Developing the token economy urgently requires an institutional framework that brings together tiered authorization in advance, full-process logging during execution, and performance-based evaluation afterward. Image generated by AI

As agents move artificial intelligence (AI) beyond content generation toward task execution, tokens are gradually becoming the basis for measuring usage and settling payments for AI services. Yet measurable usage does not necessarily establish the value created, much less clarify who bears responsibility.

Developing the token economy urgently requires an institutional framework that brings together tiered authorization in advance, full-process logging during execution, and performance-based evaluation afterward, extending measurement, pricing, and governance from individual calls to the entire task lifecycle. This would curb inefficient calls and wasted computing resources while building the trust needed to integrate AI agents into core production, transaction, and public service processes. Governance capacity would thus become a competitive advantage as the intelligent economy expands.

Tokens are the basic units that large language models (LLMs) use to process information. The token economy—built around token usage, distribution, pricing, and payment settlement—provides a new yardstick for measuring resource consumption and transaction costs in AI services. As of June 2026, China’s average daily token usage had exceeded 500 trillion, reflecting explosive growth. Nevertheless, measurability is merely the starting point for market formation, not the endpoint of value assessment. Token counts indicate how much has been used, but cannot, on their own, establish what has been accomplished, what benefits have been generated, or who should be held responsible when something goes wrong.

The same number of tokens can yield very different value depending on whether they are used for casual conversation, drug research and development, or industrial quality inspection. Even for the same task, using more tokens does not necessarily mean better results—it may instead indicate repeated searches, unproductive reasoning, or inefficient process design. If token usage is treated as a direct proxy for output in the intelligent economy, evaluations may favor scale at the expense of effectiveness. This could even incentivize verbose model outputs and redundant calls, increasing the consumption of computing resources and energy.

Tokens are therefore better suited to serving as the basic units for resource accounting and service pricing, while complete tasks should serve as key units for value assessment, risk evaluation, and accountability. The former allows AI service costs to be calculated, while the latter allows results to be verified and responsibility to be traced.

Moving from usage measurement to task governance provides the institutional foundations for large-scale deployment without imposing additional burdens on innovation. In practice, many companies are willing to use AI to draft copy or search for information, yet remain cautious about allowing it to access core data, execute transactions, or control equipment. Their key concern is not whether a model can “say” the right things, but whether an AI agent can reliably “act” within the scope of its authorization. AI agents will struggle to enter high-value business processes if the limits of their authority are unclear, their actions cannot be reconstructed for review, their results are difficult to verify, and responsibility cannot be traced.

Conversely, clearer governance rules make transactions more predictable, giving companies greater confidence to open up new applications, commit data, and reorganize processes. China’s AI industry is expanding globally as part of the country’s “new tech trio,” alongside robotics and innovative pharmaceuticals. Its competitiveness will depend on model performance and token prices, as well as its ability to deliver reliable services, adapt to regulatory frameworks, and provide sustained support.

The token economy’s methods of measuring usage are still not fully aligned with the way AI agents operate. For example, charging by usage volume can reflect some computing costs, but it does not adequately capture task success rates, the degree of human intervention, or actual returns. Inefficient usage may even generate higher revenue on paper. A one-off, blanket authorization cannot adequately address the changing risks that arise as AI agents execute tasks continuously, and the same agent may require different permissions depending on the data, tools, and scenarios involved.

In addition, AI agent tasks typically involve multiple stages, including data provision, model development, application deployment, platform operations, and end-user interaction. Identifying the source of a risk from the output alone is difficult. Addressing these challenges requires redesigning measurement, authorization, recordkeeping, and accountability mechanisms around the task as a whole.

Tokens make it possible to measure AI services with precision, but token counts do not themselves create value. What matters is whether AI can complete a task successfully while keeping costs under control. Moving the token economy from measurement to governance means bringing resource accounting, performance accounting, and accountability together into a coherent system, ensuring that every call has clear boundaries, every process is verifiable, and every outcome can be evaluated. With these safeguards, AI agents can both undertake tasks and see them through successfully, strengthening the intelligent economy by making governance integral to its development.

Jin Wuwenhao and Lu Niehai (professor) are from the Zhejiang Provincial Institute of Socialism.