The Founder Who Came Out of Retirement, Killed 80 Spreadsheets, and Spent $60M on a Warehouse Before He Touched a Chatbot

THE CRAFT

Steve Schlecht founded Duluth Trading Company in 1989 to sell a sturdier version of a canvas tool bag to tradesmen out of a mail-order catalog, retired from day-to-day operations more than once, and last year, in his late seventies, came back out of retirement to run the company as Interim CEO after the previous CEO stepped down. Duluth Trading now does about $626 million a year in revenue (FY2024), is publicly traded on NASDAQ, and sells everything from its signature “Fire Hose” work pants to a steadily growing women’s line. It is exactly the kind of business this newsletter is written for: founder-led, mid-market, traditional industry, one notch above the ICP ceiling but close enough that every decision Schlecht makes is a decision one of you is going to have to make inside the next 24 months.

Here is the sequence of moves he’s making in 2026 that I want every founder-CEO reading this to study, in order.

First, he committed $60 million in capital to a warehouse automation project with Manhattan Associates, aimed at order accuracy, fulfillment speed, and labor efficiency inside the central Duluth distribution center. Second, he put a project on the tracker to kill more than 80 internal spreadsheets that the product development team was using to run the merchandising calendar, replacing them with a single centralized product lifecycle management system (Centric PLM). Third — and this is the one the tech press is going to completely ignore — he started building a custom product recommender model on the e-commerce side, trained on the behavior of more than four million shoppers across online and retail, feeding personalized product suggestions of up to 20 items per customer.

Notice what’s not in the first two moves. There is no chatbot. There is no “AI-powered customer service transformation.” There is no generative-AI content engine writing product descriptions. There is no board-slide headline about “deploying AI across the enterprise.” The first two things Schlecht did — and these are the things the company is publicly on record about — are boring infrastructure projects that happen to be AI-enabled underneath. The recommender system, which is the closest thing to what a tech reporter would call “AI,” is third in the sequence. Not first.

This sequence is the most important thing in this issue. I am going to walk you through why.

THE OPERATOR

The company and the comeback

Duluth Trading started as Dave Fritz and Steve Schlecht’s side project in 1989, mail-ordering a canvas bag called the Bucket Boss to tradespeople out of a single catalog. Schlecht bought out his co-founder in the 1990s, ran the company privately through the 2000s, and took it public on NASDAQ in 2015 (ticker: DLTH). He moved to Chairman in 2019. He moved to Interim CEO again in 2024 after the previous CEO, Sam Sato, retired. He is, by any reasonable measure, a founder-CEO who has already been through the full lifecycle of private growth → IPO → professional management → strategic stumble → founder return. If that pattern sounds familiar to any of you, it should. You are the person this is written for.

The reason Schlecht came back matters. Duluth Trading had run into the exact problem that defines mid-market apparel retail in 2025: rising fulfillment costs, aging infrastructure, merchandising processes that had outgrown the spreadsheets they were built on, and a customer experience that felt thinner than it had ten years earlier. The previous management team had a tech strategy. The board and Schlecht concluded, reading between the lines of the public filings, that the strategy was not paired with the operational spine needed to execute it. So the founder came back, and the first thing he did was stop the front-end work and look at the back of the house.

The $60M warehouse

Duluth’s central distribution center, in Wisconsin, is the physical spine of the company’s DTC business. The $60M Manhattan Associates project — publicly disclosed in investor calls and case studies — is a warehouse management system (WMS) replacement and automation layer. It handles order routing, pick optimization, replenishment logic, labor planning, and the feedback loop between the e-commerce demand signal and the physical fulfillment capacity. Manhattan’s current product line is, to use the industry’s word, “intelligent” — meaning it embeds machine learning inside the routing and labor models, predicting order volume, anticipating bottlenecks, and dynamically reassigning pickers to zones based on where the next three hours of demand are likely to land. None of this is going to be called “AI” in a press release. It’s called “WMS.” The AI is the substrate underneath.

This is the least sexy possible project for a founder-CEO to lead with in a year when every board deck in America has a slide titled “Our Gen AI Roadmap.” Let me tell you why it was the right one.

A $60M warehouse project changes the cost structure of the company. Depending on Duluth’s specific pre-automation fulfillment cost per order, a modernized WMS with embedded optimization typically reduces per-order variable fulfillment cost somewhere between 8% and 15%, improves perfect-order rate by several percentage points, and — critically — creates the clean operational data layer that every subsequent AI project is going to need to produce real results. Without this project, nothing else Schlecht does in 2026 and 2027 has a chance of producing measurable value, because the upstream data coming out of a manual-ish WMS is too dirty, too latent, and too siloed for any model trained downstream to do useful work on it. The warehouse is not an AI project. The warehouse is the project that has to precede the AI projects, and the operators who skip this step and go straight to the chatbot are the ones whose 2027 AI postmortems will read “we couldn’t get clean enough data to make the model work.”

The 80 spreadsheets

The second move — killing 80+ internal spreadsheets and replacing them with a centralized PLM (product lifecycle management) system from Centric Software — is even less glamorous, and even more important.

Here is the sentence I want to put on a wall. The organizational readiness of a mid-market company to deploy AI is almost perfectly predicted by how many critical business decisions are currently running on shared spreadsheets. Every spreadsheet is a pocket of organizational knowledge that has been encoded by a single employee, usually with no documentation, usually with formulas that only that employee can debug, and almost always with input from other spreadsheets that nobody has mapped. A merchandising team running 80 spreadsheets is a team that cannot be automated on, optimized against, or extracted from — because the knowledge inside the spreadsheets is not in a form that any model, agentic or otherwise, can reach.

Duluth’s PLM migration is not an AI project either. It is the project that creates the conditions for AI projects to work. Once merchandising lives in a centralized system, the product development calendar, the cost structure, the vendor-to-SKU relationships, the color and size breakouts, the margin targets — all of that becomes queryable, exportable, and most importantly learnable from. A model trained on that data can do real work. A model trained on Sue-in-merchandising’s Excel file with the pivot table that breaks every time she changes the sort order can do no work at all.

So: the first two projects in Schlecht’s sequence are not AI projects. They are the prerequisite data projects that make the AI projects possible.

The recommender (the actual AI)

Only now, third in sequence, does Duluth put a real machine learning model into production. Working with a partner called Bounteous, the company built a custom recommendation model trained on the historical behavior of more than 4 million shoppers, producing up to 20 personalized product recommendations per customer across both DTC and retail touchpoints. This is the thing the tech press would call “Duluth’s AI project” if they were going to write about it, which they are not going to, because the story is boring to anyone who isn’t running a mid-market apparel operator.

What’s worth noticing: the recommender is trained on the clean-ish data that the PLM migration made possible, delivered through the e-commerce infrastructure that the warehouse project is making reliable. It is an AI project that was set up to succeed by two projects that preceded it. The recommender is the ornament on top of the tree. Schlecht spent the first two quarters building the tree.

The Craft of AI read

This is the single most important frame for founder-CEOs of $100M–$500M companies in 2026, and I am going to say it plainly.

The question is not “where should I apply AI first.” The question is “what infrastructure has to be clean before I apply AI anywhere, and which of those projects are my board going to let me do without complaining that it isn’t an AI project.”

Every board in America is going to push their CEO toward the ornament — the chatbot, the generative email writer, the AI-powered landing page — because those projects are legible to the board and look like progress. Every one of those projects will fail or underperform if the underlying operational data is dirty, siloed, or stuck in a spreadsheet. The founder’s job, in 2026, is to sequence the unsexy prerequisite projects in front of the sexy AI projects, and to defend that sequencing to a board that will want it reversed.

Schlecht is doing this in public, on the investor call, with actual capital allocation behind it. $60M to the warehouse. Months of team work killing spreadsheets. Only then does the recommender ship. You should read the Duluth 2026 capital plan as a template. It is the template.

The three specific decisions to steal:

1. Put the “AI strategy” slide at the back of the deck, and put “data prerequisites” at the front. When you present your 2026 plan to the board, make “clean operational data” the first project and “AI applications” the fourth. The board will push back. Push back harder. Every operator who reversed this order in 2024 and 2025 has a war story by now.
2. Audit your spreadsheets the way a lawyer audits contracts. How many shared spreadsheets run critical business decisions? If the answer is “more than 20,” you are not ready to deploy AI on top of those decisions, and any vendor who says you are is selling you a lie. The number is almost always higher than you think. The last founder-CEO I watched do this audit was stunned to learn her merchandising team was running 47 spreadsheets — she had thought it was 10.
3. Be the founder who says “we will do the boring project first.” This is the hardest one, because boring projects do not generate press releases and do not impress LinkedIn, and you will feel like the slowest CEO in the room. Steve Schlecht is 75 years old, back out of retirement, and taking the slow-first path in public. If he can hold that line, so can you.

Things to consider

1. Count your spreadsheets this week. Ask every functional head how many shared Excel or Google Sheets files their team uses to run a critical weekly decision. Write down the number. I promise the number is bigger than you think. That number is the size of your pre-AI data debt.
2. Find the single most physically inefficient process in your company. For Duluth it was the warehouse. For you it might be field service routing, production scheduling, order entry, or dispatch. That process is your warehouse. That is where the first capital goes, and where the first “invisible AI” — the kind embedded inside a vendor’s routing or optimization layer — is going to produce real ROI that funds the later projects.
3. Look at your last three “AI initiative” slide decks. How many of them would have been possible without the data prerequisite? If the answer is “most of them assumed the data was clean,” you have a sequencing problem, and the projects are going to underperform or fail. Reorder the deck.
4. Benchmark your warehouse/operations cost per unit against the best operator in your category. Not against yourself year-over-year. Against the best. The gap is what the first AI-enabled infrastructure project is going to close. If you don’t know the benchmark, that’s your first homework.
5. Ask whether your board is a boring-project board or a sexy-project board. This is a read you need to have before the 2026 capital plan conversation. If the board is going to reject the warehouse project because “that’s just ops,” you have a governance problem and you need to address it before you address the AI problem. Schlecht can do this because he’s the founder with board control. If you don’t have that control, the conversation is longer — but the sequencing is the same.

THE WORKBENCH

Here’s the tactical takeaway for your Monday.

Run the Duluth sequence test on your own 2026 capital plan. Take the list of every project you have on the roadmap that touches “AI,” “automation,” “digital transformation,” or “modernization.” Put them in two columns:

• Column A: Prerequisites. Projects that clean up, centralize, or standardize data, infrastructure, or workflows. WMS. ERP. PLM. CRM consolidation. Master data management. Killing spreadsheets. Instrumenting a process so it can be measured.
• Column B: Applications. Projects that deploy a model, agent, chatbot, or automation against existing data. Recommenders. Forecasters. Generative content. Support automation. Sales automation.

Now check two things. First: is every Column B project backed by at least one Column A project that is funded, staffed, and ahead of it in the sequence? If not, that Column B project is going to produce a bad result — not because the AI is bad, but because the ground it’s standing on is unstable. Second: what percentage of your 2026 capital is in Column A versus Column B? Duluth’s is weighted heavily toward Column A. If yours isn’t, you are either further along on data hygiene than most operators (possible, but rare) or you are setting up your AI projects to fail (more likely).

The template here is free, and it’s the one Schlecht is using in public. The hard part is having the institutional courage to do the unsexy project first when every board member in the room wants to hear about the chatbot.

THE QUESTION

Here is what I actually want to know from you this week, and I’d love your reply.

What is the spreadsheet in your company that, if it disappeared tomorrow, would halt a critical decision by Friday — and who is the one person on your team who knows how it works?

That spreadsheet is your AI readiness problem, compressed into a single file. The person is your biggest single point of organizational failure. I am collecting these privately for a future issue, and no names or identifying details will appear in anything I publish. I just want to know how bad the problem is across the audience, because my strong suspicion is that it is much worse than the tech press is saying and much worse than the consultants are telling you.

Hit reply, or send a note to grant@thecraftofai.com. One sentence is fine: “accounting close schedule, maintained by Karen.” That’s all I need. I read every reply.

— Grant grant@thecraftofai.com