In theory, AI should have transformed manufacturing by now. From predictive maintenance and fatigue detection to real-time quality control, the promise has always been smarter, faster, and safer operations. But in practice, the factory floor is still a place where AI ambitions often run into real-world limitations.
That’s a huge problem, especially because the size and weight of this industry are hard to ignore. U.S. manufacturing alone contributes $2.9 trillion to the economy, accounting for over 10% of total output and supporting nearly 13 million workers, according to the National Association of Manufacturers. Globally, manufacturing represents 16% of the world’s gross domestic product (GDP) and a total market value well over $16 trillion, per a new report from Cargoson, a transport management software company.
Now, as AI advances even further and policymakers push for reindustrialization in the U.S.—aiming to restore domestic production capacity, regain supply chain control, and modernize strategic infrastructure—the spotlight is back on factories. There’s momentum and money behind the movement, but without restructuring the fragmented digital systems that dominate most production floors, that momentum may stall. An estimate by market research firm MarketsandMarkets projects the global AI in the manufacturing market would grow to $155 billion by 2030, up from $34 billion in 2025—but that growth will remain theoretical unless companies solve the bottlenecks slowing down adoption.
Outdated infrastructure
According to a 2025 survey of more than 500 manufacturing leaders, 92% say outdated infrastructure is holding back generative AI progress. Another report on the state of AI infrastructure by cybersecurity firm A10 Networks found that 74% of global IT decision-makers believe their current infrastructure is not fully prepared to support AI workloads. For all the talk of digital transformation, many factories are still running on architecture that predates smartphones, most of which cannot support new AI capabilities.
“The hype around AI in manufacturing is real, but so are the technical barriers,” Shahid Ahmed, EVP of new ventures and innovation at NTT Data, tells Fast Company. “Modern connectivity is unlocking the next wave of AI-driven innovation in manufacturing. Private 5G and next-gen Wi-Fi give manufacturers the speed and reliability to finally turn AI into a productivity engine.”
However, better connectivity is just one part of the big problem with getting AI to produce optimal results on the factory floor. What’s really stopping AI from working on the ground isn’t just weak networks, but also a mismatch between how factories run and how AI systems think.
At aiOla, a conversational AI company that works with Fortune 500 manufacturers, Assaf Asbag sees a common pattern: data silos, fragmented systems, and little end-to-end accountability. Even when manufacturers bring in advanced models and top-tier talent, the results rarely scale.
“Even with expensive AI talent, teams can’t generate value if they don’t have clean, connected data,” explains Asbag, aiOla’s chief technology and product officer. “You need aligned data, integrated workflows, and clear accountability—otherwise pilots never scale.”
That’s because many manufacturing systems were never built to support AI in the first place. Legacy enterprise systems—like outdated ERP (enterprise resource planning) tools, old-school CRMs (customer relationship management platforms), and manual data entry—still dominate much of the landscape. When critical insights are buried across disconnected platforms—or worse, written down in logbooks—it becomes nearly impossible to feed AI models the context they need.
Ahmed points to a recent deployment with materials manufacturer Celanese, in which private 5G and edge AI were introduced to improve worker safety and equipment monitoring. “They were able to identify fatigue risk factors and detect hazards in real time,” he claims. “It was only possible because the infrastructure was there to support that intelligence.”
For him, the key to successful AI deployments in manufacturing isn’t just having data, but also having the right data, in the right place, and at the right time. Without that, he warns, “factories will keep seeing failed pilots, no matter how powerful the model.”
Not all use cases are built the same
While the buzz often centers on predictive maintenance and visual inspection, those aren’t plug-and-play features. They require reliable data flow, ultra-low latency, and hardware compatibility that many plants simply don’t have. In remote or offline environments, traditional cloud-based systems can’t keep up.
“Use cases that demand real-time decision-making—like voice-enabled workflows or autonomous quality checks—are especially sensitive to network and system performance,” Asbag notes. “That’s why edge computing matters. It allows speech recognition or LLM-driven tasks to happen on-site, without depending on cloud access.”
Picture a factory line that shuts down every time it loses Wi-Fi. Without local processing—meaning the ability to run AI tasks on devices in the factory instead of sending them to the cloud—even a short loss of connectivity can stop production and make AI tools more of a problem than a help.
For factories operating with limited or unreliable connectivity, edge AI offers a way forward. By processing data locally, companies can cut lag time, protect sensitive data, and reduce downtime. But again, these benefits only materialize if the surrounding infrastructure—from sensors to routers—is up to the task.
“Think of it like trying to run a modern electric vehicle on outdated roads,” Ahmed says. “No matter how powerful the engine, if the path is broken, you’re not going anywhere fast.”
Getting real ROI
One of the biggest traps in AI adoption is mistaking model accuracy for business success. Just because a model performs well during testing doesn’t mean it will drive positive outcomes on the floor.
“The most successful AI initiatives begin with a clear vision—improving quality, boosting efficiency, or unlocking insights,” Ahmed says. “From there, quick wins build momentum.”
Asbag agrees with him. “ROI in AI is not about proving that the model works or that accuracy improves on a benchmark. Those are technology goals, not business goals,” he notes. “Companies should avoid fluff by defining ROI [return on investment] in clear, specific business terms—faster processes, better decisions, or measurable savings.”
That means tracking metrics like how many more inspections a worker can perform with a voice assistant or how predictive maintenance reduced unexpected machine downtime. When AI is tied to concrete, operational key performance indicators (KPIs), it becomes a tool for transformation—not just a tech experiment.
And that’s the big difference between the hype-induced claims of faster operations in the AI space and real measurable impact. It’s one thing to say your model is 96% accurate in a test environment. It’s another to show that it actually helped to cut defect rates by 12% in real production. While the first might get a nod from the technical team, the second gets leadership to sign off on a bigger rollout.
The path forward
Getting AI to work in manufacturing isn’t about chasing the most advanced model. It’s really about understanding the problem, cleaning up the data, modernizing the systems, and making sure every deployment serves a real business need.
“Too many companies fall into endless discussions, pilots, and meetings without ever delivering value,” Asbag says. “Success with AI comes from being precise about the problem, aligning with the business outcome, and giving teams the autonomy to execute.”
Ahmed puts it even more directly: “AI without infrastructure is like trying to build a smart city with no roads. You need the foundation in place before you scale.”
Sateesh Seetharamiah, CEO of EdgeVerve, a leader in AI and automation, also agrees. “Without a defined set of use cases and outcomes, manufacturers will be stuck without a clear strategy to prioritize the right emerging tech capabilities for business success,” he says.
Conversations about building AI infrastructure in manufacturing often stall because leaders assume it means ripping everything out and starting from scratch. But meaningful progress rarely requires a full overhaul. Some of the biggest wins come from small, targeted changes—like installing local edge devices to reduce lag, connecting isolated systems, or clarifying who owns what data so teams can move faster.
Manufacturing may be one of the toughest environments for AI, but it’s also one of the most rewarding. The factories that get it right won’t just optimize how work gets done; they’ll also lead a new era of industrial work—while the ones that hesitate may fall behind. “This isn’t the time to sit on the fence,” Seetharamiah says. “Manufacturers who delay risk missing out on enormous opportunities to create digital experiences for their customers.”