Artificial intelligence has surged into nearly every industrial sector over the past three years, but the enterprises moving fastest often aren’t the ones you’d expect. And according to one industry expert, the manufacturers best positioned to extract real value from AI are those that stop thinking about “AI projects” and start thinking about data strategy.
“There’s no AI strategy without a data strategy. Let’s make sure we get all of the right context in the data pieces for your organization ready to go,” says Jeff Hollan, who was director of product at Snowflake at the time of this interview. Hollan arrived at this insight after more than a year working closely with customers deploying AI inside mission-critical operations.
Context is king
The promise of agentic AI—systems that synthesize, reason, and take action on enterprise data—depends heavily on context. Public LLMs can draft emails or summarize meeting notes, but they can rarely answer the questions that matter to a production engineer or plant manager.
“Once I start doing my job, LLMs become far less valuable. For the questions I care about, they just don’t have the correct context. If you don’t have the right context, you’re not able to build those pieces,” Hollan says.
This is where manufacturers, particularly those with fragmented MES, ERP, sensor, and quality datasets, face the first friction point: model performance directly correlates with the organization’s ability to centralize and standardize its data foundation.
Unexpected leaders in AI adoption
Despite assumptions that tightly regulated sectors would be slowest to adopt AI, Hollan has been surprised.
“Because of regulations and complexity, I probably would have put healthcare and financial services, a little bit lower. I’ve been shocked that they’ve been some of the fastest moving industries,” Hollan says.
The expected-value calculation is simple: high-stakes operations see transformative value even in small efficiency gains. That pattern is now emerging in manufacturing as well, particularly in semiconductor and advanced electronics, where scrap and defect rates have direct bottom-line impact.
What manufacturers expect from AI
Across industries, the dominant theme is productivity. How can users be more productive in what they do, and potentially, how can they remove some of the more mundane or automated tasks?
Hollan describes using agentic AI to draft internal documents by combining meeting transcripts and roadmap data, cutting three hours of work down to minutes. That same acceleration applies on the shop floor.
“A major microchip manufacturer spends so much time trying to analyze where do we have potential defects or safety concerns. We can now go spend some AI time to go through a bunch of those cycles and create a bunch of time,” he says.
Across use cases—quality, claims review, analysis, reporting—the pattern is consistent: agents do the first 80%, and people finish the last 20%.
The expectation gap
Where AI falls short is when companies overestimate the system’s autonomy.
AI, in practice, behaves a bit like a new college hire: capable, but dependent on how clearly the problem is explained. If a problem can be clearly described, AI can likely handle it. If not—even humans struggle with poorly defined problems—expectation gaps appear.
What will unlock manufacturing adoption
For manufacturing specifically, the biggest unlock is executive priority. But the second requirement is far more practical: making it easy to start.
Manufacturers face dozens of potential AI applications—quality, supply chain, maintenance, scheduling—and struggle to decide where to begin. Hollan argues that ‘turnkey’ starting points lower the barrier.
“If you can point us to the data, we can help get you up and running with the basic agent without having to write any code,” he says.
That allows manufacturers to begin with small, high-value queries, such as “What were the test results for this SKU over the last three months?” and quickly expand into action-taking workflows.
What about small and mid-sized manufacturers?
Most North American manufacturers aren’t massive, well capitalized multinational firms, yet the same principles apply.
The key is to start with where you have a real pain point today and make sure you’re building on platforms. Critically, the digital baseline required to start is lower than most assume. You could start with just an Excel file or a CSV because AI is very malleable and it can reason around the rough edges.
Clean, structured, perfectly integrated systems are ideal but not a prerequisite. Manufacturers can build value from zero to 50% very quickly, and mature from there.
The bottom line
Agentic AI isn’t magic, and it isn’t turnkey autonomy. But it is an accelerant—one that turns messy, sprawling operational data into actionable insights and measurable time savings.
Start by just doing something, even if it’s nothing more than spending a Friday afternoon with the set of people simply “playing” with some technology.
Manufacturers who start now—not those who wait for a perfect data state—will be the ones defining what AI-enabled operations look like six months from today.