Pella's AI Reads Twenty Years of Maintenance History. Lean Would Have Deleted It.

THE CRAFT

Almost every window Pella makes is different from the one before it.

That sounds like a brochure line until you stand next to the consequences. A Pella window is configured rather than picked off a list (size, frame material, glass package, grille pattern, hardware, finish, screen), and the company reckons the combinations run to something on the order of eight octillion. Twenty-seven zeros. What it means in practice is that the line almost never builds the same thing twice, at speed, on machinery that takes a long time to learn. Pella had trouble holding onto the people who ran it.

Now put that next to the other thing that has been true of Pella for most of a century. It is a lean manufacturer, and a serious one: gemba walks, kata projects, an internal tournament that ranks every plant on safety, quality, delivery, productivity and people, now in its sixth year. Lean is a discipline of subtraction. You find the steps that do not add value for the customer and you take them out, and you keep doing it forever.

Here is the collision, and it is the reason I picked this story. Writing something down is a step. Scanning a code, logging a fault, recording which part went into which frame at which hour. Every one of those is an extra transaction added to the work, and under a strict reading of lean, an extra transaction is waste. So for years, at Pella, it got stripped out. The plants ran efficiently and remembered almost nothing.

Then, over the last several years, somebody had to walk into that culture and argue for putting the transactions back. Not for abandoning lean; lean is still running there, still funded, still celebrated. For adding recording steps back into a live lean program, on purpose, knowing exactly how that argument sounds to people who have spent their careers removing them. The reason was that a machine cannot learn a factory it has no record of. This brief is about that argument, what it bought, and the order Pella did things in, because the AI everyone would want to talk about is the last thing that happened here, not the first.

THE OPERATOR

The situation

Pella Corporation has made windows and doors in Pella, Iowa since 1925, when Pete and Lucille Kuyper bought a company that made a window screen you could roll up like a blind. It is still private and still family-controlled, four generations on, which matters for this story in one specific way: nobody outside the company has ever had to be shown the numbers. There is no filing, no earnings call, no analyst asking why the maintenance budget moved. Whatever Pella did, it did without an audience.

The market it did it in has been getting harder from both directions at once. New single-family construction spending is down about 6.5% year over year, and residential improvement spending is down roughly 10%, according to Census figures released the day before I wrote this. Housing completions, the number that sits closest to somebody actually installing a window, are down almost 17%. In the last downturn, remodelling held up while new construction fell, and manufacturers leaned on the half that was working. This time both legs are bending together.

What that does to a window manufacturer’s P&L is worth being precise about, because it explains the whole shape of the decision. Look at the public comparables and the pattern is the same everywhere: the decline is volume, not price. One of the largest listed door and window makers finished last year with core revenue down about 12%: volume and mix took 13 points out of it while price added 1. The purest available read on door manufacturing, the doors segment of a large public building-products group, went from a 13% operating margin in the first half of last year to 9% in the first half of this one.

Read those two facts together and you get the operator’s actual problem. You cannot price your way out, because price is already doing all it can. You cannot sell your way out, because both demand channels are shrinking. Everything left has to come from what it costs you to build the thing, and in a plant making bespoke product at high variety, most of that cost is people, machine time, and the mistakes that happen when either of them is under strain.

Which brings us to the constraint underneath the constraint. The trade press will tell you the window industry’s hiring problem eased last year, and it did: 77% of window and door manufacturers reported less difficulty finding workers in 2025 than in 2024. I take that seriously and it complicates the easy version of this story. But there is a narrower shortage that is not easing, and it is the one that bites a plant like Pella’s. The people who can diagnose why a complex machine failed, and the operators who can run a line where no two jobs are alike, are leaving the workforce faster than they are being replaced. Federal projections put annual openings for industrial machinery mechanics, millwrights and maintenance workers at roughly 10% of the entire occupation every year, and attribute most of that to people exiting the labour force rather than changing jobs. In Marion County, where Pella is headquartered, the labour force has shrunk by nearly 8% since 2019 while unemployment has sat at 2.6%. Fewer people. Not more unemployed people.

That is the situation: margin has to come from operations, and the operational knowledge is walking out the door in the heads of the people who hold it.

The move

The order Pella did things in is the whole lesson, so follow the sequence rather than the technology.

The first move was not artificial intelligence. It was automation, and I want to be plain about that, because handing the credit to AI hides the part you can actually copy. What Pella built first was a voice-directed work system (the plant simply calls it Voice), and the mechanism is unglamorous to the point of being humble. An operator wears a headset. The system tells them which bin to pull a part from. The operator reads the check digits on that bin aloud to confirm they took the right one. The system gives the next instruction. That is it. Voice-directed picking with check-digit confirmation has existed since the early nineties.

Travis Turnbull, Pella’s VP and CIO, describes what it replaced in terms any plant manager will recognise. “We’re a very data driven organization,” he says, “[and] that data was going to paperwork. The team member had a piece of paper that said, ‘Here are the things to do.’ They were carrying this piece of paper everywhere. And if you just watched them, you could just see the inefficiency.”

Pella piloted it at its centralised parts plant in Pella, Iowa. New operators reached proficiency in a few days instead of three to six months. Over the longer run, errors and bad picks fell by 90%. Then it moved out of parts and onto the manufacturing floor, to operators picking components for the line itself, including the high-volume wood window plant at Carroll, Iowa, and it picked up Spanish-language instruction along the way for a plant with a large Latino workforce.

Two things happened in that build that matter more than the headsets.

The first is that Pella had to reorganise who owned the work. Instead of an IT function taking tickets from an operations function (the arrangement almost every mid-sized manufacturer runs, and the one that guarantees neither side ever fully understands the problem), the company put information-technology and operations-technology engineers onto the same cross-functional teams. One group works on data and insight, one on taking a proven solution and spreading it to the other plants, one on building things that do not exist yet. Turnbull’s reason for it is a single sentence: putting IT and OT on the same creative teams breaks down the natural conflict between them. That is an organisational decision, made before any of the interesting technology, and it is the least copied and most copyable part of this entire story.

The second is subtler and it is the reason I chose Pella over half a dozen better-known names. Every one of those check digits read aloud is a record. Every confirmed pick, every completed instruction, every logged exception is a small piece of evidence that the work happened in a particular way at a particular moment. Voice did not only make picking more accurate. It made the picking legible. It turned an activity that had lived in paper and in people’s hands into something a system could read afterwards.

And that is the transaction lean would have removed.

Jacey Heuer, who leads AI, data science and advanced analytics at Pella, is unusually direct about how cultural that fight was:

“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.”

Only after years of that (sensors added back, capture added back, the work made readable) does the part everyone wants to lead with become possible. Pella built what it calls an AI Maintenance Doctor: an agentic system that reads more than twenty years of accumulated equipment and maintenance history, builds a working model of how its various plant environments actually behave, then watches machinery, diagnoses faults, prescribes the fix, and recommends the parts that fix will need based on what has actually worked before. It runs across more than fourteen of Pella’s plants. The Technology Association of Iowa gave it an AI Breakthrough of the Year award in November of last year.

Notice what that system is made of. It is not made of a model. It is made of twenty years of maintenance records that somebody decided to keep, at a cost, inside a company whose operating philosophy said not to.

The result

Heuer puts a number on it, and he is specific about what earned it. Across the fleet, he says, “our machine uptime is up 10 to 15% because of it.”

For a manufacturer whose margin has to come out of operations because it cannot come out of price, uptime is close to the most useful number available. A window plant running bespoke product at speed loses money in two places when a machine stops: the output that does not get built, and the disruption to a schedule where every job is different and nothing downstream can simply be swapped in. A 10 to 15% gain in running time across more than a dozen plants, in a year when completions are down 17%, is margin that did not have to be found in headcount or price.

The groundwork paid on its own terms too, and I would argue it paid harder. 90% fewer bad picks is 90% fewer wrong parts travelling downstream into a bespoke assembly where the error surfaces late and expensively. Getting a new operator productive in days instead of three-to-six months changes what a plant can survive: it means a retirement, a resignation or a seasonal ramp stops being a quarter-long hole in capacity. In a county that has lost eight percent of its labour force since 2019, that is not a productivity statistic. It is the difference between being able to staff the line and not.

Now the honest take.

A 10 to 15% gain is a good number. It is not an extraordinary one. Predictive maintenance has been producing results in that band for years. The modern consulting benchmark for digitally enabled reliability sits at 5 to 15% asset availability, and a US Department of Energy guide has claimed 35 to 45% downtime reduction for predictive maintenance since 2004. Pella is at the top of the ordinary range, not past it. And Voice, as I said at the start, is thirty-year-old technology whose accuracy ceiling has been marketed since before most of these plants installed their current machines.

There is a second thing a fair reader should weigh. Pella was doing several things at once. The internal plant tournament ranking every site on productivity ran right through this period, as it has for six years. The company bought one business in 2023, divested another in 2024, bought Weather Shield in 2025, and has been spending capital on plant infrastructure throughout. Any one of those changes a fleet-wide average without a single process improving. Nobody at Pella has published the arithmetic separating the AI’s contribution from the continuous-improvement programme running alongside it, and a private company is never going to.

And the business itself is not having an easy year. In December, Pella closed three installation operations in Florida, cutting 52 jobs, and the company’s own explanation was that demand had declined despite efforts to improve sales and reduce costs, leading to financial losses. That is the market I described at the top, arriving at a specific set of doors. Uptime being up 10 to 15% does not make a company immune to a housing cycle. It makes it cheaper to run while the cycle does what it does.

What survives all of that is the thing I actually want you to take: the sequence. Pella’s results are ordinary results, executed across more than a dozen plants, on a foundation of recorded work that most manufacturers do not have and could not assemble quickly if they decided tomorrow that they wanted it.

The Craft of AI read

Here is what I would take from this if I ran any business where the expertise lives in people’s hands and heads rather than in a system.

The ceiling on what AI can do for you is set by what your process remembers. Not by which model you license, not by how many seats you buy, not by the sophistication of whatever you are being pitched this quarter. An agentic system that diagnoses a machine fault is doing something conceptually simple: it is reading what happened the last several thousand times something like this happened, and drawing on it. If those several thousand occasions were handled by an experienced technician who fixed the problem and went to lunch, there is nothing to read. The intelligence has no purchase. You will buy the tool, run the pilot, get a polite result, and conclude that AI does not work in your environment, when what actually happened is that your business never wrote anything down.

Pella’s twenty years of maintenance history is the asset in this story. The agentic system is the thing that finally made the asset pay.

And the uncomfortable part, the part worth your Sunday evening: the discipline that makes a company efficient is often the same discipline that makes it illegible. Lean is not wrong. Taking non-value-adding steps out of a process is one of the few management ideas of the last fifty years that reliably works. But a recording step looks exactly like waste right up until the moment you need the recording, and by then the decade in which you could have been collecting it has already gone. Pella did not resolve that by abandoning lean. It resolved it by making a deliberate, arguable, culturally awkward exception: add these transactions back, accept that they cost something, because what they produce is worth more than what they cost.

Somebody had to make that argument internally, to colleagues trained for years to delete exactly what was being proposed. That is the move. Not the headsets, not the agents.

It is the same order Tom Shorten ran at HSS ProService, where the rebuilt ordering and billing systems had to work before a single agent was pointed at anything, and the same order Dave Peacock ran at Advantage Solutions with three unglamorous years of consolidating systems first. Different industries, different technology, same sequence. Make the work knowable. Then make it intelligent. The companies getting real results from AI this year are, almost without exception, the ones who did the first part somewhere between two and five years ago, usually for a reason that had nothing to do with AI.

Things to consider

• Your AI’s ceiling is whatever your process wrote down. Before you evaluate a single vendor, ask what record exists of the work you want the AI to improve. If the last three years of that process live in email threads, in a spreadsheet somebody maintains privately, and in the judgement of two people who have been there a long time, then no model will help you yet. The honest first project is not an AI project. It is making the work produce a record as it happens, which, unlike the AI, will pay for itself in fewer errors and faster training whether or not you ever build the clever thing on top.

• Efficiency and legibility pull in opposite directions, and nobody warns you. Every improvement discipline worth running (lean, six sigma, whatever your version is called) rewards removing steps. Recording is a step. Which means the better you have been at continuous improvement, the more systematically you may have deleted your own evidence. That is not an argument against the discipline. It is an argument for one deliberate exception, made consciously, with somebody senior willing to defend it to people who will correctly point out that it looks like waste.

• Do the unglamorous capture first, and judge it on its own numbers. Pella’s voice-directed system is not AI, and the company does not pretend otherwise. It still cut bad picks by 90% and took new-operator ramp from months to days. If the groundwork only makes sense as a prerequisite for a future AI project, you have not built the business case properly. Build the capture so that it pays standing alone. Then the AI on top is upside rather than justification.

• Ordinary results executed everywhere beat frontier results executed once. A 10 to 15% uptime gain is not a number that wins a conference keynote. Getting it across more than a dozen plants is. The mid-market pattern I keep seeing is the reverse: one spectacular pilot in one facility, presented to the board, never replicated, quietly retired eighteen months later. Pella’s structure tells you they were thinking about this from the start: one of their cross-functional teams exists specifically to take a solution that already works and spread it to the other plants.

• Put the technology people and the plant people on one team, not on two sides of a ticket. This is the cheapest thing in the entire story and the one most likely to be ignored, because it costs no capital and requires an actual organisational decision. Pella stopped having operations file requests with IT and started staffing shared teams that owned the problem together. Every manufacturer I talk to describes some version of the conflict that arrangement removes. Almost none of them have removed it.

THE WORKBENCH

Do this tomorrow

An hour, one workflow, and somebody from the floor in the room with you.

Map what your process forgets. Pick the workflow where expertise matters most, the one where you would be genuinely worried if two specific people left. Walk one real job through it end to end on a whiteboard, step by step, the way it actually happened rather than the way the procedure says. At every step, ask one question out loud: when this step is done, what record exists that it happened this way? Mark each step with a symbol. A tick where a system captured it. A cross where the only record is that a person did it and knows they did it. Do not skip the steps where somebody made a judgement call. Those are the ones you most want an AI to learn from, and they are almost always crosses.

Now count the crosses. That count is the real answer to “are we ready for AI,” and it is a more honest answer than any assessment you will be sold.

Then price what the crosses cost you today. Take the three biggest crosses and put a number on each, using this year’s figures rather than a hypothetical. How many hours a month do your people spend rediscovering something the business already knew once? How much rework, scrap or credit came out of a decision made without information that existed somewhere in the building? How long does a new person take to reach full productivity, and what does that ramp cost you in loaded salary before they are earning it? Add them up and write the annual figure at the top of the board.

That number is not waiting on a technology decision. It is coming out of your operating margin this quarter, in a workflow nobody has ever put a price on, and it is the number that tells you whether your next dollar should buy intelligence or should first buy a memory.

THE QUESTION

Travis Turnbull did not begin with a model. He began by watching people carry pieces of paper around a factory and concluding that the inefficiency was visible from across the room. What his team built first was a headset that tells an operator which bin to reach into and asks them to read two digits back. Technology so old it predates most of the machines it now runs alongside. What it produced, quietly, over years, was a record of how the work actually happens. Only then was there anything for an agentic system to read, and only then did twenty years of maintenance history turn into a 10 to 15% gain in running time across more than a dozen plants.

So here is the question, and it is worth being honest with yourself for thirty seconds rather than answering it the way you would answer it in a board meeting. If you pointed the best AI available at the workflow that matters most in your business tomorrow morning (the claim, the order, the schedule, the repair, whatever actually carries your revenue), what would it find to read?

Not what would you want it to find. What is actually there. A decade of structured history about how that work goes right and how it goes wrong, or three years of email, one heroic spreadsheet, and the accumulated judgement of a handful of people who are closer to retirement than to their first day.

If it is the second, then the constraint was never the technology, and no amount of budget applied this quarter will change that. The good news is the fix does not require anyone’s permission or anyone’s platform. It requires deciding that the record is worth what the record costs, and then defending that decision to a room full of people who have spent their careers being told to take exactly those steps out.

Pella made that argument years before it had anything to show for it. That is the part nobody puts in the case study.

---

Want this done for you?

You just ran the whiteboard version: one workflow, an hour, and a count of the steps your business does not remember.

My Ground-Up Workshop is the rigorous version of the same exercise, run on your business with the small group of your people closest to that workflow. Every AI strategy you’ve been sold starts with software. This one starts with the people who do the work.

One to two days in person, then a synthesis week, and you walk out with a plan: a ground-truth map of how your core workflow actually runs, the one where a rebuild would release real operating margin, a target operating model with AI built in from the ground up, and a 30-60-90 day plan your own team can start on. I do a small number of these a quarter, for $20,000.

I’m a Partner at geniant. Since 2003, our craftspeople have done one thing: understand how work actually happens, design how it should happen, and build software that works the way people do. When a workshop gives you the roadmap to execute or turns into a build, they’re who does it — one senior-led team, layered on top of the systems you already run, value in weeks rather than months.

Book a discovery call →

— Grant K. Baldwin