Jacey Heuer, Head of AI at Pella Corporation, Answers Your Questions about Agentic AI

In a recent IIA session, Jacey Heuer — head of AI, data science, and advanced analytics at Pella Corporation — sat down with IIA’s Jason Larson for a candid practitioner account of what it takes to get an enterprise using AI: not in a pilot, but in production, and increasingly through agents that are reshaping how operational decisions get made on the plant floor. Pella is one of North America’s largest premium window and door manufacturers, with more than 20 manufacturing sites, over 10,000 employees, and a 101-year history, headquartered in Pella, Iowa.

What makes Pella a distinctive test case for enterprise AI is the complexity hiding inside a deceptively simple product. As a large-volume custom manufacturer, Pella produces units with something on the order of eight octillion possible combinations of attributes and features — meaning the supply chain, value stream, and manufacturing behind any given window or door are anything but simple. That complexity, Heuer noted, is precisely why AI matters so much to the business: it’s what helps the organization manage it. For instance, Pella’s award-winning agentic AI system, AI Maintenance Doctor, is now deployed across 14+ of its plants.

But as Heuer was quick to point out, the path from pilot to production to agentic was not frictionless. The conversation that followed focused on that friction — the organizational obstacles, the workforce-readiness work, the change-management decisions that made the difference, and where Pella’s agentic roadmap goes from here. The Q&A below captures that discussion, including an honest look at what failed before something worked, and what Heuer believes separates the enterprises that compound AI value from those that stay stuck cycling through pilots. 

What did Pella’s data and analytics operating baseline look like before its first steps into enterprise-grade AI?

I’ve been at Pella for just over six years, and the change in that window has been dramatic. When I joined, we were running on-prem databases — some people still had towers under their desks running their own little local databases. Today we’re a multi-cloud environment with strong connectivity and movement of data from the shop floor up through every layer into our cloud. We went through a real modernization in a relatively short span, and the through-line of all of it is the same: the foundation is data. Access to data, access to information.

Part of that journey was cultural, and it ran against our own grain in an interesting way. Pella was born on lean manufacturing principles and still adheres to them, and lean is historically about taking out waste. The complication is that any additional transaction added to the manufacturing process used to be seen as waste — so it got stripped out. Over the last several years we’ve deliberately gone the other direction, reintroducing sensors and data capture into different parts of the process, because that data is exactly what gives an AI the rich context it needs to understand the environment and be effective.

The other significant change has been in our people. We’ve leaned much harder into true cross-functional and product teams. On the manufacturing side, we’ve built what we call an intelligent operations environment — rooted in Industry 4.0 concepts — where operational technology engineering, IT, AI engineering, data science, and data engineering all sit on the same team. That blend has significantly accelerated our ability to build the data foundation layer, put AI on top of it, understand the business requirements, and actually deliver outcomes to the organization.

Who’s building these agents — the general workforce, or AI specialists?

It’s a blend of both, and that blend is by design.

I’ve worked at a few different organizations over my career, and one thing Pella does especially well at the enterprise level is set strategy clearly and then align initiatives and projects down through every level to pursue it. The governance structures are in place, the prioritization structures are in place — the mechanisms to decide yes or no on different efforts and missions, and then allocate resources to deliver those strategic outcomes.

Out of that process, you get a balance. On one side, you have the broad enterprise driving the creation of agents to help in their own day-to-day work. I’ve got someone on my team who owns our AI literacy and workforce enablement, educating people on Microsoft Copilot — and we’re at 97% adoption, roughly 1,200 active users, across the population that isn’t on the manufacturing floor. For an organization Pella’s size, that’s significant. Those people are building agents that accelerate how they work and drive real value.

On the other side, you have projects that represent true differentiation or innovation. Those tend to be more complicated, and they may not be solved by an agentic AI solution alone. A good example is the work we’re doing in computer vision — I have someone on my team with a PhD in mechanical engineering and computer vision who’s a genuine expert at building those end-to-end solutions. So it’s that balance: very technically deep AI capability on one end, and enabling the broader business on the other. That’s the environment we’re working to foster.

When you deployed pilots, what was the biggest source of friction to acceptance — the technology, human trust, workflow change, or governance?

All of those showed up to some degree through deployment. But the most common one, in our experience, is demographic. The older cohort, the people with longer tenure, generally tend to be a little more reticent to accept the change — there’s nothing novel in that observation. So the people earlier in their careers are typically where we like to start. We get them on board, build the evidence that maybe there’s something better here, and that, in turn, brings along the people who are later in their careers. That’s been a consistent through-line across our different deployments as we’ve worked to grow the user base for the applications and solutions we’re putting out into the organization.

The other challenge — and I’ll be candid, I still don’t have a perfect answer to it — is tribal knowledge extraction. I inevitably run into people wrestling with the question, “Am I just teaching this thing to do my job?” We go back and forth trying to toe that line. Ultimately, our people and talent are a cultural pillar of the organization, and we genuinely view our pursuit of AI as augmenting and enhancing the people we already have. The goal is to build it as a true partner in the environment — something people can share information with and get value from, a tool that’s part of our broader strategy. But getting people over that hump has not always been an easy task for us.

What metrics tell you you’re becoming more productive and not just generating AI activity?

A couple of thoughts, starting on the software development side. The metric we’re really leaning into is how the bottleneck is moving further down the development value stream. To reframe it: historically, the bottleneck might have been on the developers — building something took a while, took forever, however you want to put it. Now we have real examples of work that used to take two or three weeks being done in an afternoon. So the bottleneck shifts out to our product owners — how quickly can they give requirements? — and now we’re seeing it shift again, to the business users who can state what they need and work with the product owners directly.

That comes to life through lines of code, velocity, the typical agile measurements. We’re tracking where the bottleneck keeps shifting, and it’s challenging some of our traditional perspectives. We’d typically build an application for a human to click through and interface with. What we’re evolving toward is the human doing less and less, as the agent does more and more of the decision-making and the clicking-through. So the real question becomes: how do we keep remedying that bottleneck?

The business side is where the real benefit comes through. Take AI Maintenance Doctor — across the fleet, our machine uptime is up 10 to 15% because of it. That’s a key metric, and the business measures it directly as value, because machine, people, and material all feed our ability to produce units on time for our customers. Machine uptime goes up, and that follows through to on-time completion and on-time shipment — the metrics we really care about on the manufacturing side. Another big area is warranty claims. As we get greater quality out of our manufacturing processes, are those claims changing? The areas where we’ve directly deployed AI to our assembly lines and parts-making have a clear through-line to warranty claims, and we’re seeing significant benefit there. Those are the real, P&L-level impacts we’re measuring.

How are you thinking about talent as the work shifts from traditional predictive analytics toward agentic development?

Honestly, no one has the answer on this — we’re continuing to reflect and wrestle with it. But here’s what I see evolving. It’s increasingly about people who are systems thinkers, who have a genuine willingness to experiment and learn new technologies. And ultimately, I think the skill set that becomes most critical is clarity of thought.

Can I clearly convey the target — the outcome I’m trying to achieve? Because agentic AI is going to keep abstracting away more and more of the technical challenges that exist today. So it increasingly comes down to articulating the ideas, the value to the organization, and what the customer needs, and then leveraging the ecosystem of AI technologies to build toward it. We’re still wrestling with exactly how that evolves through the talent profile. But those traits — critical thinking, clarity of thought — are being elevated in significance for anyone who comes into the organization or grows through it.