The AI Training Economy: The Human Expertise Behind the Agent Revolution

The challenge of training AI models on human expertise is not new. What is new is that three college dropouts just built a $10 billion business by productizing that exact problem at scale.

The effectiveness of AI is all about the data. The first generation came from the internet, which enabled companies to build powerful general-purpose models. But for AI to have a major economic impact, someone has to get humans to tell the model, step by step, how they actually do their work.

Bloomberg Businessweek published “The $10 Billion Startup Training AI to Replace the White-Collar Workforce” last Thursday — and it is essential reading for anyone helping clients navigate the transformational impact of AI.

Commercializing AI Training

Mercor.io is a San Francisco-based startup, co-founded by three 19-year-old high school friends, believing that a large portion of what humans do in companies is going to transform to training the models that power AI agents. And while this is an audacious claim, the scale of what they’ve built so far is impressive.

What Mercor's contractors are actually doing is training the underlying AI models by teaching them judgment, domain expertise, and nuance. Those trained models then become the engines that power AI agents: autonomous systems that can research, analyze, write, and execute tasks on behalf of professionals. You can't have capable agents without well-trained models underneath them.

In less than three years, Mercor has tens of thousands of contractors on its platform with annualized recurring revenue of approximately $500M. With their latest funding round in October, they were valued at $10 billion — five times what it was thought to be worth just a few months earlier — and they have been profitable from day one.

Mercor seeks experts across fields, including legal specialists, investment bankers, and venture capitalists. A dermatologist could earn up to $250 an hour, while poets enhancing AI’s understanding of literary nuance can make as much as $150 an hour. Ironically, while many contractors are doing this work to escape unemployment, they may be training AI to take their job. The optimistic view is that AI will remove the tedious and administrative parts of their work, freeing them to focus on higher-value aspects of their roles.

The next key question is: are the machines getting good enough to do the work themselves? In the past year the AI industry has been trying to address this through a growing ecosystem of evaluation frameworks. Mercor has developed its own benchmark, called the AI Productivity Index (APEX), better suited for the judgment-heavy work that white-collar professionals do every day. It measures professional performance and publishes the results publicly.

The Bloomberg article raises a question worth contemplating: if training AI has been a core requirement for nearly a decade, why did it take three college dropouts to commercialize it — and why didn’t the existing AI ecosystem get there first?

Why Training AI Agents Requires Expertise

Almost a decade ago we started building and implementing AI-based enterprise solutions at IBM with Watson across industries — energy, financial services, insurance, and more. One of the most public examples was GEICO. What made GEICO’s implementation particularly challenging, and instructive, was that, unlike expert-facing applications in medicine or law, customers are unpredictable. They don’t know the right next question to ask. Getting Watson to handle that required intensive human expertise to structure and train.

As I learned with Watson, the hardest part of that process is finding the right domain experts and structuring their input. Mercor essentially productized that exact problem.

Existing AI Ecosystem Blind Spot

OpenAI, Anthropic, and Google DeepMind were laser-focused on model capability. Human feedback was viewed as an operational cost to manage, not a business to build. They outsourced it to Scale AI, Surge AI, and internal contractors.

Structurally, this probably was a challenge for them too. OpenAI couldn’t build a marketplace that also serves Anthropic and Google. The moment you become a neutral platform, you’re commoditizing your own training advantage. An independent third party like Mercor has no such conflict — they’ll work for everyone simultaneously. In addition, a staffing marketplace with 30,000 contractors is messy, operationally intensive, and doesn’t map to any of the metrics of the big model developers.

Where Fast Followers Can Still Win

The significant demand for training AI agents should create an opportunity for fast followers. The most credible candidates aren’t likely to be tech companies, but professional services and staffing firms who already have the credentialed expert networks. For example, companies like Axiom Law (legal outsourcing) could expand into AI training for legal models — they have the supply side, they just need the AI-matching infrastructure. You would think that the big consulting firms like McKinsey, Deloitte, and Accenture would be naturals, as they all have massive expert networks, domain credibility, and existing relationships with the AI labs. Any of them could build a premium version of Mercor targeting their high-stakes domains (healthcare, financial services, legal) and charge significantly more per hour. The fact that none have moved aggressively suggests they don’t see this fitting their business model.

The real moat Mercor is building isn’t the platform, but the proprietary data on what good human feedback looks like across thousands of tasks and domains. That evaluation intelligence will compound over time and get harder to replicate. Fast followers who wait another year and a half may find they’ve missed the window to compete on quality, not just scale.

I believe the scope and speed of this workforce transformation is underestimated by most organizations. At Chief Outsiders, we don’t claim to have all the answers, as the landscape is evolving too fast for anyone to keep up stride-for-stride. What we are doing is actively investing: building AI fluency across our consulting and professional services teams, and embedding AI capability with deep sales and marketing expertise directly into our GrowthGears® OS platform so our clients can translate this moment into competitive advantage. If you are wrestling with where and how AI fits in your go-to-market strategy, we welcome the conversation.

​