For the past two years, the AI economy has been measured in tokens.
How many tokens are generated? How much does inference cost? How fast are token prices falling? Which model gives you the best output per dollar? The rising consumption of tokens and costs have also been the source of controversy as more US companies turn to Chinese Open Source models to save costs.
That metric still matters. But it may not be the whole story.
A recent thread by Benchmark Partner Everett Randle framed the next shift as the rise of the Task Economy: a market where the scarce input is not just compute or tokens, but the human work needed to make AI useful in the real world. Inference tokens have become the default proxy for tracking AI growth, but the next trillion-dollar category may be built around data and tasks, not tokens alone. But wait a second, wasn’t AI supposed to automate work?
Yes. But to automate work, you first need to define it, price it, evaluate it, supervise it and continuously improve it. That creates demand for a new kind of infrastructure: platforms that break knowledge work into tasks, match those tasks with experts, capture their judgment, and turn that judgment into training data, evaluations, workflows and eventually automated services.
This is where companies like Mercor, Scale AI, Surge AI, expert networks like GLG, Fiverr, Deel and vertical AI-native services start to look connected.
They are all part of the same shift: from hiring people to buying outcomes, and from buying outcomes to training machines to deliver them.
From SaaS to services to tasks
Earlier this year I wrote on VC Cafe about AI-native services: companies that do not sell software seats, but sell completed work. Not software that helps the accountant. Software that closes the books. Not software that helps the recruiter. Software that finds and qualifies the candidate. Not software that helps the broker. Software that gets the customer insured.
The task economy is the next layer underneath that thesis.
If AI-native services sell the finished work, the task economy is the supply chain that makes that work possible.
A legal AI service still needs lawyers to create examples, judge edge cases and review high-stakes outputs. A healthcare AI service still needs clinicians to label symptoms, evaluate responses and define safe escalation paths. A self-driving car company still needs humans to label road scenes, identify corner cases and review ambiguous footage. A customer support AI company still needs real human operators to define what “good resolution” looks like.
This is why Sequoia’s “services are the new software” thesis matters. Julien Bek argued that replacing an outsourced service with an AI-native services provider is much cleaner than replacing internal headcount, because the customer already accepts that the work can be done externally and already has a budget line for the outcome.
YC made a similar point in its 2026 requests for startups, highlighting AI-native service companies in areas like insurance brokerage, accounting, tax, audit, compliance and healthcare administration. YC’s point was simple: services spend is much larger than software spend, and many services are already outsourced.
The missing piece is that “services” are not delivered as one big blob. They are delivered as thousands of tasks.
That is the atomic unit.
Mercor as the clearest signal
Mercor is probably the most visible example of this new market.
The company started as an AI recruiting platform and evolved into a marketplace for expert human work used to train, evaluate and improve AI models. In October 2025, Mercor announced a $350 million Series C led by Felicis, valuing the company at $10 billion.
By June 2026, Sacra estimated that Mercor had reached $2 billion in annualized gross revenue, up from $760 million at the end of 2025. Importantly, Sacra notes that this is gross revenue, before contractor payouts, and estimates that contractors receive 60–70% of top-line revenue.
Mercor’s CEO Brendan Foody also claimed on X that the company grew from $1 million to $2 billion in ARR in 24 months, with enterprise demand becoming the fastest-growing segment. The public evidence around that specific claim is still largely company/X-led, and some observers have questioned whether “ARR” is the right frame for a marketplace-heavy model.
But even with those caveats, the signal is hard to ignore.
This is not just AI labs buying cheap data labels. It is Fortune 500 companies, AI application developers and frontier labs buying structured human expertise at scale.
Mercor has also published benchmarks showing that current AI agents still struggle with complex professional services work. In one APEX-Agents benchmark, AI agents completed less than 25% of real-world consulting, banking and legal tasks on the first try, and only 40% after multiple attempts. The harder the task, the more human planning, context and file navigation still mattered.
That is exactly why the task economy exists.
The models are getting better, but the path to improvement still runs through expert human judgment.
Why this is not Mechanical Turk
It is tempting to compare the task economy to Amazon Mechanical Turk. That would miss the point.
Mechanical Turk was built for small, generic, low-context microtasks. The new task economy is built for high-context, domain-specific, often regulated work.
The early internet data-labeling model was: “Can a crowd worker tag this image?”
The AI-native task model is: “Can a former investment banker evaluate this financial model?” “Can a lawyer identify whether this contract clause creates real risk?” “Can a physician judge whether this AI-generated triage answer is safe?” “Can a Hebrew-speaking operator with local context evaluate whether this customer support answer sounds natural?” That is a different market.
In the words Everett:
Tasks are the “unit of practice” in reinforcement learning: a model is given an initial state and an environment to act in, and its behavior is scored by a reward signal/verifier. Across many tasks, those scores are aggregated into a training signal that shifts the model’s behavior toward what scored well. Strictly speaking, “task” refers to this RL post-training substrate. But I’ll use it more loosely to stand for the unit of data-driven improvement generally, since the industry is rapidly inventing new forms that data takes in the service of making models better, and candidly because Task Economy has such a nice ring to it. I also want to distinguish this category from the dated moniker of “data labeling”, which brings to mind bounding boxes and thumbs up/down for LLM responses — the market has evolved well beyond these primitives over the last couple years into much more complex & high value tasks.
Grand View Research estimates that the global data collection and labeling market was $3.77 billion in 2024 and could reach $17.1 billion by 2030, growing at a 28.4% CAGR.
At the same time, the expert network industry reached roughly $3 billion in 2025, with Inex One estimating 12% annual growth between 2023 and 2025. Corporate customers are becoming a bigger adoption driver, making up around 45% of clients by number. Those two worlds are converging.
Expert networks historically monetized access to human judgment through calls, transcripts and surveys. AI data companies monetize expert judgment through labels, rubrics, evaluations, red-teaming and training datasets. AI-native services monetize expert judgment by embedding it into workflows that eventually become more automated. The task economy combines all three.
The platform stack of the task economy
The winners in this category will probably not look like old freelance marketplaces.
They will need to solve several layers at once:
- Task decomposition
Turning messy business workflows into clear units of work that can be assigned, measured and improved. - Expert sourcing
Finding people with real domain knowledge, not just available labor. - Verification and quality control
Knowing which expert is good, which answer is correct, and which tasks require multiple reviewers. - Workflow and data capture
Turning human work into reusable training data, evaluation sets, playbooks and system prompts. - Compliance and payments
Managing contractors across countries, currencies, tax regimes and employment classifications. - Automation over time
This is why Deel is already positioning around AI data-labeling workforces. Deel’s own materials describe AI data labeling as a use case for global contingent workforces, with emphasis on fast scaling, local linguistic and cultural expertise, compliance, onboarding and bulk payments.
That is also why Scale AI became strategically important enough for Meta to invest $14.3 billion for a 49% stake, valuing Scale at $29 billion. Reuters reported that Scale’s Remotasks and Outlier platforms recruit and manage gig workers who manually label data for AI models.
In other words, the task economy is not just a labor marketplace. It is part of the AI supply chain.
The Israel angle
Israel has a few reasons to pay attention to the task economy. It already produced one of the defining companies of the online freelance era: Fiverr. Its Q1 2026 results show both the pressure on the old marketplace model and the opportunity ahead: active buyers declined 17.8% year over year, but annual spend per buyer rose 15.4%, services revenue grew 30%, and the company is now talking about combining expert freelancers with GenAI models and agents. That is the transition in one public company: fewer simple tasks, more complex work, and more AI-enabled delivery.
Israel also has relevant infrastructure companies. Dataloop, acquired by Dell for $120 million, built technology for managing, labeling and processing unstructured data used to train AI models. Papaya Global operates in the workforce/payments layer, helping companies manage global payroll and workers across multiple countries. Calcalist reported in January 2026 that Papaya was in advanced acquisition talks at a potential $3.5–$4.5 billion valuation. Add Israel’s strengths in cyber, data infrastructure, fintech, developer tools, healthcare and enterprise automation, and the opportunity becomes clear: not to build another Mechanical Turk, but to build trusted task infrastructure for complex verticals.
The most interesting wedges for Israeli founders may be AI data operations for regulated industries; expert-in-the-loop evaluation for Hebrew, Arabic and multilingual enterprise AI; compliance and payroll infrastructure for distributed AI taskforces; data provenance for training workflows; and vertical AI-native services that start with expert humans and automate more of the workflow over time. Israel’s edge is not cheap labor. It is domain depth, speed, technical talent and the ability to turn operational pain into software.
What changes for founders
For founders, the task economy changes the startup playbook. The old SaaS question was: “What tool can I sell?” The AI-native services question is: “What work can I take off the customer’s plate?” The task economy question is more specific: “What is the smallest repeatable unit of valuable work, who can do it best today, and how does every completed task make the system smarter?”
That last question is where the moat may come from. If all you do is route tasks to people, you are a staffing company. But if every task produces proprietary workflow data, expert feedback, evaluation sets and automation loops, you may be building something much more defensible.
Pricing will likely change too. SaaS priced by seat. Marketplaces priced by take rate. AI-native services may price by outcome. The task economy may price by completed task, verified judgment, successful evaluation, resolved ticket, closed workflow or training data value. Tokens measure model usage. Tasks measure economic work.
The risk: platform work without dignity
There is also a darker side. Platform work has always created tension between flexibility and precarity, and the AI task economy could make that worse if expert work is fragmented into disposable tasks, workers are treated as interchangeable, and platforms capture most of the upside from models trained on human judgment.
Recent reporting from The Verge and Business Insider has highlighted concerns from AI training workers around instability, pay changes and the experience of training tools that may eventually replace parts of their own work. Founders in this category should take that seriously. The best companies will not just arbitrage labor. They will build trust with both sides of the market through transparent pay, clear project expectations, portable reputation, worker protections where appropriate, strong privacy controls and a credible answer to the question: “What happens to the humans once the model improves?”
If the task economy becomes a race to the bottom, it will be fragile. If it becomes a way for experts to monetize judgment, train better systems and participate in new forms of AI-enabled work, it could be one of the most important labor-market shifts of the decade.
The next AI metric
A few years ago, we measured AI progress by model size. Then we measured it by benchmark scores. Now we measure it by tokens, inference costs and revenue run-rate. The next metric may be more practical: tasks completed.
How many useful tasks can an AI system complete? How much expert supervision does each task require? How much does each completed task improve the system? What percentage of the workflow can move from human-led, to human-reviewed, to agent-led?
That is where the task economy becomes interesting. Not because humans disappear from the loop, but because for now, humans are the loop. The companies that organize, price, verify and learn from that loop may become some of the most important infrastructure companies in AI.
For Israeli founders, this is a category worth studying closely. The opportunity is not to copy Mercor. It is to ask where Israel has domain expertise, trusted talent, complex workflows and global customers willing to pay for outcomes. The next great AI company may not sell software. It may not even sell services. It may sell the task graph behind the work.
He is a former General Partner at Google Ventures (GV) in Europe, former head of Google for Entrepreneurs in Europe, and founding head of Campus London, Google's first startup hub. Eze writes on Israeli tech, venture capital, artificial intelligence, and founder strategy.
He is also the founder of Techbikers, a nonprofit that brings together the startup ecosystem on cycling challenges in support of Room to Read.
- The Task Economy: Why AI’s Next Big Market May Be Human Work, Not Tokens - July 8, 2026
- Every Founder Eventually Looks for the Exit Sign - July 7, 2026
- Weekly FIRGUN Newsletter – July 3 2026 - July 3, 2026
