Reducing Waste and Improving Efficiency with Multi AI Agent Quality Control in Manufacturing

Most AI projects in manufacturing never make it past the pilot stage. A recent MIT study found that more than 90% of generative AI pilots in industrial settings fail. Only about 10% succeed, and those tend to be led by teams with deep expertise in both the technology and the specific use case.

So what separates the projects that work from the ones that quietly collect dust?

In a recent conversation with Willem Klein, CEO and co-founder of Zetamotion, we explored why AI-powered quality control is so hard to get right, what manufacturers keep underestimating, and how multi-agent inspection tools could finally close the gap between proof of concept and production.

The Real Reason AI Pilots Stall on the Factory Floor

The failure rate is alarming, but the reasons behind it are surprisingly human. One of the most overlooked factors is what the MIT report calls the "dark number" — unsanctioned AI projects built by skilled engineers working outside official channels.

These aren't rogue employees. They're often a company's best people. They see inefficiencies, they know a tool like ChatGPT exists, and they start hacking together their own solutions. The problem is that these grassroots efforts rarely get the support or infrastructure they need to scale.

As Klein explains, this signals something important: the people closest to the work are often the ones best positioned to translate operational knowledge into automated processes. But most AI tooling today doesn't meet them where they are. It demands data science skills they don't have, or forces them into rigid workflows that don't match real-world conditions.

The result is a painful gap. Either you hand engineers a do-it-yourself toolkit and push all the hard work onto them — collecting thousands of images, labeling data, training models — or you hire an expensive solution provider who takes months and charges accordingly.

Why the Best AI Model Still Isn't Enough

There's a common assumption in manufacturing leadership that the path to successful AI is simply finding or building a better model. Klein pushes back hard on this idea.

He points to real-world cases where companies built extremely accurate inspection models — outperforming any human — but the projects still failed. The reason had nothing to do with model performance. It was everything around it: clunky interfaces, awkward loading steps, and reporting workflows so cumbersome that operators found it faster to just do the inspection themselves.

This is the "Excel script from 2006" problem. Every manufacturer has a stack of legacy systems and internal tools that have evolved over decades. If you drop a high-performance AI model into that environment without thinking about how it fits into the operator's actual workflow, you've built something impressive that nobody uses.

Klein argues that the competitive advantage doesn't come from the model itself — it comes from the end-to-end system. That means modular deployment, clean interfaces that show operators exactly what they need (and nothing more), and tight integration with existing infrastructure. The AI has to make someone's job easier on day one, not just perform well on a benchmark.

What Multi-Agent Quality Inspection Actually Looks Like

This system-level thinking is what led Zetamotion to build Zelia, an AI assistant designed to handle the full lifecycle of inspection — from training to deployment — without requiring a data science team.

Here's how it works. Instead of asking a manufacturer to collect 100,000 images and hand-label each one (Klein cites a glass manufacturer who did exactly this), Zelia starts with roughly five good product samples and five examples of each defect type. From there, it synthesizes its understanding of what the product should look like and what defects look like, then verifies that understanding with the human operator. The alignment process typically takes less than an hour.

Once Zelia has confirmed its understanding, it generates synthetic training data, trains accurate models, and plugs them into Zetamotion's modular inspection platform, Spectron. The time savings are dramatic. What used to take weeks or months of manual data collection and labeling can be compressed into a single working session.

The broader roadmap is even more ambitious. By end of year, Klein says, Zelia will handle not just model training but the full solution build — custom dashboards, API connections, architecture configuration — all driven by conversation with the user. Describe what you need to inspect, provide a few examples, and the system builds the entire inspection pipeline.

Why You Can't Copy-Paste Between Production Lines

One of the most underestimated challenges in scaling AI quality control is the assumption that what works on one line can simply be replicated on another. Klein is blunt about this: no two production lines are the same, even when they're running the same product.

The differences are often invisible to upper management but obvious to the people operating the equipment. One machine might be from 1988, upgraded in 2006. The next line's machine might be ten years newer with no upgrades. One line runs hotter at a certain stage. Environmental factors affect how products present visually at different steps.

This matters enormously for multi-step quality control — checking products at multiple points along a production line. Because conditions vary at each step and across each line, you cannot take a model trained for one context and expect it to work in another without intelligent fine-tuning. Scaling AI in manufacturing is not a copy-paste exercise.

Edge Deployment and the Data Sensitivity Problem

Where the AI runs is another critical decision. Klein reports that most manufacturers in quality control strongly prefer edge deployment over cloud solutions, and the reason is simple: QC data is some of the most sensitive information a manufacturer has. Full product scans including defect parameters, measurements, and QC results reveal things companies don't want outside their walls. As Klein puts it, it's like looking into someone's drawers — you see everything, including the skeletons.

Any AI quality control solution that requires pushing detailed product data to the cloud is going to face serious resistance from manufacturers who see this as both a security and competitive risk.

How to Give Your Best People a Better Tool

The most compelling insight from this conversation isn't technical — it's philosophical. Klein, whose academic background is in technology ethics and philosophy, frames the human-AI relationship in manufacturing not as replacement but as amplification.

His image is worth remembering: don't think of AI as a robot walking onto the factory floor telling everyone to go home. Think of it as handing your best people a supercharged magnifying glass that draws their attention exactly where they need to apply their expertise.

In inspection work, edge cases and outliers show up constantly. Experienced operators make excellent judgment calls about these situations, and that feedback makes the AI system better over time. The most successful deployments are human-in-the-loop systems where the technology handles the repetitive, high-volume work while humans handle the exceptions that require real judgment.

For manufacturing leaders looking to move AI quality control from pilot to production, the takeaway is clear: stop chasing better models and start building better systems. Invest in tools that meet your operators where they are, respect the complexity of your production environment, and keep humans in the loop where their expertise matters most

‍

‍