, part position, cycle time, mechanical tolerances, network latency, and downstream dependencies.
Edge systems are critical for fast detection and action, while the cloud layer supports root-cause analysis, model improvement, image storage, and cross-line learning. The goal is not just to correct one defect, but to understand why it happened and prevent it from recurring.
R&AN: Automotive factories operate at extremely high speeds with very little tolerance for disruption. How do you ensure that AI-driven robotic intervention systems can operate consistently without slowing production?
BS: Fast detection and immediate intervention need to happen at the edge, close to the line, where latency is minimal. The cloud layer is better suited for image storage, root-cause analysis, model improvement, reporting, and broader operational learning. It also helps to specialize the workflow.
One system may inspect, another may classify or mark, and another may physically intervene. By separating those responsibilities, each step becomes more predictable and less likely to create bottlenecks.
The system also isn’t trying to make every call autonomously. When the AI is confident, it acts, but when it isn’t, the part is pulled for human review rather than stopping the line.
This allows the assembly line to continue to move at speed while routing the harder judgment calls to the people best positioned to make them.
R&AN: Your system not only identifies defects but also automatically repositions or removes parts. Do you see this as the beginning of a broader shift toward more autonomous quality-control systems in manufacturing?
BS: Yes, and we see things moving towards more closed-loop quality systems versus fully autonomous factories. For a long time, quality control was largely downstream. A defect was identified, documented, and analyzed later.
What is changing now is that quality signals can be captured earlier, acted on faster, and connected to root-cause analysis in near real time. When a robotic system removes or repositions a defective part, that is only one part of the value.
The larger value is that every inspection, intervention, escalation, and outcome can become part of a learning system. The organization can preserve what happened, why it happened, what action was taken, and whether it solved the issue.
This creates a stronger foundation for continuous improvement. Instead of only catching mistakes, manufacturers can begin to prevent repeat issues by linking quality events to upstream production conditions, supplier inputs, tooling, maintenance history, and operator workflows.
R&AN: AI in manufacturing is often discussed in terms of analytics dashboards and software optimization. How important is it for AI to move into the “physical world” through robotics and direct machine interaction?
BS: It is critical, because manufacturing ultimately happens in the physical world. For the last few years, the AI conversation in manufacturing has centered on dashboards and software, as these systems deliver value through new levels of visibility and pattern recognition that weren’t previously feasible.
But the manufacturers I talk to are getting impatient with those use cases because they’ve already seen multiple different iterations of the same dashboard. What they want to know now is when their AI investments will lead to new efficiencies and productivity.
If AI can detect a defect but cannot help prevent, route, correct, or escalate it, much of the value remains unrealized. That does not mean every AI system needs to directly control a machine.
In many cases, the most valuable role for AI is to guide human work, recommend interventions, capture evidence, or coordinate workflows. But the direction is clear: AI has to become more embedded in how work is actually performed.
R&AN: One challenge with AI systems is trust. How do manufacturers react when AI is allowed not just to recommend actions, but to physically alter production outcomes in real time?
BS: Manufacturers are pragmatic about this. Trust is not a yes-or-no question. It depends on what the system is allowed to do, how confident it is, what evidence it provides, and how quickly a human can intervene when needed.
In our experience, manufacturers are generally comfortable with AI making autonomous calls on things that are clearly within the system’s wheelhouse, such as a part that’s obviously misaligned or a label that’s clearly unreadable.
Where they push back, is on the ambiguous cases. For example, it comes to 50/50 calls, they don’t want the machine to make those on its own. They want those decisions routed to a person. Trust also depends on auditability.
Operators and leaders need to see what the system detected, what evidence it used, what action it took, and what happened afterward. That evidence trail is what allows trust to grow over time.
R&AN: GFT mentions using AI agents for automated root-cause analysis. How close are manufacturers to achieving truly self-optimizing production systems that can identify and correct process problems autonomously?
BS: Real progress is being made, particularly in building the digital infrastructure necessary for fully self-optimizing production systems, but realizing those systems is still a longer-term goal.
Today, when the system catches a defect, it points back to where the problem likely originated and routes the flag to the team that owns it in real time.
This is still a significant improvement over historical root cause analysis processes, which usually occur after the fact and require manual intervention. The harder piece is the closed loop.
Moving from a system that identifies a problem to one that automatically corrects the upstream process is a bigger leap than it sounds.
It requires the AI to have authority over operational systems it currently only observes, and it requires those systems to be integrated, goverened, and trustworthy enough to act on AI input without a human review step.
Most plants aren’t there yet, and likely won’t be for a while. For example, if a defect is caused by a bad paint batch, we are still some distance from AI determining that the paint batch is the problem, notifying the supplier, and automatically identifying which other paint inventory or workarounds the factory floor needs to keep running.
The manufacturers we are working with today aren’t pushing for fully autonomous factory floors. Most of them want AI to operate autonomously in areas where it has earned their trust, and to keep human judgment in place for the decisions that matter most.
R&AN: Many factories still operate with legacy equipment and fragmented data environments. How difficult is it to integrate modern AI robotics systems into existing automotive manufacturing infrastructure?
BS: It is difficult, and the challenge almost always starts with data rather than AI. Automotive plants, especially established ones, are running on a patchwork of systems that were never designed to talk to each other.
Production scheduling, quality tracking, supplier management and logistics systems were likely implemented by different teams at different times, often with different vendors.
The data is there, but it’s fragmented and inconsistent. When you try to drop an AI system on top of that, the AI is only as good as the data it can access, and the integration work to make that data usable is usually larger than the AI work itself.
That is why successful AI deployments usually require an operational data foundation. The goal is to contextualize production, quality, maintenance, supplier, and workforce signals into a common model that AI systems can reason over.
Without that, AI is working from isolated data points rather than operational reality. Legacy equipment adds another layer. You will find machines on the floor that are 20 or 30 years old and still perform well, but they were not designed with modern sensors, APIs, or cloud connectivity in mind.
The successful approach is usually incremental: start with a high-value workflow, connect the minimum data needed, capture evidence at the point of execution, and build outward from there.
R&AN: Looking ahead, where do you think AI-enabled robotics will have the greatest impact in automotive manufacturing over the next five years – inspection, assembly, logistics, predictive maintenance, or fully autonomous production workflows?
BS: Quality and inspection will likely see the greatest impact first, followed closely by predictive maintenance. Assembly and logistics will continue to advance, but fully autonomous production workflows are further out than what some projections suggest. The reason quality is at the top is that the economics are clearest.
Defects are expensive, and the earlier they are caught, corrected, and understood, the more value manufacturers can capture. When AI-enabled robotics can not only detect a defect but also act on it and preserve the evidence, the ROI becomes easier to justify.
Predictive maintenance is close behind because the data infrastructure is already largely there. Plants have been instrumenting equipment for years, and the shift now is from systems that can predict when a machine will fail to systems that can route work around the machine or trigger a maintenance order before it does.
Assembly and logistics will follow, but those are heavier lifts. Assembly involves more variation and more physical complexity, and logistics depend as much on the supplier ecosystem as on the plant.
Both will move, but the gains could compound more slowly. Fully autonomous production workflows are the longest arc. The vision is compelling, but the operational, regulatory and workforce realities introduce complexities not easily addressed.
The plants that get furthest will likely be the ones that earn trust in narrower domains first, including inspection, intervention, guided execution, maintenance, and evidence-based escalation, and let broader autonomy emerge from that foundation rather than trying to leap to it.
