Organizing human intelligence to power the AI economy

In August 2023, Brendan Foody, Adarsh Hiremath, and Surya Midha found themselves in an unlikely situation for college sophomores.

Just months after founding their company, Mercor, from their college dorm rooms, they were at Tesla’s offices in the San Francisco Bay Area, sitting across from the entire co-founding team of xAI (sans Elon Musk).

Mercor, Latin for “to trade” or “to buy,” contracted out highly-skilled software engineers located in India on a per-hour basis. They’d developed the idea at a three-week hackathon in São Paulo, Brazil. Brendan, Adarsh, and Surya knew there was an abundance of talented engineers at Indian Institutes of Technology who struggled to find job opportunities, even though their skills were in high demand with American tech companies. They automated the processes for reviewing resumes, conducting interviews, and making hiring decisions. They started contracting engineers to friends at Harvard for projects and then hustled to scale it up.

Just two days before the meeting, a Mercor customer introduced the co-founders to the xAI team over Zoom so the cofounders could explain how they’d amassed a bench of Indian software engineers who excelled at math and coding. xAI wanted to know more.

So there they were, second-year college students meeting with the founders of a frontier lab, who were captivated by the caliber of hires they’d accessed.

“This can’t be normal,” Brendan recalls thinking. “We just felt like there must be such an anomaly that these people are excited to talk to us when we’re not qualified in the same way as the people they would normally spend time with. That’s when we knew there was something in the market we had to chase down.”

xAI was still pre-training its model and not ready to introduce human data, but the conversation made them realize that a radical change was coming. Models would need to be trained on real human expertise. They could be at the forefront of how that happened.

“Almost immediately, we realized that what the human data market needed was a talent assessment offering,” says Adarsh. “And because we had built up this network of really, really high-caliber, high-end knowledge workers and developed the ability to match and onboard them in an automated way, it ended up being a great fit to actually provide that offering to the foundation models.”

Hunting opportunities – the bigger, the better

Sitting in a room surrounded by elite technology leaders. Chasing a massive market that had yet to actually materialize. Many potential founders – especially those as young as Brendan and Adarsh – would be intimidated by these experiences. But Brendan and Adarsh have always been unafraid to chase the biggest possible opportunity.

The two met freshman year in rhetoric class at the all-boys high school they attended along with Mercor’s other founder, Surya Midha. All three bonded on the debate team. Brendan, a natural talker, enjoyed the verbal parts of debate but struggled with the heavy reading load because of his dyslexia. Adarsh, conversely, was pushed into debate in middle school because his parents worried he was too quiet. He and Surya formed a historically formidable duo, becoming the first team to ever win all three national tournaments in policy debate. 

Even from a young age, Brendan was always on the hunt for business opportunities, whether it was selling Safeway donuts that cost $5 a dozen for $2 a pop, or buying up whatever was going the fastest at a middle school dance and doubling the price. “I called them my little startups,” says Brendan. “My parents called them my schemes.”

The startups grew as Brendan did. During Covid, he launched a company that built software interfaces for startups so they could better interact with AWS servers, and Adarsh joined the fray.

“I got drawn to startups in part because Brendan roped me in,” says Adarsh.

The pair, who now serve as Mercor’s co-CEOs, complement each other and sometimes take different paths to get to the same destination. Brendan, the enthusiastic ideas man, is customer- and ROI-obsessed. He effectively dropped out of Georgetown by missing his finals because he was too busy with Mercor. Adarsh, who thought his future would be in research, has found a natural home leading engineering and product, and building the company’s new Enterprise (opens in new tab) business. He weighed the situation more carefully and opted to take a leave of absence from Harvard.

Both founders excel at recognizing a problem and then dissecting it to find the real issues and opportunities. First principles are second nature to them. Each opportunity they’ve pursued has revealed a new, bigger one that demands their attention. Indian engineers were underemployed because labor markets are wildly inefficient at identifying and matching skills. Why not go after that opportunity?

“Brendan and Adarsh have the remarkable ability to stay focused and disciplined, while also spotting the changes that have not yet come into view,” says Felicis managing partner Sundeep Peechu. “They’ve already proven to be visionary and skillful operators.”

Capturing the expertise in everything

After the xAI meeting brought their attention to the human data opportunity, Brendan and Adarsh spent months working on a new version of their product that could manage each step of the hiring process at a faster pace and greater volume.

Most human-involved AI model training to that point was low-skilled work: labeling and annotating data, and giving some basic feedback on output. Such work remains a big part of the data market, and is essentially a volume play, both in terms of data and human reviewers. Even though this type of work has been critical to get LLMs to where they are today, the returns have been diminishing for some time as AI’s capabilities become more sophisticated.

By bringing in top-flight engineers at scale, Brendan and Adarsh were introducing a new, valuable element to the training process. Initially, they struck a deal with data-labeling pioneer Scale AI to hire more than 1,000 engineers for work with frontier labs. “We thought it would be more efficient to partner with a company that works directly with the labs,” says Brendan.

But the engineers Mercor hired started complaining about missed pay and a poorly managed platform. Brendan and Adarsh decided to cut out the middleman. Within a couple of months, they started working with Cognition and a leading frontier lab. High quality training data was an urgent bottleneck for every major lab, and Mercor quickly counted them all as customers. To meet demand, they refined the product so it could autonomously process hundreds of thousands of resumes, conduct just as many interviews, and accurately decide who to hire.

“There were things that just had to mature very, very quickly,” says Adarsh. “Building a cloud network where you could source, vet, screen, and onboard these people in an automated way is just incredibly challenging to do.”

Having that robust infrastructure in place positioned Mercor as a critical part of the AI ecosystem as labs shifted training to reinforcement learning environments. Whereas consumer models trained on information ingested from free, readily available sources, AI agents need to be trained on real human expertise to develop more valuable capabilities.

“We don't have these giant corpuses of knowledge online that answer exactly what the model should be doing in areas like law, banking, and consulting,” Brendan explains. “And so that requires an enormous amount of humans putting in the work to create the right context to build the reinforcement learning environments needed to train agents across all these different verticals.”

Using its network of over 5 million vetted experts, Mercor generates that context through evals (opens in new tab): detailed rubrics to gauge how models do at tasks compared to professionals in those fields. If the model is the product, the eval is the product requirement document.

“We view evals as a foundational infrastructure investment for all AI deployment and development more generally, because they answer the questions of how well the model is actually doing, and why it’s doing well, or why it’s doing poorly,” says Adarsh.

Getting an eval dialed in just right is what ultimately makes a model successful at a given task.

“Reinforcement learning is becoming so effective that once we can build an evaluation for something, the models can saturate it and learn that capability,” says Brendan. “So that means that the bottleneck to applying agents to every workflow in the economy is, how do we build evals for everything? How do we mobilize hundreds of millions of people to enable us to do that?”

In July 2026, Mercor moved on to the other half of that problem. The company announced it would acquire Deeptune (opens in new tab), a startup that has spent two years recreating hundreds of enterprise applications, from spreadsheets to Salesforce, as simulated environments where AI agents can practice before they touch a live system. Mercor was already a Deeptune customer, and Brendan had backed the company as an angel investor before Mercor acquired it.

A training environment has to do three things at once: imitate the software, set the agent a real task, and judge how it did. Mercor's expert network already handles the second and the third. Deeptune handles the first, and now it handles it exclusively for Mercor.

More important than Mercor’s sheer number of contributors is their quality. Brendan says model improvement is driven by the power law, where the top 10 to 20% of people drive the majority of advancements. Mercor’s job is to find those high achievers and keep them on the platform. The company pays out more than $4 million a day to experts that range from math Olympians and chess masters to shopping experts, chefs, and radiologists.

Building the future of work and fixing labor markets

Chasing increasingly bigger opportunities has driven Brendan, Adarsh, and Mercor to where they are today. They’ve surged past the 10x moment others dream about. In September 2025, Mercor became the fastest startup to ever reach a $500 million revenue run rate. In March 2026, they reached $1 billion. And by June, Mercor said it had reached $2 billion in annualized revenue run-rate. It has all happened so fast that it almost seemed easy. In reality, it has been anything but.

In March 2026, for example, Mercor experienced a data breach after a widely used open-source tool was caught up in a broader supply chain attack. The incident hit close to the core of Mercor’s business, which runs on data and trust, and led to a thorough evaluation of how the breach occurred, what was accessed, and how to ensure something like it never happens again. Just as importantly, Mercor communicated directly with its customers every step of the way as rumors swirled on social media about the incident “Having a complete understanding of what actually happened and having really strong relationships with customers gave us a lot of confidence that we would get through the incident and be on the other side even stronger,” says Brendan. In the 60 days following the breach, Mercor expanded its relationship with all the frontier labs and added $300 million to its revenue run rate.

Operational threats aside, there’s nothing easy about the problems Brendan and Adarsh want to solve: helping AI reach expert-level capabilities, unlocking that expertise for enterprises, and making labor markets more efficient.

“With knowledge work, there is this huge long tail of context that lives in people's heads that models will need to learn and understand over time,” Brendan says. “Making that happen in the enterprise space will be especially difficult and a lot of work. It feels like there will be enormous demand for human labor and human jobs for many decades to come.”

As models become more capable, Brendan and Adarsh believe the future of work looks like what they’ve built: skilled workers training AI agents to work alongside them and take on redundant tasks while they focus on higher-value work.

To accelerate toward that future, they want more experts on their platform, better skill assessment and job matching, and to make each step in the process faster through automation. It’s essentially the same approach they took three years ago in their dorm rooms when they saw an imbalance in the market for Indian software engineers, but at a profoundly greater scale.

Brendan and Adarsh believe the market for knowledge workers is wildly inefficient in part because there is no marketplace that provides a true network effect for the unique skills workers have to offer. (Don’t get Brendan started on LinkedIn, which he describes as “the worst product I use every day.”) Drivers have Uber. Homeowners have Airbnb. Expertise needs its own network. Creating what would effectively be the definitive network for organizing human intelligence to power the AI economy is an audacious goal. It’s also an opportunity to solve problems that have plagued labor markets for as long as there’s been work.

“Building the network to understand what everyone is working on, what their skills and expertise are, what they’re capable of, and then codifying that in the models and the systems for every company is the most important network effect that hasn't been built,” says Brendan.