A Practical Guide to Implementing Industrial AI Agents in Factories

Most manufacturers assume their data infrastructure is the primary barrier to deploying AI. They're wrong. The real constraint is knowledge—the procedural, tribal, and contextual understanding that lives in the heads of your best engineers and technicians, not in your databases.

In my conversation with James Zheng, co-founder and Chief Product Officer of Optimate AI, on the AI in Manufacturing podcast, we unpacked why the agentic AI opportunity in manufacturing isn't about adding chatbots to existing systems. It's about building an entirely new decision intelligence layer that captures how your best people think, reason, and solve problems—then scales that capability across your entire operation.

‍

‍

Why Haven't Two Decades of Digital Investment Solved Manufacturing's Productivity Problem?

The numbers tell a damning story. According to U.S. Bureau of Labor Statistics data, total factor productivity in manufacturing from 2008 to 2023 has been flat—even declining in some periods. This is the same window during which the industry poured billions into industrial IoT, machine learning, automation, MES, and ERP upgrades.

So where did the money go? James's observation, drawn from a career that spans SAP, PTC, and years of consulting, is blunt: the technology investments addressed capital and tooling, but they never addressed the actual limiting factor. That factor is skilled labor—not frontline operators doing physical work, but the engineers, technicians, CI specialists, quality engineers, and frontline leaders who represent maybe 10-20% of cost of revenue but determine quality, safety, efficiency, and delivery for the entire operation.

James references what Augmentir CEO Russ Fadel calls the "skilled labor crisis." It's not just that these people are hard to find. It's that the entire operating model still depends on them personally interpreting dashboards, reasoning through root causes, and carrying institutional knowledge that exists nowhere else. Every tool we've built in the last two decades still assumes a highly skilled human is sitting at the end of the pipeline, making sense of it all.

‍

What Changed to Make Legacy Approaches to Manufacturing Intelligence Obsolete?

The traditional manufacturing technology stack was built in layers that made sense at the time. James describes his career as essentially two phases. The first phase—the SAP era—was about collecting data and creating a single source of truth. MES, ERP, historians: systems of record. The second phase—the PTC and industrial IoT era—was about turning that data into dashboards and reports. What happened, what's happening, what will happen next week.

Both layers delivered value. But they shared a critical assumption: that there would always be enough skilled humans to act on the insights. That assumption has collapsed. The workforce that was supposed to consume all those dashboards is shrinking, aging out, and becoming impossible to replace at the rate manufacturers need.

Meanwhile, the complexity hasn't decreased. Every company is different. Every factory is different. Even within the same factory, every line behaves differently, and the problems on Monday are completely different from the problems on Friday. James makes the point sharply: manufacturing is far less standardizable than customer support or financial operations. You can't write a universal playbook. The knowledge is local, contextual, and constantly shifting.

‍

What Is the Real Cost of the "Skilled Labor Crisis" on the Factory Floor?

Consider what actually happens when a technician gets a machine alarm. James walked through the workflow in detail, and it's worth sitting with. The technician checks vendor manuals. They search Google or ChatGPT. They pull up historical records—maybe a binder of copper records, maybe a CMMS system. They check temperature and pressure trends in the historian. They talk to peers, or they find Bob, who's been running that machine for 30 years. Then they check spare parts inventory in yet another system. If the part isn't in stock, they trigger a purchase request in still another system.

This process routinely takes 30 minutes to several hours. And it happens every day, in every factory, across every shift.

When James's team runs time studies with customers, they find that a large portion of a frontline leader's 40-hour week is consumed by non-value-added activities. A significant chunk is just finding information. Another significant chunk is administrative work—collecting data, generating reports, closing out work orders. One customer had 20 technicians closing 300 work orders per month, each taking 10 minutes of pure administrative effort. That's 50 hours a month of skilled labor spent on paperwork.

The compounding effect is what kills you. Every hour a skilled worker spends hunting for information or filling out forms is an hour they're not solving the next problem. And when that worker retires, the tribal knowledge they carried walks out the door permanently.

‍

How Does an AI Decision Intelligence Layer Work Differently Than Adding Copilots to Existing Systems?

James draws a clear distinction between two approaches to agentic AI in manufacturing. The first approach adds copilot capabilities inside existing tools—a chat interface in your MES, AI-assisted code generation in your IoT platform. This makes your current stack incrementally better.

The second approach is fundamentally different. Instead of giving workers ten systems with ten copilots, you create a single orchestration layer that works the way a skilled human works: accessing multiple systems, reasoning across data sources, following troubleshooting procedures, and handling the administrative aftermath.

This is what James calls the decision intelligence and execution layer. It sits on top of your systems of record and your systems of intelligence. It doesn't replace your MES, ERP, CMMS, or SCADA. Those become tools that the AI agent uses, exactly the way your best technician uses them today.

The practical difference is striking. James described a proof-of-value deployment where a reliability team was manually exporting two spreadsheets—one of work orders, one of parts consumed—then comparing them column by column to identify spare part gaps. The traditional approach would require a developer to write Python or SQL scripts, normalize the data, and build a report. The agentic approach lets someone upload both spreadsheets and ask in natural language: "What are the top three spare part gaps?" The AI reasons through the column mappings, writes and executes code on the fly, validates its own output, and delivers the answer. No data model required.

This is the mindset shift James emphasizes most forcefully: you don't need perfect data to get started. Legacy machine learning demanded clean, structured, perfectly modeled data before you could do anything. Large language model-powered agents operate probabilistically, like humans. As James put it, "If you and I don't need perfect data to make a decision and take action, the AI agent doesn't need it either."

‍

How Are Manufacturers Actually Deploying AI Agents on the Shop Floor?

The pattern James describes starts simple and compounds. Step one is knowledge assistance—a natural language interface that gives the factory team a single point of access to everything they need. Deploy it against a bottleneck process. This alone can save an hour per person per shift just by eliminating information hunting.

Step two is task automation—closing work orders, converting customer specs to internal specs, generating reports. James cited a quality engineering workflow where converting a customer specification into an executable internal spec took two days. With agentic automation, it dropped to 20 minutes.

Step three is agentic workflows—AI that monitors a line, runs diagnostics when something goes wrong, and guides a technician through corrective action step by step. Optimate AI pre-builds common workflow patterns like structured troubleshooting (scope the problem, derive hypotheses, validate with data, implement corrective action, standardize) so teams aren't starting from scratch.

One customer example illustrates the scale of impact. A discrete manufacturer serving AI data center construction had workers spending 8-10 hours on box wiring jobs involving 70-80 terminations. Each individual termination took only two minutes of actual work, but interpreting wire diagrams inflated the average to 10-15 minutes per connection. An AI coach sitting alongside the worker, guiding them step by step through the diagram, could cut that time from 10 minutes to 3-5 minutes—a massive reduction in cycle time for a company with years of backlog.

The critical insight about factory deployment is that the people doing the daily work must be the ones building and shaping the agents. Manufacturing is too variable, too contextual, and too dynamic for a centralized team to define every workflow. James is emphatic on this point: "The only way is that we have to put this power in the hands of the people doing the daily job."

‍

Why Should Manufacturing Leaders Rethink Both Top-Down and Bottom-Up AI Adoption?

The strategic question isn't "how do we implement agentic AI?" It's "how do we scale the reasoning capability of our best people across the entire operation?"

James made an observation that stopped me: in his entire career in manufacturing technology, he has never seen a technology that pulls adoption from both the top and the bottom of an organization simultaneously. ERP and MES were top-down management systems. The frontline always asked, "What's in it for me?" Agentic AI is different because it delivers immediate individual productivity gains—not just enterprise metrics. Factory workers who use ChatGPT and Gemini in their personal lives are already asking for something similar at work.

This changes the adoption model entirely. Yes, you need top-down governance. But the real magic, as James describes it, happens when you put the tools in the hands of your frontline teams and let them explore. The combination of top-down structure and bottom-up innovation, iterating quickly, is what separates this wave from everything that came before.

If you're evaluating AI investments for your factory, stop asking which system needs a copilot bolted on. Start asking where your best people's knowledge is trapped, where your skilled workers are spending hours on tasks that should take minutes, and what happens to your operation when those people leave.

‍

‍