Ariat Cut Warehouse Steps From 40,000 a Day to 7,000 — And Never Once Called It an 'AI Strategy'

THE CRAFT

The most interesting AI deployment at a founder-led mid-market company in 2026 is one that nobody is calling an AI deployment.

Beth Cross is the co-founder and CEO of Ariat International, a 36-year-old performance footwear and apparel company based in Union City, California. Ariat makes the boots the U.S. Olympic equestrian team wears, the boots most working cowboys in America are now wearing instead of their grandfather’s Justins, and a growing wall of steel-toed safety footwear that ships into warehouses and oil fields. It is privately held, founder-controlled, and by most available estimates doing somewhere around $1 billion in revenue — a little above the hard ceiling of this newsletter’s ICP, but close enough that the playbook travels, and honestly the playbook is the only reason I am writing this.

Here is the sentence I keep coming back to. In the last two years, the average Ariat distribution-center worker’s daily step count dropped from 40,000 steps to 7,000 steps. That’s an 82.5% reduction in how much a human body has to move to fulfill the same number of orders. If you want a single metric that captures what AI-plus-robotics actually looks like when it’s installed by an operator who knows what she’s doing, that’s the metric. Not productivity per labor hour. Not “AI spend as percentage of opex.” A number that a 58-year-old pick-and-pack worker can feel in their knees at the end of a shift.

Now here’s the part I want you to sit with: Beth Cross did not set out to deploy AI. There was no offsite. No consulting engagement with a Big Four “AI transformation” partner. No strategy memo titled “Ariat in the Agentic Era.” The warehouse automation project that produced the 40,000→7,000 number started as a conversation about worker retention and physical injury in a labor market where she couldn’t find people willing to do the job. The AI showed up because the AI was the only way to solve the problem. The strategy document, such as it exists, is the problem itself.

I want to walk you through how this happened, because for a founder-CEO of a $100M–$500M traditional-industry business, this is the cleanest template I have seen for how a first AI deployment should actually look — and it’s the exact opposite of the slide deck your board advisor is going to hand you in Q3.

THE OPERATOR

The company and the founder

Beth Cross co-founded Ariat in 1990 with Pam Parker, a fellow Stanford Graduate School of Business classmate. The founding insight was embarrassingly simple: the equestrian boot category had been ignored by every serious athletic-footwear company, and riders were putting up with stiff, heavy, no-support leather boots that hadn’t been redesigned since the 1940s. Cross and Parker put athletic-shoe construction — shock-absorbing midsoles, pronation-controlled heel cups, breathable linings — inside a traditional western silhouette. Competitive riders switched, and then working cowboys switched, and then oilfield and ranch workers switched, and then the line extended into workwear. Cross has been CEO the entire time. Pam Parker is still on the executive team. Ariat has never had a non-founder CEO.

That matters for this story, because the thing Beth Cross does well — and she has been doing it for 36 years — is watch one worker do one job and ask what the hardest part of their day is. It is a product-development discipline. It’s also, it turns out, an operations discipline. And in the last two years, it became an AI-deployment discipline, without her ever having to rename it.

What she actually installed, and why

The Ariat distribution center in the Midwest — the main one, which handles the majority of the company’s online and wholesale fulfillment — used to run on a standard pick-pack-ship workflow. A worker would walk to a shelf, scan a box, put it in a cart, walk to the next shelf, repeat. The walking was the job. In a big warehouse with a long SKU tail (Ariat has somewhere north of 3,000 actively stocked items across boot sizes, widths, and colorways) the walking became most of the job. Forty thousand steps a day is roughly 19 miles. That is a professional dog walker’s mileage, performed by a warehouse worker, in a building with concrete floors.

Two things happened at once. The first was that Ariat partnered with Infios (a warehouse-management-system vendor that had recently integrated a tight handoff with Geekplus, a robotics company that makes autonomous mobile robots — the squat, wheeled kind that drive themselves around a warehouse floor carrying shelves). The second was that the Geekplus AMRs do not just follow fixed paths. They use a learning layer — call it AI, call it reinforcement learning with a route planner, call it what you want — that optimizes travel patterns continuously based on where the orders are, where the workers are, and where the congestion is. In 2023 this was a novelty. By mid-2025 it was mature enough that a mid-market operator could install it without betting the warehouse on a three-year vendor engagement.

The specific operational change: the worker no longer walks to the product. The robot brings the product to the worker. The worker stands at a pick station, and the robots deliver shelves in a sequence optimized by the AI layer. The old 19-mile daily walk becomes a roughly 3.3-mile walk — still more than most desk workers, but within the range of a shift a human body can sustain for 30 years instead of six.

That’s the visible change. The invisible change is the one I want you to notice. The planning layer — the thing that decides which shelf goes to which station in which sequence — is an AI system that learns continuously from order velocity, worker throughput, and SKU correlation. If you asked the person running the warehouse floor “is there AI in your building,” they would probably say “there’s some algorithm in the robots.” If you asked the vendor, they would say “our agentic orchestration layer continuously optimizes…” — the word “agentic” doing most of the work. These are the same system. What changes is whose vocabulary you’re trapped inside.

The results

I will not claim to have a complete audited set of Ariat’s internal metrics. What is public and verified is enough to be interesting.

• Steps per worker per day: 40,000 → 7,000 (82.5% reduction).
• Physical injury rates: publicly described as reduced, though Ariat has not released specific numbers.
• Order volume through the same facility: grew while headcount was held roughly flat.
• The retention story: Ariat has been publicly on record — Cross herself, in interviews — about the automation project being driven by an inability to find and retain workers in a tight Midwest labor market, not by a cost-cutting mandate. This is the part that should make you sit up.

That last point is the interesting one. The “we did this because we couldn’t find workers” frame is the opposite of the “we did this to replace workers” frame that every AI-replacement piece in the consumer tech press trades in. It’s also much more honest about what’s actually happening in mid-market manufacturing, logistics, and distribution right now. The labor market for physical work at a $100M–$500M operator is not tight because of AI; it was tight before AI showed up, and AI is the first tool in 40 years that has let operators like Beth Cross solve a problem that wasn’t solvable with another training program, another shift differential, or another pallet-jack redesign.

The Craft of AI read

Three things to steal from this, and two warnings about what not to.

Steal #1: Pick one worker, measure their hardest day. Cross’s method is not “where can we apply AI.” It’s “what is the most physically punishing 30 minutes in our operation, and who is doing it, and what would it take to remove it.” That is a falsifiable, observable, quantifiable question. It produces a target (“reduce the walking”) that can be measured by a step counter, not by a McKinsey deck. Every founder-CEO reading this can do this exercise Monday morning without hiring anyone. Walk your most physical operation. Find the worker whose shift is hardest. Ask them what they would remove if they could remove one thing. That conversation is your AI roadmap.

Steal #2: Let the vocabulary come last. The Ariat deployment is, technically, an “agentic AI” deployment. It fits the definition: the system makes autonomous decisions, executes them, and learns from the results. But Cross never needed to hold a board meeting titled “Agentic AI: Our Path Forward” to get it approved. She needed one titled “Warehouse Retention.” The rule: when you are the operator, you choose the problem; the vendor chooses the vocabulary to describe the solution. Never let those get reversed, because once the vocabulary comes first, the problem becomes “we need an AI strategy” — and that problem has no solution, only more vocabulary.

Steal #3: Install the capability at the layer closest to the physical work. The AI in the Ariat warehouse is inside the robot fleet’s planning layer. It is not running on a laptop in finance. It is not orchestrating a chatbot on the website. It is inside the physical system where the value is created and where the cost lives. This is the single biggest mistake I see founder-CEOs make in first deployments: they put the AI somewhere that’s cheap to install and easy to demo (marketing copy, customer service, sales email) instead of somewhere that changes what a human body has to do on a Tuesday. If your first AI deployment does not change the shape of somebody’s workday in a way their spouse notices at dinner, it is marketing. Install it closer to the work.

Warning #1: This is not cheap. Ariat’s warehouse automation project is, by conservative estimate, a multimillion-dollar capital commitment — robots, WMS integration, racking reconfiguration, worker retraining. A $100M operator cannot write that check in one quarter, and shouldn’t try to. What a $100M operator can do is run a one-zone pilot in the most punishing 10% of their operation, measured by the same step-count discipline. The ROI case at small scale is often stronger than at full scale, because the baseline inefficiency in a manual zone is usually worst at the worst-performing zone.

Warning #2: The robotics vendor will try to sell you the planning layer as “their AI.” It’s not really. The planning layer is the thing that creates the step-reduction value. It is also the thing that becomes your dependency for the next fifteen years. Negotiate it like a core system, not like a feature. Ask explicitly who owns the data the system learns from, how the model is updated, and what happens if you want to switch robot vendors in 2031. These are normal questions for an ERP. They are going to feel rude to ask a robotics vendor in 2026. Ask them anyway.

Things to consider

1. Walk your hardest operation this week. Not your warehouse — your hardest operation. For you that might be a plant floor, a field service route, a customer onboarding handoff, a returns desk. Find the worker whose shift is the hardest and ask them, on the floor, what the worst 30 minutes of their day is. Do not bring a notebook. Do not call it an AI project. Call it a conversation.
2. Translate the answer into a measurable physical unit. Steps. Lifts. Minutes on hold. Boxes reopened. The metric has to be something a worker feels in their body or their gut at the end of a shift. If it doesn’t, the problem isn’t real enough to deserve a solution.
3. Ask your current systems vendor whether their roadmap has the thing you need within 18 months. If the answer is yes, you do not need to buy a new vendor. If the answer is no, you need to know now, because the replacement cycle is the most expensive thing you will do in the decade. Do not let a greenfield AI project be your path out of a legacy WMS, ERP, or CRM — unless you are ready for the full replacement cost.
4. Do not write a strategy document. Write a problem statement. One paragraph. The problem is that worker X is walking Y miles a day to move Z pounds of product, and the retention math does not work anymore. That paragraph, on one page, is your entire AI strategy. Everything else is implementation.
5. Beth Cross has been doing this for 36 years and has not been called an AI visionary once. That should be a comfort, not an anxiety. The operators who get quoted in AI strategy pieces are almost never the ones who got the biggest step-count wins. Choose which group you want to be in.

THE WORKBENCH

Here’s the tactical takeaway for your Monday.

Run the step-count test on your own operation this week. Literally. Pick one worker in your most physical role — a distribution picker, a line technician, a service tech, a field installer — and ask them to wear a phone or a fitness tracker through a full shift. You are looking for two numbers:

1. Steps per shift (or the nearest physical-work proxy for your operation — lifts, stops, handoffs, mouse clicks, whatever).
2. The worker’s own estimate of how much of that is “necessary” versus “because the system is stupid.”

That second number is the one the robots can take. The first number is the one you report to the board.

Then do one more thing: look at your most recent three vendor pitches that contained the word “agentic.” Cross out the word every time it appears, and reread the pitch. Ask yourself: is there still a problem being solved, or was the problem “we don’t have agentic AI yet”? If the pitch collapses without the word, the pitch was always about the word. Move on.

If the pitch still stands up, send it to me. I want to see the ones that do.

THE QUESTION

This is the one I am honestly curious about, and I’d love your reply.

What is the worst 30 minutes of the hardest job in your company, and when is the last time you personally watched someone do it?

Not “heard about it in a skip-level.” Not “read about it in an employee survey.” Personally stood there. If it’s been more than a year, that is fine — I am not trying to catch you out, I am trying to figure out how many founder-CEOs in this audience are still close enough to the physical work of their companies to do what Beth Cross does. I suspect more than the tech press thinks, and fewer than we’d all like. The answer, I think, is going to tell us a lot about which $100M–$500M operators are going to get real AI wins in the next 18 months and which are going to get PowerPoint wins.

Hit reply, or send a note to grant@thecraftofai.com. One sentence is fine. I read every reply, and the best ones will shape the next issue.

— Grant grant@thecraftofai.com