Summary of Key Points
Following the 2026 World Artificial Intelligence Conference, the focus in the intelligent agent industry has shifted from whether the technology is feasible to how to calculate its commercial viability. Enterprises—whether they are state-owned or private, large or small, even single-person companies—are accelerating their entry into this field, but they face three core challenges: How to optimize the cost of computing power? How to measure ROI (Return on Investment)? And how to handle the hidden costs associated with organizational integration? The industry is transitioning from selling models to selling results, and it is increasingly recognized that humans and intelligent agents are not in a competitive relationship but rather in a symbiotic one, where the skills of intelligent agents become the long-term intangible assets for businesses.
1. Enterprises Fully Embrace Intelligent Agents: State-Owned and Private Companies Exploring Different Approaches
Nowadays, both large state-owned enterprises and small companies are eager to utilize intelligent agents:
- State-Owned Enterprises: Exploring Deep Water (Paradigm Upgrades)
In the past, enterprise data was used for human purposes; now it is being utilized by intelligent agents. For example, China Electronics has established an “AI-native data factory” and a “software factory,” transforming data into actionable capabilities for intelligent agents, which has helped more than 20 leading companies increase their efficiency by 40%. However, they also point out that deploying AI is just the beginning; the real challenge lies in reorganizing the company—retraining employees and altering processes, costs that are often underestimated.
- Private Enterprises: Focusing on Scalable Implementation
Companies like Haidilao’s customer service have moved from having hundreds of staff handling orders and location inquiries to using intelligent agents to handle these tasks independently. More disruptive is the role of single-person companies, where one entrepreneur along with several intelligent agents can perform team tasks such as design, sales, and administration, eliminating the need for hiring and paying salaries, thus reducing the cost of trial and error. However, if an intelligent agent makes a mistake, a human still has to take responsibility.
2. The Revolution in Cost Calculation: From “Building Computing Power” to “Achieving Results”—ROI Is the Real Metric
In the past, evaluating the cost of intelligent agents focused on how many GPUs were purchased; now, it’s about what they actually achieve:
- Shift from Scaling Computing Power to Improving Efficiency
Enterprises used to focus on the hardware when purchasing computing power; now, they consider the value of “tokens”—the cost per call or interaction with an intelligent agent (similar to mobile data usage fees). Companies like Suiyuan Technology are working on “core-model collaboration,” designing chips and models together to make token generation faster and more cost-effective. Some also aim to deploy computing power at the edge (e.g., in phones and devices) to reduce token costs.
- Avoid the Token Trap: Measure ROI, Not Just Token Volume
Many companies initially overused intelligent agents, resulting in higher token costs than employee salaries (for example, AI programming costs could be ten times that of manual labor). The industry is now using “DAA” (Daily Active Agents) as a metric for success: the number of active agents multiplied by the amount of work each agent performs and the quality of that work divided by the cost. For instance, a fresh food supermarket that uses replenishment intelligent agents can reduce loss rates by 1%, resulting in additional profits of tens of millions per year, making AI investments almost negligible.
- Business Models Changing: From Selling Models to Selling Results
Previously, AI models were sold based on the number of tokens used; now, payment is based on the results. For example, Ronglian Cloud charges only if the intelligent agents perform well; otherwise, there is no charge. A human agent may cost 5,000-8,000 yuan per month, while an intelligent agent costs 2,000-3,000 yuan and can work 24/7.
3. Hidden Costs: Organizational Integration Is More Troublesome Than Technical Deployment
The financial implications of intelligent agents involve many invisible costs:
- Longer Human-Agent Integration Period
Optimizing models takes time, and employees need to transition from performing tasks to supervising intelligent agents. The costs associated with training, psychological adjustment, and employee turnover are difficult to quantify. For example, after implementing intelligent agents, it’s not the agents that adapt to company processes but humans who must adjust their CRM and ERP systems and change their work habits, often increasing cognitive burdens.
- Cross-Business Barriers
Those familiar with the industry lack knowledge of AI, and those knowledgeable about AI do not understand the business. Data is not shared across departments, and evaluations still focus on technical metrics (e.g., computing power) rather than business value (e.g., efficiency improvements), creating barriers to the successful implementation of AI.
4. The Future: Not Replacing Humans, but a Symbiosis of Humans and Intelligent Agents
Intelligent agents are not meant to replace jobs but to work alongside humans:
- Humans Moving from Executors to Value Definers
For example, Ronglian Cloud’s customer service agents no longer just answer calls; they train intelligent agents. Each confirmation or correction helps the agents improve their performance. Humans no longer need to align their output with that of the agents but ensure that the agents meet business standards.
- Intelligent Agents as Intangible Assets
Computing power and tokens may depreciate, but the industry knowledge accumulated by intelligent agents, the completed business processes, and the skills they develop become the true long-term assets for enterprises. For instance, Zhongshu Ruizhi’s “Meta-Causal” framework transforms business rules into digital foundations, enabling intelligent agents to go from describing what happened to suggesting how to proceed, while also being validated by humans, creating a symbiotic cycle.
Conclusion
The era of intelligent agents has just begun. The conference did not provide all the answers, but it brought these issues out of the laboratory into real-world business scenarios—how to accurately calculate costs and manage the relationship between humans and intelligent agents. These questions can only be answered through practical application by enterprises. The intelligent agents that will truly thrive are those that have been valued by humans, validated by business processes, and accepted by organizations. Computing power and tokens may eventually become obsolete, but the skills of intelligent agents will be the enduring assets of businesses.