Zero Hallucinations on a Live Robotic Process: What Bryan Zorn Learned About AI in Advanced Manufacturing

Today most manufacturing operations and the AI story are nearly identical. A few ChatGPT subscriptions, maybe a dashboard pilot, and a lingering uncertainty about what comes next. The promise of advanced manufacturing technology gets discussed constantly in boardrooms and trade publications. What actually reaches the plant floor, in most cases, is far thinner than the conversation around it suggests.

Bryan Zorn has spent more than two decades inside that gap. He has grown businesses, worked with emerging technology, and built manufacturing programs from the operational side rather than the vendor side. Today he leads business development at the Commonwealth Center for Advanced Manufacturing, known as CCAM, a non-profit research and technology center in Virginia founded in 2013 by Rolls-Royce with support from the Commonwealth of Virginia. CCAM runs industrial-scale test beds using real machine data, not lab simulations, and partners with Virginia Tech, UVA, VSU, VCU, and federal agencies on projects spanning nearly every corner of manufacturing.

When Zorn talks about AI adoption in manufacturing, he is not selling a platform. He is describing what happens when advanced manufacturing technology gets tested properly before anyone bets a production line on it.

Inside a Manufacturing Center of Excellence

CCAM is not a consultancy, and it is not a vendor. It operates as a manufacturing center of excellence: a non-profit where industry members drive research projects using real industrial equipment rather than theoretical models.

That sentence captures what separates CCAM's work from most corporate AI pilots. The center exists to move a technology along both the Technology Readiness Level and the Adoption Readiness Level, meaning it validates not just whether something works technically, but whether an organization is actually prepared to run it day to day.

The center was founded in 2013 after Rolls-Royce established a manufacturing research facility in Disputanta, Virginia, just south of Richmond, with backing from the state. CCAM has since built out federal contract work alongside university partnerships across Virginia Tech, UVA, VSU, and VCU, giving it access to research talent in addition to its industrial test beds.

That combination matters for a simple reason. When Zorn describes a system performing well, he is describing something tested on live, industrial-scale machinery generating real operational data, not a simulated environment built to make a demo look clean. For manufacturers evaluating advanced technology in manufacturing on their own, that difference between a lab demo and a live production test is often the entire question that decides whether a project succeeds. It also sets up one of the more striking claims later in this conversation: a live deployment that returned zero failures and zero hallucinations.

Why AI Adoption in Manufacturing Is Still Stuck

Ask Zorn why AI adoption in manufacturing has not moved faster, given that generative AI has been commercially available for several years now, and he does not point to a single cause. He describes four barriers that tend to compound.

The first is capital, though not only in the financial sense. Many smaller manufacturers do not have the dollar capital or the human capital required to run a structured AI implementation. Sometimes, Zorn says, that shortage is genuinely the whole story. There is no spare time, no spare headcount, and no budget to sit down and map out where a new system would even apply.

The second is friction between operational technology and information technology teams, a divide that predates AI by decades but complicates nearly every new deployment. "One side wants security, one side wants to run fast," Zorn explains, "and those are two things that haven't really gone together." Both teams have legitimate priorities. Neither one is wrong. The problem is that those priorities rarely get reconciled before a project stalls out. That friction looks a little different at enterprise scale, where how Nicole Hilgenkamp navigated AI adoption at enterprise scale covers a separate conversation on the same underlying tension, but the core dynamic Zorn describes holds up across company size.

The third barrier is workforce. Senior operators are retiring with process knowledge that was never fully documented, and fewer people are stepping in to replace them. AI adoption needs an internal owner, someone who understands the plant well enough to guide the deployment. When that person does not exist inside an organization, adoption slows regardless of how much budget is available.

The fourth is perception. Much of the skepticism Zorn hears is not really about AI's capability. It is about trust that has not yet been earned. "I don't want to say it's unproven," he says. "It's been around for a long time, and now we just have a lot of compute power to really make it do really cool stuff." The technology has moved faster than the confidence surrounding it.

Zorn points to a widely reported episode at Microsoft, where a team was let go following AI-driven changes and then largely rehired after the technology ended up costing more than it saved, as an example that circulates often in manufacturing circles. His read on it is not that the failure proves AI does not work. It proves that deployment decisions matter more than the technology itself. "If you set it up right," he says, "it'll really have a big impact."

Industry 5.0: Putting Humans Back Into the Loop

Much of CCAM's AI work sits inside what Zorn calls industry 5.0 manufacturing, the shift beyond pure automation toward systems that put people and robots into a genuinely collaborative loop rather than replacing one with the other. The goal is not to remove the operator. It is to give that operator better information, faster.

Two examples come up repeatedly in Zorn's work. The first is ergonomics monitoring: AI systems that track repeated motions on the floor and flag when a worker may need support, a different tool, or a rotation, before strain turns into an injury. The second is additive manufacturing, where AI in additive manufacturing calculations can meaningfully speed up print quality and throughput. Zorn describes a case where a print speed increase of roughly 10 percent turned into a real competitive advantage, compressing a part's development timeline from years to months, a difference that matters enormously in defense and industrial supply chains.

This is where human in the loop manufacturing becomes more than a phrase. The AI is not asked to make the final call. It is asked to surface where attention is needed. Zorn describes it as finding the specific area where a large share of problems concentrate, then putting a trained human on that exact spot. The system narrows the search. The person still makes the decision.

What 100% Accuracy on a Live Robotic Process Actually Means

The most striking result Zorn shared involves a live deployment at CCAM: a real robotic process running multiple machines with people working alongside them. According to Zorn, the test returned zero failures and zero hallucinations, with 100 percent accuracy in that specific application.

That result deserves context rather than repetition. Zorn is careful to frame it as a narrow, well-scoped outcome, not a blanket claim about AI running unsupervised across an entire planet. What made it possible was not a generic subscription tool. It was a system built on the operation's own data, checked across multiple models, and layered with agents built for that exact workflow rather than a general-purpose assistant. "You can get one hundred percent or darn near one hundred percent accuracy," he says, "if you're putting in your own data, running off of multiple models inside, and then building your own agents."

Industry-wide, most AI pilots in manufacturing still fail to reach production. Zorn's point is not that his result contradicts those numbers. It is that both outcomes describe the same technology deployed in opposite ways. One is a subscription bolted onto an existing workflow. The other is a purpose-built system tested against real operational conditions before anyone trusted it with production. For manufacturers weighing whether that kind of investment pays off, AI cost reduction case studies from other sectors show a similar pattern of narrow, well-tested deployments outperforming broad, generic ones.

Where Manufacturers Should Actually Start

Not every manufacturer has access to CCAM-scale test beds, and Zorn's advice for those companies follows a clear progression rather than a single leap.

The first step is using what already exists. Camera systems with built-in AI, or dashboard tools that connect to equipment a plant already owns, often solve a specific, expensive problem without requiring a full system build. "There are tools out there that they can use to bring that in now without a lot of headache," Zorn says.

The second step is a dashboard built on a company's own data, not generic analytics. Zorn describes the kind of question a plant operator should be able to ask once that dashboard exists: give me an update on cell five, or a list of specific reports. Fast, accurate, and tied directly to how that plant actually runs. This same progression shows up in other sectors too, including AI in food and beverage manufacturing operations, where dashboard-first deployments have followed a nearly identical path before moving to anything more advanced.

The third step, once the data is clean and the dashboard works, is purpose-built agents for narrow workflows such as inventory tracking, maintenance alerts, or procurement. This is where Zorn's earlier point about combining a company's own data with multiple models and custom agents becomes a practical roadmap rather than an abstract claim.

His underlying principle ties the whole sequence together. "Hey look, you're already using it," he says. "It's already in the stuff you're doing." The goal is not to bolt AI onto an operation and hope. It is to show people what they are already relying on, build comfort with how it behaves when configured correctly, and then let it do the sorting work humans do not want to do by hand. "Tell it what to do," he says, "instead of sorting through 150 files, it gets you down to where you need to be really quickly."

Bryan's Read on Where Manufacturing AI Stands in 2026

Zorn's closing read on the industry is measured. AI adoption in manufacturing is real and accelerating, but the distance between organizations moving quickly and those still running on a handful of subscriptions is widening rather than closing. The technology itself is proven; CCAM's zero-hallucination result is not a unique fluke. What remains unevenly distributed is the knowledge of how to deploy it correctly.

Zorn credits consortia, public institutions, and industry leadership for getting that message out at the top of organizations. The harder work, he says, is happening lower down, where adoption actually gets implemented. But the proof points exist now. Manufacturers still waiting for more evidence before they start are, in Zorn's view, no longer really waiting for the technology. They are waiting for permission.

Frequently Asked Questions

Bio Block

About the Interviewer