If AI agents become more efficient, who profits?

Author: Zhang FengWhen OpenAI's latest model achieves a 10-fold leap in inference efficiency, JPMorgan Chase's COIN intelligent agent improves the efficiency of legal document parsing by 3600 times, and ByteDance's Coze platform's monthly active developers surpass 100,000—AI agents have moved from the laboratory to large-scale applications, becoming a core force in reshaping the productivity landscape.AI is no longer a passive interaction tool, but an autonomous execution system capable of understanding goals, breaking down tasks, calling tools, and providing dynamic feedback. The efficiency revolution is sweeping across all industries, but a fundamental question is becoming increasingly acute: to whom does the enormous value created by AI agents actually go? And what do ordinary users who contribute massive amounts of data gain from it?1. AI The Rise and Penetration of Agent Efficiency EngineAI The essence of an Agent is an autonomous execution system built upon a large model, equipped with a complete architecture of "brain (large model) + memory (vector database) + tool invocation (API interface)". Compared to traditional AI, its core advantages are reflected in three dimensions: "autonomy", "coordination", and "evolution". From a functional iteration perspective, Agent has completed three stages of evolution. The first stage is "single-task execution," focusing on replacing repetitive tasks such as document generation and basic customer service. The second stage is "multi-task collaboration," capable of breaking down complex goals and linking multiple tools to complete the entire process. Google's A2A protocol further enables intelligent agents to collaborate across systems. The third stage is "multi-agent collaboration and autonomous evolution," where multiple agents collaborate and continuously optimize decision-making logic through user behavior feedback, forming a closed loop of "autonomous thinking—execution—reflection—iteration." By 2025, AI Agents have four core capabilities: goal understanding, task planning, tool invocation, and dynamic feedback, and can connect to thousands of tools such as ERP and medical imaging equipment. Application scenarios have achieved comprehensive penetration from the consumer (C-end) to the business (B-end). The global market size is projected to reach $32 billion by 2025, with enterprise services accounting for over 60%. In the financial sector, JPMorgan Chase's COIN intelligent agent saves over $200 million annually, while Boosted.ai's financial research intelligent agent can quickly analyze market data. In the healthcare sector, OpenAI's o1-mini, after enhancement and fine-tuning, improves gene analysis accuracy by 80%, and diagnostic logic and raw data hashes can be traced on the blockchain. In the industrial sector, Siemens' Industrial Agent reduces factory downtime by 30%, and Alibaba Cloud's ET Industrial Brain helps manufacturing companies improve yield rates by 5%. In the office sector, Kingsoft Office's WPS AI-powered writing service has reached 618 million monthly active users, improving document generation efficiency by 300%, and ServiceNow, based on "cost reduction value," has 44 million-dollar customers by 2024. Consumer-facing applications are also rapidly penetrating the market. The Honor Magic7's YOYO AI agent can automatically organize meeting minutes; the Thunderbird V3 smart glasses achieve a real-time translation accuracy of 95%; ByteDance's Doubao AI, leveraging Douyin's traffic, has become the largest native AI application in China; the AI programming assistant Cursor achieved an ARR of over $100 million in 12 months, with a 45% paid user rate. Low-code platforms such as Coze and iFlytek's Xinghuo AI agent platform allow ordinary users to build their own agents with zero barriers. Technological breakthroughs support all of this. Large-scale models, from GPT-4 to Claude 3.7 and Gemini 2.5 Pro, have seen continuous improvements in inference efficiency. OpenAI o3's dynamic computing power allocation is 10 times better than o1. Standardized communication protocols (MCP, A2A) enable efficient interaction between agents and tools, and between agents themselves. Computing costs have decreased, with the proportion of computing power required in the inference stage increasing from 20% to 60%, laying the foundation for large-scale applications. AI agents are propelling productivity into a new era of "intelligent automation." II. Profit Model and Beneficiary Structure: Dividends Concentrated Among Leading Players The profit model of AI Agents has formed a diversified monetization system of "basic layer - intermediate layer - application layer", but the distribution of efficiency dividends shows a clear "concentration among leading players" characteristic. (I) Current Mainstream Profit Models Upstream Model and Infrastructure Layer: The core monetization model is "technology licensing + computing power services." This is dominated by giants such as OpenAI, Google, Microsoft, and domestic companies like iFlytek, ByteDance, and Alibaba Cloud. Large model licensing is charged based on usage or annual fees, while computing power leasing is charged based on consumption. Tencent Cloud TBaaS provides blockchain traceability services. Data annotation services target vertical industries, with medical annotation costs reaching 800 RMB/hour. Midstream Mass Production and Customization Layer: Monetization methods include "customized services + platform commissions + subscription fees." Accenture, ServiceNow, and other enterprise-level solution providers have average order values ranging from hundreds of thousands to millions of dollars; platforms like Coze take commissions from developers; general-purpose agents are charged based on subscriptions or usage counts, with Salesforce Agentforce charging $2 per use, and ServiceNow taking a 10% cut of the customer's cost reduction revenue. Downstream Scenario Implementation Layer: Monetization relies on "scenario monetization + advertising revenue." AI-powered sales agents facilitate transactions and generate commissions. Mingyuan Cloud helps enterprises reduce customer acquisition costs by 60% and extracts commissions on completed transactions. Free C-end agents such as Doubao AI and Quark AI Super Box embed ads and charge based on impressions or clicks. AI terminal hardware (AI Pin, Rabbit R1) has also become an important monetization channel. (II) Core BeneficiariesCategory 1: Underlying technology and platform giants, the biggest beneficiaries. Control large models, computing power, and data resources to seize profits at the top of the industry chain. OpenAI's revenue exceeded $10 billion in 2024; Microsoft Azure + OpenAI related businesses grew by more than 50% annually; Google seized the enterprise market with its Gemini large model and A2A protocol. Domestically, ByteDance has built a complete "technology + scenario + ecosystem" layout; iFlytek's Xinghuo Big Data Model saw a 213% increase in paying customers in 2024; Alibaba Cloud and Tencent Cloud share the computing power dividend; Yonyou's BIP intelligent agent platform serves over 50,000 enterprises. The second category: large enterprises and vertical industry leaders, direct beneficiaries. By replacing repetitive labor with agents and optimizing processes, they significantly reduce costs, increase efficiency, and expand market share. JPMorgan Chase's COIN saves $200 million annually; Goldman Sachs optimizes trading strategies to improve revenue; Siemens and BMW optimize their supply chains; Kingsoft Office Enterprise Edition's subscription revenue share increased from 15% to 35%; ServiceNow has 44 million-dollar customers with a 92% renewal rate. The third category: Developers and small and medium-sized service providers, with limited benefits. LangChain has been downloaded over 100 million times, and 80% of enterprise-level intelligent agents are developed based on it. Some developers generate revenue through customization and plugin sales, but lack core technology, computing power, and traffic, resulting in mostly niche applications. The profit margins of supporting service providers such as data annotation and computing power leasing are also squeezed by industry giants. Crucially, the vast majority of ordinary users—the core data providers supporting AI agents—receive virtually no benefit from the current profit model. User chat logs, behavioral patterns, and usage feedback are key materials for agent training and iteration, yet they receive no reward and instead face risks of data leaks and privacy violations. III. Actual Impact on Users: A Mixed Bag of Advantages and Disadvantages, with a Severe Lack of Rights Protection The efficient operation of AI Agents is essentially "data-driven," with users as implicit contributors. However, their contributions are not respected, resulting in a dilemma of "more disadvantages than advantages." (I) Positive Impacts: Convenience and Efficiency Improvement For ordinary users, the most direct positive impact is the convenience of life and the improvement of personal efficiency. In work scenarios, AI programming assistants automatically complete code and check for errors; in life scenarios, smart glasses provide real-time translation and intelligent agents organize minutes; in learning scenarios, personalized assistants create plans and provide tutoring. Zero-code platforms lower the technical threshold, and free basic services allow users to enjoy the benefits of AI at zero cost. (II) Negative Impacts: Data Plunder and Rights Infringement Firstly, data is collected and used without compensation, resulting in zero benefit for users. To optimize their products, platforms collect user usage records, interaction content, behavioral trajectories, and even private information for model training and iteration, becoming their core profit-generating resource, while users receive no economic return. Medical consultation data improves diagnostic accuracy, but users face the risk of data leakage. This kind of "uncompensated data appropriation" is a plunder of users' labor. Secondly, the risk of privacy leaks is prominent. Some platforms collect data covertly without clearly disclosing the scope and purpose; data storage and transmission mechanisms are imperfect, leading to frequent leaks and thefts. Leaks of medical records and genetic data infringe on patient privacy; theft of financial information leads to financial losses; leaked conversation content damages personal image. AI's autonomous iteration is more likely to lead to the "secondary use" of privacy information, indirectly leaking it to other users. Thirdly, there is data abuse and algorithmic bias. Platforms abuse data to precisely push advertisements, induce consumption, and even engage in "big data price discrimination," damaging the right to fair trade. Training data is biased; recruitment agents discriminate against women and older job seekers; credit agents discriminate based on region, depriving users of equal rights. Users lack control over their data, unable to decide the scope and duration of its use, and find it difficult to request deletion. Fourth, over-reliance leads to a decline in personal abilities. Long-term reliance on agents to generate reports, translate documents, and make decisions may lead to a decline in independent thinking and problem-solving abilities, especially for teenagers, affecting knowledge accumulation and cognitive development. Homogeneous responses also limit the diversity of thought. Overall, users face the reality that "limited convenience cannot outweigh the deprivation of data rights, invasion of privacy, and threats to security." If these contradictions remain unresolved, they will not only harm user rights but also erode user trust and hinder the long-term development of AI agents. IV. Solution: Building a Multi-faceted Collaborative Rights Protection System with Blockchain as the Core To address the issues of "uneven distribution of efficiency dividends" and "insufficient protection of user data rights," it is not enough to rely solely on market self-regulation or platform self-discipline. A multi-faceted solution of "technical support + institutional guarantees + ecological collaboration" is needed. Blockchain technology, with its immutable, traceable, and decentralized characteristics, serves as the core support. (I) Solving the Data Rights and Distribution Problem with Blockchain Technology as the Core Firstly, data ownership confirmation and traceability clarify user ownership. User data is encrypted and uploaded to the blockchain, generating a unique "digital ID card" (hash value). The entire process of data collection, use, transmission, and deletion is recorded on the blockchain, giving users complete control. In case of leakage, the source can be quickly traced. Medical diagnostic agents can upload diagnostic logic and raw data hashes to the blockchain; supply chain optimization agents upload decision-making data to the blockchain for verification by partners. Tencent Cloud TBaaS already supports consortium blockchain/private blockchain deployment. Secondly, decentralized storage ensures privacy and security. Traditional centralized storage, once attacked, can lead to massive data leaks. Blockchain encrypts and distributes data across multiple nodes, ensuring no single node can control all data; access requires user private key authorization. Employing an "on-chain/off-chain collaboration" approach, sensitive data is stored in a private off-chain database, with only data fingerprints (SHA-256 hashes) and operation logs uploaded to the blockchain, balancing privacy and traceability. Tencent Cloud COS object storage + Data Variants service can achieve this model. Thirdly, smart contracts enable fair profit distribution. Pre-defined data usage rules and profit distribution ratios are established. When the platform uses user data, the smart contract automatically pays the user tokens or cash. Platform profits are automatically distributed proportionally to data-providing users, realizing "whoever provides the data, benefits." A blockchain-based AI Agent data trading platform can be built, where users upload encrypted data, the platform pays to acquire it, and the entire transaction is traceable on the blockchain. (II) Improve Laws and Regulations, and Strengthen Data Supervision and Rights Protection First, clarify user data rights. Accelerate the revision of the "Personal Information Protection Law" and the "Data Security Law," clarifying that users have ownership, usage rights, revenue rights, and deletion rights to their personal data. Platforms must inform users of the scope, purpose, and revenue distribution method in advance of data collection and use, and obtain explicit authorization. Hidden collection is prohibited. Any leakage, abuse, or illegal sale will incur civil, administrative, and even criminal liability. Formulate a special "AI Agent Data Governance Regulations" to regulate all aspects of data collection, training, and iteration. Second, construct a diversified regulatory system. Establish a framework of "government regulation + industry self-regulation + social supervision." The Cyberspace Administration of China, the Ministry of Industry and Information Technology, and the State Administration for Market Regulation will strengthen enforcement and conduct regular inspections of data security, privacy protection, and revenue distribution; industry associations will formulate self-regulatory guidelines; and users and media will be encouraged to report violations. A regulatory traceability platform will be built using blockchain technology to achieve real-time monitoring of the entire data process. Thirdly, strengthen algorithm regulation. An algorithm registration and evaluation system will be established. Platforms must register their algorithm models, training data, and algorithm logic. Regulatory departments will organize experts to conduct regular evaluations, focusing on issues such as algorithm bias, misuse of user data, and manipulation of decision-making. Platforms are encouraged to disclose their algorithm logic (after anonymization) to improve transparency. (III) Building a Collaborative Ecosystem to Promote the Fair Distribution of Efficiency Dividends First, break the monopoly of leading platforms and support small and medium-sized developers and users. The government provides funding, technology, and computing power support to small and medium-sized developers; it guides leading platforms to open up technology, interfaces, and traffic, lowering the threshold for entrepreneurship. It encourages major platforms to lower entry barriers, provide free computing power, and expand the scope of interface opening. Second, establish a user participation mechanism to guarantee their right to speak and their right to benefit. The platform establishes a user representative committee to participate in product design, data usage, and revenue distribution decisions. In addition to automatic revenue distribution via smart contracts, the platform can return a portion of its profits to users in the form of dividends, coupons, and free services. ServiceNow can return a percentage of its commission revenue to enterprise users who provide data; the C-end platform offers free premium services and cash rewards based on data contribution. Thirdly, it strengthens technological innovation to balance efficiency and rights. It develops technologies such as federated learning and differential privacy to achieve "data usable but invisible," allowing the platform to train models without obtaining raw data. It optimizes algorithm logic to curb algorithmic bias. It promotes multi-agent collaborative technology to improve overall efficiency and expand the scope of beneficiaries. V. Sustainable Development in the Digital Age The efficiency revolution of AI agents has created enormous commercial and social value, but behind this efficiency improvement lies the unpaid support of massive amounts of user data, a severe imbalance in the distribution of benefits, and the continuous infringement of users' data rights and privacy. The answer to the question, "AI agents are more efficient, but who profits?" should not only be given to tech giants and large enterprises, but also to every ordinary user who provides data—without user data support, an efficiency revolution is impossible. To achieve sustainable development of AI agents, it is essential to break the pattern of "top performers benefiting while users suffer losses." This requires building a multi-faceted solution based on blockchain technology, encompassing "technical support, institutional guarantees, and ecological collaboration." This solution must clearly define user data rights, ensure fair distribution of benefits, curb data abuse and privacy leaks, and ensure that efficiency benefits reach every participant. Only in this way can AI agents, while improving efficiency and creating value, also uphold fairness and justice, truly becoming a powerful driving force for human societal progress and achieving a win-win situation of "efficiency improvement" and "rights protection."