Humans-in-the-loop is likely the fastest way to AI value

Building the machines that power AI comes with an unusual manufacturing problem: some of the parts are worth as much as a house.

Sviat Dulianinov, CEO of data center infrastructure maker Bright Machines, remembers the first time his company started building GPU servers and realized that a single component could be worth about $250,000.

“If you damage or break one component…you run into a big problem,” said Dulianinov, CEO of Bright Machines.

That kind of risk helps explain the company’s approach to manufacturing AI infrastructure. Bright Machines is highly automated, but Dulianinov isn’t interested in automation simply for automation’s sake. And despite the company’s heavy use of machine learning and artificial intelligence, he doesn’t envision factories where machines make every decision themselves.

People remain in the loop

Bright Machines sits in an unusual part of the AI supply chain. It doesn’t make the chips or design the servers. Instead, it assembles components—including motherboards, memory and cooling hardware—into the GPU compute systems, CPU platforms and storage racks that ultimately end up in data centers.

The challenge is doing that quickly and efficiently, as demand for AI infrastructure grows while manufacturers struggle to find enough workers.

Dulianinov estimates the U.S. manufacturing workforce is short roughly three million people. That makes automation increasingly necessary for companies bringing production back to the United States.

But his starting point is different from the traditional approach.

“We start with robotics and software before we engage people,” he said. “There are still processes not designed ideally for robots so there you have to use humans as well.”

Start with the product, not the robot

One of the biggest barriers to automation, Dulianinov said, often hits before a product ever reaches the factory floor—it’s the product itself.

Historically, product designers and manufacturing engineers haven’t always worked together closely enough to account for robotic assembly during the design process, resulting in a product that looks perfectly reasonable on paper but turns out to be difficult—or prohibitively expensive—to automate.

“You look at all the processes and you realize some of them just you can’t automate,” Dulianinov said. “You don’t want to do automation for the sake of automation because sometimes it’s really expensive.”

Bright Machines tries to identify those problems earlier with an application called Bright Designer.

Before a new server or storage product goes into production, engineers can use the application to analyze the design and determine whether changes could make it easier for robots to assemble.

“[We] do different analysis and tests in this app and actually provide feedback how to improve designs to improve the level of automation and to make it easier for robots to assemble,” Dulianinov said.

The idea is essentially design for automated assembly. Automation becomes considerably harder when engineers are trying to make a robot work around a product that was designed entirely with human assembly in mind.

A factory made of cells

That same philosophy carries onto the factory floor. Rather than building one long, fixed production line, Bright Machines uses modular manufacturing cells. Each cell is roughly the size of a household refrigerator and typically contains one or two robotic arms performing a particular operation. A server might require anywhere from 10 to 30 separate assembly steps, depending on the complexity.

“If you have a particular server…and let’s say it has 10 steps of the process…each station would be a process,” Dulianinov said.

Put those stations together and a product moves through the sequence until it emerges at the end of the line. The advantage of this modular approach becomes apparent when the product changes. Instead of rebuilding an entire production line, Bright Machines modifies individual cells. End-of-arm tooling is changed to accommodate new components, while people can temporarily step into a station while equipment is being reconfigured.

That flexibility is particularly important in AI infrastructure, where server configurations and components can change quickly. But the robots aren’t the only thing moving through the cells.

Data is moving through them, too.

Every station tells a story

Dulianinov sees each manufacturing station as both a place where work gets done and a source of information.

“At each station you’re going to collect data whether it’s an automated station or not,” he said. “Even if you have people in the loop you’re going to collect data…so that you have traceability.”

The data can include inspection results, images, force-control measurements and information about individual components. It gives engineers a record of what happened to a product as it moved through the factory.

If something fails, they can go and look for patterns. Was the problem associated with a particular supplier? Did something unusual happen at a particular station? Was there a defect in a component? The goal is eventually to incorporate those lessons back into the next product.

“We do our best to collect data from the very start to the very end because that data can be used to drive more efficient operations later,” Dulianinov said. “Then finally close the loop because all that data you collect from robots and from the products that you build, you should be able to use for the next generation.”

For expensive AI hardware, that can be the difference between catching a problem early and destroying a component worth hundreds of thousands of dollars.

To address this, Bright Machines uses 3D navigation, inspection systems and force-control monitoring as part of what Dulianinov describes as a “zero scrap” philosophy. If a process begins deviating from what is expected, the system can stop production before the problem becomes an expensive failure.

AI doesn’t get the final word

For a company building the infrastructure behind the AI boom, Bright Machines’ own use of AI is surprisingly pragmatic.

“AI is a fancy word today because everything became AI,” Dulianinov said. “We’ve been using machine learning from the start.”

Machine learning has long been part of Bright Machines’ robotics systems, supporting inspection, navigation and robot decision-making through what the company calls its Smart Skills.

More recently, the company has started using generative AI and large language models to help engineers work faster. One application is creating the instructions—or “recipes”—that tell robots how to perform assembly tasks.

“We started exploring how to use AI for better training the robot and better creation of recipes,” Dulianinov said. Instead of having engineers manually build every workflow, AI models can help create those flows faster.

The company is also using AI-assisted coding tools and experimenting with agentic AI to diagnose manufacturing problems. That’s where the mountains of manufacturing data become particularly useful. Instead of an engineer manually working through production logs after a failure, an AI agent can analyze the information and suggest likely causes.

“If you have an agent, it can present you with three assumptions from the very start because it went already through the data,” Dulianinov said. “Then the engineer can only test A, B and C instead of A to Z.”

But there’s a critical qualifier: the engineer still makes the decision.

“Even if AI does 70%, you still have an engineer going to finish, and that’s going to be the final person who’s going to say, ‘Okay, it’s good to go,’” he said. “The agent would not act on those assumptions until a system engineer is going to have a look and say…that’s the right thing. You have human in the loop,” he said. “I think that’s the easiest answer.”

The factory of the future still has people

That philosophy runs counter to the idea that increasingly capable AI and robotics will eventually eliminate manufacturing jobs.

“People ask, ‘What is the factory of the future?’” he said. “The answer is the factory of the future still has people in the loop. So you still create jobs.”

The people may do different things. Robots may perform more of the repetitive physical work. AI may sift through data that would take engineers hours to analyze. But people remain responsible for understanding what the systems are doing and deciding what happens next.

Bright Machines has deployed more than 130 microfactories, he said, giving the company years of experience watching automated systems operate in real manufacturing environments.

“I think one, it’s about experience, you just see and learn how people use that, how it works on the floor, where you learn about mistakes, where machine learning worked, where it didn’t work.”

That experience has also made him skeptical of AI hype.

“You shouldn’t be delusional that if you trust AI to do everything, it’s going to,” he said.

Even sophisticated systems make mistakes. The answer isn’t necessarily to abandon them, but to understand where they can fail and build processes that account for those failures. That is also why Dulianinov isn’t particularly impressed by technology demonstrations that don’t translate into measurable manufacturing improvements. Manufacturers care about uptime, downtime and quality, he said. Those are the metrics that should determine whether technology belongs on the factory floor.

Making engineers better

AI is already making a difference inside Bright Machines, but Dulianinov doesn’t describe the benefit as replacing engineers. He describes it as amplifying them.

“It definitely makes engineers more efficient,” he said. “Especially good engineers become much better.”

But he sees a similar lesson when he looks beyond Bright Machines. The biggest companies may have more money to experiment with AI, but that doesn’t necessarily make implementation easier.

“I don’t agree that large companies have the resources to do that,” he said. “Large companies have the resources to have trials and make an announcement and buy all the licenses. But the implementation and engagement…it’s really hard to drive real engagement.”

Indeed, having access to AI isn’t the same as knowing how to use it effectively.

“When you have completely new tech, yes, everybody has a chance to make similar mistakes,” Dulianinov said. For manufacturers, the lesson from Bright Machines’ experience is relatively simple. Don’t automate just because you can and don’t trust AI simply because it sounds intelligent. Keep people in the process, even when you’re sure the technology

The post Humans-in-the-loop is likely the fastest way to AI value appeared first on Engineering.com.