For the past few years, the AI community has been obsessed with tokens. We measured the success of a model by its context window, the cost per million tokens, and the sheer volume of the training corpus. The prevailing logic was simple: more data equals more intelligence. But as the industry pushes toward the next frontier of reasoning, a quiet realization has set in among the engineers at the world's leading labs. The era of scraping the open internet for raw text has hit a ceiling. The low-hanging fruit of general knowledge is gone, and the models are beginning to plateau because the internet does not contain the one thing they need most: the tacit knowledge of a human expert.
The Rise of the Task Economy
This bottleneck has given birth to what is now being called the task economy. In this new paradigm, the fundamental unit of value is no longer the token, which is the smallest unit of text processing, but the task, which is the smallest unit of model improvement. While tokens are the currency of inference, tasks are the currency of evolution. The task economy is a specialized market dedicated to generating and supplying high-quality, procedural data that teaches a model not just what the right answer is, but exactly how a professional arrives at that answer.
The explosive growth of this market is best illustrated by the trajectory of Mercor, a platform designed to supply this expert-level data. The financial metrics are staggering. In February, Mercor achieved an annual recurring revenue (ARR) of $1 billion. Only four months later, that figure doubled to $2 billion. This rapid acceleration is not a fluke of venture capital hype but a direct reflection of the desperation among AI labs to acquire specialized human intelligence. In this market, the value is derived from the actual working hours of experts—lawyers, doctors, software architects, and scientists—whose problem-solving processes are converted into structured data.
Unlike the previous era of data collection, which relied on massive crawls of Common Crawl or Wikipedia, the task economy focuses on the creation of synthetic and human-curated datasets that capture the nuances of professional expertise. The goal is to move beyond generalist capabilities and into the realm of deep domain mastery, where the model can replicate the sophisticated reasoning paths used by the top 1% of practitioners in any given field.
From Static Text to Reinforcement Learning
To understand why this shift is happening, one must look at the difference between how a model learns a fact and how it learns a skill. Traditional supervised learning is about pattern matching; the model sees a prompt and a corresponding answer and learns to predict the next token. However, professional expertise is rarely a straight line from A to B. It is a series of iterative steps, corrections, and strategic pivots. This is where the task economy diverges from the token economy.
In a task-based framework, the data provided to the model consists of an initial state, a specific work environment, and a set of reward signals or verifiers. Instead of simply being told that the answer to a legal query is X, the model is trained on the entire process: how to search the case law, how to weigh conflicting precedents, and how to refine the argument based on a verifier's feedback. This structure allows the model to utilize Reinforcement Learning (RL) to find the optimal path to a solution. The model treats the task as an environment where it can experiment, receive a reward signal (a high score from an expert or a programmatic verifier), and adjust its internal weights to favor the successful behavior.
This transition represents a fundamental change in the AI moat. For years, the competitive advantage was held by those with the most compute or the largest dataset. Now, the advantage belongs to those who can build the most effective data infrastructure for RL. This is why OpenAI and Anthropic are reportedly increasing their spending on expert data and agent training materials by ten times every year. The cost of intelligence is shifting from the cost of electricity and GPUs to the cost of high-end human labor. Some AI application companies have already scaled their spending on specific task-related data to over $100 million in short windows, recognizing that a model trained on 1,000 high-quality expert tasks is more valuable than one trained on a billion tokens of mediocre web text.
This shift is particularly critical for high-stakes domains like medicine, law, and finance. In these fields, a hallucination is not just a quirk; it is a liability. By training on tasks that include verification steps and reward signals, models can move away from probabilistic guessing and toward verifiable reasoning. The infrastructure that supports this—the platforms that can recruit experts, manage their workflows, and turn their cognitive processes into machine-readable reward signals—is becoming the most critical layer of the AI stack.
The transition from the token era to the task era marks the end of the brute-force approach to AI scaling. We are moving from a world of quantity to a world of quality, where the ability to capture and scale human expertise is the only way to break through the current performance plateaus. The companies that dominate the task economy will not just build better chatbots; they will build the first truly professional-grade autonomous agents.
