Pedro Arnt measures the progress of artificial intelligence at his company with a figure that rarely leads an investor presentation: how many fewer people are needed to do the work.
He arrived at that measure after reviewing the initial results. He was shown reports with enormous numbers of automated hours. Then he would ask how much costs had fallen. The reduction was far smaller than those figures suggested.
There was an explanation: they were automating many things they had not even been doing before.
He changed the way they measured progress. Today, every department has headcount reduction targets. Arnt believes it is a figure that is difficult to disguise: if the work stops getting done, it shows. And if automation truly works, it should allow the work to be done with fewer people.
He also looks at what percentage of pull requests, proposed changes to the code, move forward without human intervention. But he suspects that indicator will soon tell him little. Almost all teams are reaching very high percentages.
Arnt is CEO of dLocal, the Uruguayan company that exceeded one billion dollars in revenue for the first time in 2025. It ended the year with revenue of $1.093 billion, up 47% from 2024. It processed around $40.8 billion in payments, up 60%, and earned net income of $196.9 million.
He arrived in 2023, after twenty-five years at Mercado Libre. For the last twelve, he had been its chief financial officer.
A Lobotomy
When discussing his arrival at dLocal, he dwells on how they handled the leadership transition.
The company was founded in Uruguay in 2016 to offer local payment methods to global companies. Five years later, it was already listed on Nasdaq, a trajectory Arnt considers unusual.
Arnt joined as co-CEO alongside Sebastián Kanovich, who had led the company since its founding. They shared the position until March 2024. Arnt then took charge, and Kanovich moved to the board.
He recommends that founders who believe they will eventually want to step down prepare someone inside the company. To explain the risk of bringing in an outside successor, he uses a forceful comparison: a lobotomy. You cannot remove one brain, put in another and expect the body to keep functioning in the same way.
In his case, sharing the role gave them time to get to know each other through work. Arnt could observe how Kanovich made decisions. Kanovich could do the same with him and point out what he thought needed correcting.
That period can also reveal that the relationship will not work. If that happens, he says, it is better to rip off the bandage early.
The hardest part was changing certain habits at a company that had done very well.
Arnt explains that, in its early years, a company like dLocal needs to grow quickly. Size matters for the business to work. But that same growth brings obligations: regulators begin looking more closely, and its operations take on an importance they did not have before.
Some things can no longer be done in the same way. The problem is explaining that when there has been no crisis and the team has good reasons to be proud of what it has achieved.
He wanted to preserve attention to merchants, speed and the ability to manage risk. dLocal moves money in countries where it is not always possible to work as one would in Switzerland, he says. Nor can it afford problems with the regulator.
Going public brought new obligations. Oversight also increased: more than thirty authorities supervise its activities.
He wants dLocal to preserve its speed without losing the trust that licenses and oversight provide. When speaking with a Fortune 500 company, part of his argument is that it will be entrusting its financial infrastructure to a regulated institution. If he can also resolve things faster than a traditional institution, he believes he has a significant advantage. He calls it “escape velocity.”
A Horizontal Layer
dLocal generates more than a billion dollars in revenue doing work that most people making payments never see.
A company like Google, Netflix, Spotify or Amazon wants to grow in a country and discovers that collecting payments is not as straightforward as expected. In the United States or Europe, it was accustomed to accepting credit cards. In Uruguay, it has to accommodate Oca or local transfers; in Brazil, Pix or Mercado Pago.
Solving that independently in every market requires an investment that is difficult to justify.
dLocal already has those connections in more than sixty markets and works with more than a thousand payment methods. Its clients complete a single integration and need to add little code. Uruguay accounts for less than 1% of revenue.
Arnt believes there is a lesson here for other companies. The payments companies he admires in Silicon Valley typically build and control all their technology, intending to offer something better than established players.
In emerging markets, they encounter a different difficulty: there are many systems, providers and ways to pay. Building all of them is not viable.
That is why dLocal connects what already exists. The global merchant works with a single platform, and the company handles the differences between the systems behind it.
Before entering a country, they send a team to speak with regulators, banks and fintechs. They look for partners with whom they can integrate quickly and address adjustments as they arise.
Sometimes they decide to wait. If the only alternative is a century-old bank that takes six months to make each change, Arnt understands that the time has not yet come.
Before launching a merchant, they may also conduct a premortem: reviewing what could go wrong before it happens. Experience helps. They have already made mistakes in twenty markets, he says, and know many of the problems they might encounter.
Two Questions
Arnt remembers 2025 as the year they decided to commit fully to artificial intelligence. He saw two possibilities: transform the company or fall behind a company born with that technology.
He did not believe they could experiment a little and wait comfortably.
Nor was it a year of major immediate results. They learned a great deal and began seeing benefits when they reviewed what they had done.
To decide where to allocate resources, they asked themselves two questions. The first was whether AI could open a revenue stream or a new business. In payments, they had not yet found an obvious large, near-term opportunity.
The second was whether it could change costs, margins and productivity. There, they had no doubts.
They devoted almost all their effort to reviewing how the organization worked. They left a few people monitoring new business models so they could alert them if something appeared that they should not overlook.
Arnt acknowledges the company’s limits. They are not Stripe, nor are they at the frontier of research. He does not think trying to occupy that position would go well for them either.
Instead, he sees much to do within dLocal. Financial services require reconciliations, audits and controls, some because of regulation and others because they are moving other people’s money. There are thousands of tasks surrounding every transaction.
If they were starting from scratch today, he believes, they would have a smaller company, with better margins and fewer errors.
The Mistake in the First Attempt
The first time, they asked teams to help transform their own work.
They were clear about what might happen: some functions could change so much that a position would cease to exist as they knew it. They also said they would recognize that effort and that learning to work with these tools would be useful afterward.
Before automating, they needed to know what each person did and how they did it. They asked people to document their processes, something they had not previously done in that way.
They learned a lot but made little progress.
Arnt admits the request was unfair. People had to maintain their everyday work and also describe, step by step, tasks that could end up being automated. In some cases, they were helping eliminate their own positions.
An engineer proposed another solution: instead of asking teams to conduct that entire review, they could build an agent to do it.
They created agents that reviewed emails, tickets and available information. With that, they could describe each position: its tasks, how it handled them and whom it needed to work with.
They then ranked positions according to how much could be automated. That information fed into the platform they used to develop the agents responsible for doing the work.
Arnt describes it as an internal consultant. Later, he saw what consultancies were offering in this field and recognized the same idea. He wonders why it had not occurred to them earlier.
In some areas, he says, they managed to have agents do the work of 30% of positions and reassigned those people.
His way of thinking about processes also changed. He says many exist to reduce differences between how one person works and another. They organize tasks and make outcomes more predictable, although they sometimes limit someone who could do better.
An agent, he argues, does not need to follow exactly that route. It can look for another way to achieve the requested result.
The Leverage Program
In programming teams, the change led them to seek more experience.
Arnt observes that agents still struggle with more abstract decisions: designing a system well or defining its architecture. By contrast, they perform many of the tasks previously assigned to junior engineers very well.
That is changing the composition of teams. They have a program called leverage, through which they are replacing less-experienced positions with more senior ones.
He imagines a smaller and considerably more productive team, although not necessarily a much cheaper one. Fewer engineers would be needed, but with greater ability to make the decisions the tools still do not handle well.
He himself points out the problem that logic creates: if nobody hires juniors or gives them work through which to learn, where will the seniors of the future come from?
He jokes that he is already old enough to leave the question to someone else. He immediately clarifies that it is a real problem.
To get people to use the tools, they were explicit about their expectations. They also did things Arnt describes as somewhat unpopular.
Every month or two, he sent a ranking of the departments consuming the most and fewest tokens, alongside a list of the twenty and one hundred people using them most. He wanted to show that they were monitoring adoption and generate some competition.
The engineers established the necessary controls to provide access to information and manage costs. People could begin experimenting without obtaining permission becoming a constant obstacle.
Arnt recounts a scene he finds amusing: someone stands up on an office floor and asks who can lend them tokens.
He takes one distinction from that experience. For an individual to learn, they need tools and room to experiment. To change how the entire company works, leadership must set objectives, demand results and track progress.
What AI Does Not Disintermediate
When asked whether all this makes dLocal a better business, Arnt considers which parts of the company could lose value and which remain necessary.
They have no physical assets. Everything is bits and bytes, and he believes artificial intelligence will profoundly change that work.
But clients also pay them for something that cannot be resolved through software alone: being present in each market. dLocal deals with processors and local partners, answers to regulators and meets fiscal, tax and legal obligations in sixty countries.
He does not expect AI to replace all of that soon.
For that reason, he feels more comfortable serving emerging markets for clients from developed countries. If he sold financial services in those clients’ own markets, he would compete with companies that already handle many of those tasks. He worries that differentiating one product from another will become increasingly difficult there.
He is less certain about the barriers protecting a business. Using AI well provides an initial advantage, and many companies will fall behind because they fail to do so. But he believes that between companies, where purchasing decisions are made by comparing costs and results, technology also makes it easier for new competitors to emerge.
Personal relationships still count. Someone decides to hire the company, and someone calls when there is a problem. That part of the work still matters.
Even so, what worries him most is not a fintech he already knows. It is one born with artificial intelligence that organizes its entire business around it from day one.
