Unlocking Productivity in Manufacturing With Casual Models and Agentic AI

The most expensive thing happening inside your company right now isn't a technology failure. It's the time leaking out between a decision and the permission to act on it.

In my conversation with Michael Carroll on the AI in Manufacturing podcast, we dug into a question that should unsettle every manufacturing executive: why has total U.S. manufacturing productivity been flat—and at times declining—since 2008, despite massive investment in digital tools? Carroll, a global executive in industrial innovation who spent 15 years driving transformation at Georgia-Pacific and now serves as strategic advisor at LNS Research and Chief Strategy Officer at Trek AI, offered an answer that has nothing to do with buying better software and everything to do with how we've structured the relationship between people, decisions, and permission.

‍

‍

Why Has More Data Led to Worse Manufacturing Performance?

‍

The conventional assumption is straightforward: more sensors, more analytics, more insights should equal better performance. The data says otherwise. The St. Louis Fed chart shows productivity flattening around 2010 and declining in stretches through 2023. Carroll and his collaborators initially hypothesized that rising workforce attrition—climbing from 16.5% to over 25% year-over-year—was draining institutional knowledge faster than technology could compensate. That seemed plausible until COVID broke the model.

‍

When COVID hit, even more people left the workforce. If the knowledge-drain hypothesis held, productivity should have cratered. Instead, it spiked upward. Carroll's team had to throw out their first explanation and ask harder questions about what actually changed. The answer was disarmingly simple: companies turned off the badges. Corporate visitors stopped coming to plants. The phone calls dropped off. People got focused. And productivity went up.

‍

This pattern wasn't new. Carroll described learning from General Mills, where plants slated for closure consistently saw 15-20% productivity gains. Not because of new technology. Because of focus. The lesson was uncomfortable: the primary drag on manufacturing productivity wasn't insufficient technology. It was distraction—much of it created by the technology itself.

‍

What Is the Cognitive Tipping Point Destroying Frontline Productivity?

‍

Today's frontline operators see eight times more information across 50% more equipment than at any point since 1975. They also have 50% less experience than their predecessors in the equipment they're responsible for. Carroll calls this a cognitive tipping point—the moment where more insights stop helping and start hurting.

‍

The compounding effect is brutal. MES and ERP systems need data, so companies transferred transactional work to the front line. Operators who should be using their adaptive capacity—reading conditions, making adjustments, keeping downstream processes running—are instead task-switching between data entry, system navigation, and actual operations. Carroll put it bluntly: "We're asking them to text and drive." The safety data confirms it. Incident rates climbed during the same period that digital investment accelerated.

‍

This isn't a training problem or a change management problem. It's a structural problem. The architecture of how we deliver information and demand transactions from frontline workers is consuming the very cognitive capacity those workers need to perform. Every additional dashboard, every new data entry requirement, every extra system to check is another tax on the people least equipped to absorb it—because they're the ones with the least experience.

‍

What Does It Cost When Decision Velocity Stalls Inside Permission Architectures?

‍

Carroll pointed to a 2003 study by J. Robert Baum examining 318 companies across industries. The highest-performing companies shared three characteristics: they made decisions faster than competitors, they centralized strategy while decentralizing operations, and they only standardized what was easy—refusing to over-index on standardizing complex, context-dependent work. The finding that matters most: decision speed and performance were proportional. Not decision quality. Decision speed.

‍

Most manufacturing organizations today are permission architectures. The value leaks out in the staircase between "when do I need to act" and "do I have permission to act." Carroll described a vicious cycle familiar to anyone who's worked at the intersection of IT and operations: the operations team needs capability, so they ask IT for permission. IT needs to add capacity and capability before granting that permission. Both sides are waiting on the other while competitive advantage evaporates.

‍

Look at your calendar, Carroll suggested. Count how many meetings exist purely for alignment. Probably half or more. Those alignment meetings exist because the numbers aren't trusted, so politics fill the gap. "You know you have a company that can't trust its numbers when you have an alignment meeting," he said, "because alignment meetings mean the politics matter more than the numbers."

‍

How Does Agentic AI Differ From Traditional Manufacturing Software?

‍

Agentic AI is not a faster version of your existing software stack. Carroll described it as a fundamentally different species—one that treats decisions, not transactions, as the atomic unit of work. The distinction matters because it changes what the technology is actually doing.

‍

An agent, in Carroll's framing, is something that shapes an outcome on your behalf while you retain responsibility. Something that automates a task is not an agent. Something that carries out a predefined workflow is not an agent. An agent evaluates what's true about the world, considers what interventions could create a desired outcome, selects an action, observes the evidence, and learns. That's the creation of agency—and it mirrors exactly what experienced operators do when they're not buried in transactional work.

‍

The critical enabler is causal reasoning, rooted in the work of Turing Prize winner Judea Pearl. Where an LLM produces correlations based on vector math—only as good as yesterday's patterns—causal models produce chains of reasoning that explain *why* a decision was made, not just what was decided. This distinction is everything in a governance-heavy environment. An explanation is litigable; everyone interprets it differently, which is precisely why alignment meetings exist. A chain of reasoning is defensible. It carries the logic of why a decision was made within the guardrails the organization has already established.

‍

For the operator, this means the permission arrives with the recommendation. The chain of reasoning is auditable. The frontline worker becomes the feedback mechanism—challenging, adjusting, validating—rather than the bottleneck waiting for approval from three levels up.

‍

What Architecture Do Manufacturers Need for Causal AI at the Edge?

‍

Carroll described a world becoming organized around three layers: a durable core that defines what's true about the company and what it does to make things true, an architecture of trust and permission in the middle, and disposable context applications at the edge that get rewritten as conditions change. Connectivity becomes ubiquitous. Reasoning happens at the edge. Where decisions are high-fidelity and high-risk, causal reasoning provides the trust layer. Where temporal proximity is close and stakes are lower, lighter approaches suffice.

‍

This is a direct challenge to the dominant architecture pattern of centralizing everything into cloud data lakes for data scientists to analyze. Carroll's critique was pointed: if nationwide data shows performance getting worse as data accumulation increases, "probably more data models is not going to help." Instead, he argued for collecting only the variables that matter for future decisions and only the data needed as evidence to prove or disprove hypotheses about those decisions. The subscription model for that, he said, is causality—because causal models tell you what variables matter and which ones don't. You stop being a news organization reporting what yesterday looked like and start building optionality for decisions you'll need to make tomorrow.

‍

Where Should Manufacturing Leaders Start Mapping Decision Friction?

‍

Carroll's practical advice for COOs and VPs of operations was refreshingly concrete. First, map how you currently get things done—the actual flow, not the org chart version. Then identify where the inferencing load is highest: where people must gather multiple streams of information, synthesize them, and convert them into decisions through interaction with others. Next, overlay where the permission load is highest in that same map. Where inferencing and permission loads converge, that's where value is hemorrhaging from your organization.

‍

The fix starts with the permission structure, not the technology. Ask why so many permissions exist. Then ask which inferencing burdens can be automated through AI that produces defensible chains of reasoning rather than just explanations that require more alignment meetings. The organizational structure likely needs to change—not to eliminate people, but to remove people from decision gates where they contribute opinions rather than accountability. "If somebody has to be involved in a decision that is not responsible," Carroll said, "you need to start removing these folks. Not because they don't add value, just because they don't need to be a part of the permission decision."

‍

How Should Decision-Makers Reframe the AI Question for Manufacturing?

‍

Stop asking "what AI tools should we deploy?" Start asking "where is time collapsing between decisions and actions, and where is it artificially expanding?" The strategic question isn't about technology selection. It's about whether your organizational architecture—the interplay of structure, governance, and permission—is designed for a world where everyone and everything is one degree of separation apart, or for a world that no longer exists.

‍

Carroll's reframe is that roughly 20% of value comes from doing the right things, 20% from doing those things right, and a full 60% from staying focused on doing them. If your architecture generates more distraction than focus—through permission stairways, alignment meetings, and cognitive overload at the front line—no amount of digital investment will close the productivity gap. The companies that win will be the ones that collapse the distance between insight and action, not by moving faster through the same broken structure, but by rebuilding the structure around decision velocity and earned trust.

‍