Oilfield AI Cut 35% of Trucks. Zero Layoffs

THE CRAFT

Liberty Energy now delivers the same sand on 35% fewer trucks than it used to. Every frac job runs on that sand, and Liberty, a Denver company that fracks oil and gas wells across North America, moves more than 20 million tons of it a year. That is more than a million truckloads, and every one of them has to leave the right mine and reach the right well pad at the right hour, because a frac fleet that runs out of sand stops, and a stopped fleet is the most expensive thing in the business.

For years, people handled that by hand. Ron Gusek, Liberty’s chief executive, described the old way to Fortune last year: it used to be somebody’s job on location to control the cadence of trucks, and the safe habit was to line up eight or ten of them waiting to unload. Drivers sat for about an hour when they could have been driving. The people doing it were being careful. Waiting trucks were what it cost to make sure the pumps never stopped.

That was the first job Liberty gave to AI. Before it tried to automate the frac job itself, it built a system that forecasts how much sand each well site will need up to a day ahead and schedules every truck against that forecast. In August, Gusek told a room of investors in Denver what came of it: “We reduced the number of trucks that we required to deliver all of that sand by 35% with the deployment of AI.” A year earlier, he had put the number at 30%. In the same talk he added, “I haven’t laid off a single person because of AI.”

The ambitious project, AI that runs the pumps on a frac job by itself, is still being tested on about a third of Liberty’s fleets, later than Gusek once expected. The plain one, the truck schedule, has been paying for over a year.

This week’s brief is about that order of events. Liberty is a $4 billion company in a hard business, and its AI program is broad: the trucks, the repair shop, the pumps, the fuel. But the clearest lesson for a leader running a smaller company is where it started. It started at a desk most people outside the company would never think about, on a decision that gets made thousands of times a day.

So I want to start you on a question, and come back to it at the end: what is your company paying, every day, to protect itself from a guess?

THE OPERATOR

The situation

Liberty was founded in 2011 by Chris Wright, who left in early 2025 to become US Secretary of Energy, and Jim Brady, its first employee, who is now its senior vice president of operations. It grew from a tent in North Dakota into one of the largest frac companies in North America, with about 6,000 people today. Gusek became chief executive in 2025, when Wright left for Washington.

A frac company sells a service that is hard to tell apart from its rivals’. An oil producer hires a fleet of pumps, engines, and crews to force water and sand into a well at high pressure so the oil and gas can flow. The producer cares about three things: that the job is done right, that it is done fast, and that it never stops. Liberty’s whole operation is built around the third one.

The industry has also been shrinking. Gusek pointed out that when Liberty started there were about 450 frac crews working in North America. Today there are about 200, and the continent is producing record amounts of oil and gas with them. Each fleet does far more work than it used to. That is good for the producers, and it means a frac company only keeps growing by being better at the work than the fleet down the road.

Liberty’s 2025 showed how hard that is. Prices for frac work fell, and revenue dropped from $4.3 billion to $4.0 billion. Adjusted earnings before interest, taxes, depreciation, and amortization fell further, from $922 million to $634 million, a drop of about 31%. In a business like that, the costs a company controls decide who comes through a downturn in good shape.

Two of those costs are easy to underrate from the outside. The first is logistics: sand, fuel, water, and the trucks that carry them. The second is wear. Frac equipment runs in brutal conditions, and the engines, transmissions, and pumps wear out on a schedule nobody fully controls. When one fails in the middle of a job, the crew waits, the customer waits, and the repair costs more than it would have if someone had seen it coming.

Both of those problems were being managed by skilled people working with the information they could see. Gusek is respectful about that. “We did that with people, of course, with a lot of people and a lot of mind share dedicated to making sure that those trucks were flowing as optimally as possible,” he said in August. “The people were very, very good at that.”

The move

Liberty built its AI tools itself. Its annual report says so plainly: the company designs and builds each piece of its software in-house. That matters for this story, because it means the order in which things were built reflects Liberty’s own judgment about where the payoff was, and nobody else’s sales plan.

The first piece was the truck schedule. Liberty calls the system Sentinel. It pulls together what a good dispatcher would want to know if they could hold it all at once: how much sand is sitting at each well site, how fast each fleet is using it, where every truck is, how long trucks are waiting at each loading facility, the traffic, and the weather. From that it forecasts each site’s needs up to 24 hours ahead and schedules the trucks to match. The job used to be an educated guess made under pressure on location. Now it is a forecast, made the same way every time, with the whole picture in view.

The result Gusek reports is fewer trucks for the same sand: 30% fewer by September 2025, and 35% fewer by August 2026. He made a point about who else benefits. Most sand trucks belong to independent owners whose costs are mostly fixed: an expensive truck, a driver, insurance. Fuel is the main thing that goes up with each trip. “If they get to make three trips a day instead of two trips a day and they get paid for one extra trip, there is a meaningful amount of fall through in that,” he said. “So we make a trucker’s life better by doing that.” Liberty has since put the same scheduling system to work in a second business, its compressed natural gas delivery division.

The second piece was the repair shop. Liberty streams about 2 billion data points a day from its engines, transmissions, and pumps into a monitoring system it calls FracPulse. The models behind it were trained, in Gusek’s words, “with every problem we have ever had on an engine or transmission or power end.” They look for the small patterns that come before a breakdown. A small team near Denver, in a room Liberty calls The Hive, reviews what the system flags and calls the crews in the field before the problem becomes a failure. The people in it don’t run the equipment. They sit between the data and the field, deciding what is worth acting on.

The third piece runs the pumps themselves. Liberty’s pump-control software automates how fast and how hard the pumps push during a job, and a newer system Liberty calls Forge learns from data across the whole fleet to cut fuel use. On the July earnings call, Gusek said Forge has started finding improvements to how fleets are designed and used, beyond the fuel savings it was built for. Gusek has also been careful about what these tools are. When an earlier company announcement described Forge as a large language model, he corrected it on the next earnings call: “We mistakenly called it a large language model in our press release.” Few chief executives correct their own press release in front of analysts. I read it as a good sign about how the rest of the program is run.

The fourth piece is the one that gets the headlines, and it is not finished. Liberty is building AI that would run an entire frac job: the pumps, the blender, the pace, with people supervising. In September 2025, Gusek told Fortune that “we are rapidly getting to full deployment” and that “I expect by the end of this year we’ll be there.” In August 2026 he told investors the system was “on test now on probably a third of our fleets.”

Most of the credit belongs to people whose names are not in any of this: the dispatchers whose habit of lining up trucks showed exactly what the forecast had to replace, the in-house engineers who built it, and the small team in The Hive who decide every day which warnings matter.

Put those four in a line and the sequence says a lot. The first thing Liberty pointed AI at was a scheduling decision that happens thousands of times a day. The second was keeping equipment running. The flagship came last, and it is the one running late.

The result

The truck numbers are the cleanest result in the story. Fewer trucks means less waiting, fewer trucks on rural roads, less fuel burned, and a supply of sand the frac fleets can count on. Gusek has given the number in two places a year apart, first 30% and then 35%, and both times he credited the AI directly. His later remarks point in the same direction: the improvement has held up and kept growing as the system learned.

The maintenance results are broader, and Liberty is careful about how much credit it gives the AI. In its 2025 annual report, the company says that “in part through the implementation of FracPulse” it has seen 23% fewer transmission failures over two years and 61% longer life for its power ends, the heavy pump components that take the most punishment. In April 2025, Gusek reported engine life up 27% over three years and fluid-end life up 40%, again “in part” because of AI-driven maintenance. By August 2026, his own description was that component life had climbed “by as much as 100% since we have deployed an AI model that oversees that maintenance.”

The money shows up in one number. For 2025, Liberty says total maintenance cost for each unit of work fell about 14%. Gusek credits that to two things together: the AI maintenance software and the company’s move to newer natural-gas fleets. Nobody can separate the two cleanly, and Liberty does not try. What can be said is that the cost fell during the year revenue fell 7% and pricing went against the company, which is exactly when a cost you control matters most.

Then there are the people. Gusek’s line about layoffs is worth reading next to something he told Fortune a year earlier: “The assets are getting larger and more automated. That means we don’t need as many people out there.” Both can be true at once. Each fleet needs fewer hands as the equipment gets bigger and smarter, and the people Liberty already has are doing different work. “Yes, their job looks a little different today than it did two years ago,” Gusek said in August, “but they are more productive, and they are working on different problems of a different scale now.” The Hive is an example: a new job, created by the AI, held by people whose work now is deciding which warnings matter.

Here is the hard part of the story.

The AI did not save the year. None of this has made Liberty’s margins safe. In the second quarter of 2026, revenue grew to $1.2 billion, but the cost of delivering the work rose faster: about 82.5% of revenue went to the cost of services, against 77.9% a year earlier. The AI did not stop prices from falling, and it did not stop costs from rising. My read is that it softened a hard year, and the business would look worse without it. Nobody, Liberty included, has put a number on how much. Meanwhile most of the questions on Liberty’s earnings calls are now about its newer business, building power plants for data centers.

The other honest point is about the flagship. A CEO who says the AI-run frac job will be fully deployed by year end, and a year later says it is being tested on a third of the fleets, has learned something. The big system is harder than it looked. The modest one, the truck schedule, worked on the timeline it was given.

The Craft of AI read

A lot of AI work starts with something to buy: a planning agent, a maintenance copilot, a platform with a dashboard full of predictions. It usually arrives before anyone has watched how the work actually happens at the level of the person on the radio telling the next truck where to go, on the hope that the understanding will catch up after the software is installed. It is an easy order to fall into, because the software is the part you can see.

Liberty did it the other way around. Before there was a forecasting model, there was a company that knew exactly why eight or ten trucks were lined up at a well: because waiting trucks were the only insurance anyone had against a stopped fleet. A line of trucks idling at a well is the kind of tell anyone can spot on a walk-through, if they know to look. That is a specific understanding of a specific habit, and it told Liberty precisely what the AI had to replace. What it needed was a forecast good enough that people could stop paying for insurance in the form of idle trucks. When the forecast got good enough, the insurance could shrink, and so did the number of trucks.

I think there are two lessons in that for a leader running a smaller company, and they pull in different directions on purpose.

The first is about where to start. Liberty’s best AI result came from the most repetitive decision in the business, the one made thousands of times a day by people doing their honest best with partial information. Most companies have one of those. It is rarely the thing anyone wants to present to the board. The second is about patience with the big idea. Liberty kept building toward an AI-run frac job, and it is still building. It did not have to choose between the plain project and the ambitious one. It let the plain one pay while the ambitious one took the time it needed. I don’t know whether the AI-run frac job will work as promised, and the company’s own timeline has already moved once. What the company has shown is how to keep going while it finds out.

This connects to the twenty minutes PITT OHIO got back per driver, per day. Both companies found their money in the gap between a truck’s time and a schedule’s accuracy. Neither found it by buying something with “AI” in its name and hoping the schedule would sort itself out.

Things to consider

• Find your waiting trucks. Every business keeps a buffer against uncertainty: extra inventory, padded schedules, a second person who double-checks, trucks lined up at a well. Each one is a sensible price for not knowing what will happen next. Make a list of yours and put a rough yearly cost next to each. The most expensive buffer that exists only because of a forecast nobody trusts is a strong candidate for your first AI project.

• Start with the decision you make the most, not the one you’re proudest of. Liberty’s flagship AI project is the one that would run a frac job. Its first proven result came from scheduling trucks. Frequency is what made the difference: a decision made thousands of times a day gives a model a lot to learn from, and a small improvement on each one adds up quickly. Ask which decision in your business gets made more often than any other, and how much of it is still a guess. If your team rushes orders through constantly because nobody can see what’s coming, you have probably found it.

• Build the job that sits between the data and the people. The Hive is a small team whose work is deciding which of the system’s warnings deserve a phone call to the field. That job did not exist before. Most companies that add monitoring end up with alarms nobody trusts, because nobody owns the step between the alert and the action. Before you buy a system that flags problems, decide who will read the flags and what they are allowed to do about them.

• Count the work that disappears, then count the people who stay. Liberty can point to fewer trucks and longer-lasting parts, and its CEO can say nobody was laid off because of AI. Gusek says the truckers came out ahead too, with more paid trips in a day. Both claims are about real things a reader can check. When you make the case for an AI project inside your company, describe it in the same terms: what work goes away, what the people doing it will do instead, and what the business pays for that work today.

• Keep your public promises smaller than your private ambitions. Liberty said its AI-run frac job would be fully deployed by the end of 2025, and it was not. The company kept going, and the claims it makes now are more careful. Your own team will hear your AI timeline the way investors heard Gusek’s. Promise the result you can already see, and let the larger goal prove itself.

THE WORKBENCH

Do this tomorrow

One hour, a whiteboard, and two people: the person who runs your most frequent scheduling or allocation decision, and the person who pays for what happens when it goes wrong.

Name your buffers. Ask the first person a simple question: where do we keep extra capacity just in case? It might be people on standby, inventory above what you sell, a delivery window padded by a day, or a second check on every order. Write each buffer on the board, with what it protects against.

Find the guess behind each one. For every buffer, ask what forecast would let you shrink it, and who makes that forecast today. If the honest answer is “nobody, we just keep extra,” circle it. Those are the places where a better forecast pays first.

Price your most expensive buffer. Pick the circled buffer that happens most often. Using this year’s real records, add up what it costs: the idle hours, the carrying cost of the extra stock, the overtime, the trucks or vans or people paid to wait. Write the yearly total on the board.

That number is operating margin you are spending right now to make up for a forecast you don’t have. It is also the most concrete brief you can give anyone who wants to sell you AI: here is what we pay for not knowing, and here is the decision that would have to get better.

THE QUESTION

Ron Gusek’s company is building AI that could someday run a frac job on its own. The result he talks about first is still the trucks: the same sand, a third fewer trucks, and drivers getting in three trips where they used to get two.

So, back to the question from the top. What is your company paying, every day, to protect itself from a guess?

If you can name it, you know where to start. It probably isn’t glamorous, and it probably isn’t what your board expects to hear about. It is likely where the operating margin is hiding, and it is where Liberty found its first proof.

If you can’t name it yet, that is useful too. It means the buffers in your business have been there so long they look like the way things are done. Spend the hour in the Workbench and look for the trucks lined up at your well.

---

Want this done for you?

You just named the buffer your business pays for because a forecast is missing, and put a yearly price on it.

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