The enterprise-level AI Agent market is becoming a hot spot for capital pursuit, but its business model and landing problems still need to be solved. This paper delves into the evolution of enterprise-level agents from tools to “digital employees”, analyzes their characteristics in terms of certainty, verticality and landing, and points out the limitations of MCP protocols in enterprise-level scenarios and the challenges of business model transformation, revealing the opportunities and dilemmas of the enterprise-level agent market.
Enterprise-level agents are becoming the consensus of the entire toB market.
Enterprise-level agents (toB AI Agents) are becoming one of the most sought-after tracks by capital in the world.
- In October last year, SierraAI, an enterprise AI intelligent customer service product co-founded by OpenAI Chairman Bret Taylor and former Google executive ClayBavor, completed $175 million in financing at a valuation of $4.5 billion;
- In May this year, RelevanceAI, a startup that helps companies build AIAgent, completed a $24 million Series B financing;
- Last month, Glean, an enterprise-level AI platform, completed a $150 million Series F financing, claiming that the enterprise has achieved “Google+AI knowledge graph + workflow automation”, with a valuation of more than $7.2 billion……
From the initial general-purpose agents that help you do analysis and research, to the “digital employees” that claim to be able to reconstruct enterprise processes, this wave has swept in at an unprecedented speed.
“Jiazi Lightyear” clearly pointed out in the previously released “China AI Agent Industry Research Report (2)” thatBy 2026, cognitive agents will cover 70% of complex decision-making scenarios in enterprises, redefining the productivity revolution.
Capital is popular, giants have entered the game, startups have sprung up, and enterprise-level agents seem to have become the consensus of the entire AI and even toB market.
However, under the hustle and sound, some questions still need to be clarified: Manus, flowith Neo and other general-purpose agents are in front, why do enterprises still need enterprise-level agents? What are the difficulties in implementing enterprise-level agents in business scenarios? Can it truly change the business model of toB software?
After 10 years of interaction design, why did I transfer to product manager?
After the real job transfer, I found that many jobs were still beyond my imagination. The work of a product manager is indeed more complicated. Theoretically, the work of a product manager includes all aspects of the product, from market research, user research, data analysis…
View details >
In other words: Is an enterprise-level agent a good business?
1. The rise of enterprise-level agents: from tools to “digital employees”
One sure fact is that there are more and more enterprise-level agents.
Take the domestic market as an example: in the CRM track, Sales Easy and Xiangxiang Sellers have launched a series of agents around sales scenarios in their products; Service providers such as Laiya and Jin Zhiwei are upgrading from RPA to digital employee platforms based on agent capabilities; In the field of HR, companies such as Beisen have also launched various enterprise-level agent products.
There are also many enterprise-level agents that are not listed.
From the perspective of technical architecture, a complete enterprise-level agent usually includes several core modules:Environmental perception, decision engine, memory system, and executive tool.
- Perception (multimodal input and environmental awareness): Enterprise business involves not only text, but also bills, contracts, drawings, and even audio and video. Agents need to have the ability to process multimodal information. Ren Xinqi, founder of Yuedian Technology, pointed out that Agent technology is migrating from plain text to multi-modality such as images and voice, which is an important direction for the evolution of its capabilities.
- Brain (Decision Engine): This is where the Agent’s core intelligence lies, typically driven by large language models (LLMs). It is responsible for understanding user intent, breaking down complex tasks into executable subtasks, and making dynamic decisions based on real-time feedback. Lin Fan, the core R&D of MinimaxAgent, mentioned that excellent agents must have excellent “planning and execution capabilities”, which is the key to measuring their intelligence.
- Memory (Long-Term vs. Short-Term Memory System): Agents need memory to maintain task context and learn from past experiences.
- Hands and feet (tool calls vs. actuators): The decisions of the “brain” ultimately require the hands and feet to execute. This includes calling external APIs, operating software interfaces, reading and writing databases, etc.
“Jiazi Lightyear” believes that the core value of enterprise-level AI Agents lies in “large model scheduling capabilities and full-link automation closed loops”.
The reason why a large number of toB companies have added Agent to their products is that it is not just an AI upgrade of traditional software, but has the potential to disrupt. itTruly synergize the “thinking power” of large models with the “execution power” of tools to form a complete closed loop that can solve practical problems.
In the interview of “Jiazi Lightyear”, almost all respondents agreed:The core value of enterprise-level agents is that they are evolving from an “auxiliary tool” to a “digital employee” who can complete tasks independently, thereby reconstructing the collaborative relationship between humans and machines. This is mutually confirmed by the judgment of the “Jiazi Lightyear” think tank report.
Source: Jiazi Lightyear Think Tank “China AI Agent Industry Research Report (2)”
For example, the lead generation agent can automatically search, analyze, and compare potential customers in search engines, customs data, and bidding websites, and automatically write and send development letters, and then transfer the task to real sales until the customer replies to the agreed communication time. This enables end-to-end lead generation and delivers the results of the “new lead” work.
When an enterprise-level agent starts to “work”, it is no longer just a tool in the hands of employees, but becomes a “colleague” of employees. This change in role is giving birth to a new paradigm of human-machine collaboration.
From a technical perspective, this iteration can be summarized as the evolution of the digital workforce. From RPA that relies on manual configuration in the 1.0 stage, to intelligent automation that integrates conventional AI capabilities in the 2.0 stage, and then to Agents driven by large models and with autonomous decision-making capabilities in the 3.0 stage, the capabilities of digital employees are constantly leaping.
The final picture is:Each business unit and even individual employees will form a symbiotic relationship with the exclusive AlAgent, realize cross-system and cross-departmental collaboration through the agent network, and build a new intelligent organization that responds dynamically and evolves independently.
2. Determination, vertical, landing, three labels of enterprise-level agent
But the problem is, today’s users may have already used general-purpose AI Agent products to some extent, so why do we still need enterprise-grade agents?
If we look at the product itself, whether it is Manus, Zhipu’s COCO, Kunlun Wanwei’s Tiangong superintelligence, or the above-mentioned enterprise-level Agent, they do have similar product forms and interactive experiences: a dialog box that completes tasks through natural language interaction. However, the difference in usage scenarios leads to two diametrically opposite underlying logics.
The first is the difference between users’ requirements for “certainty”.
Looking at most general-purpose toC agents on the market, their tasks are usually research, creation, and even entertainment. For example, write articles, research reports, generate PPT with one click, generate travel guides, etc. What these tasks have in common is a high fault tolerance rate, and users can even tolerate some degree of hallucination.
“For example, a user asked a general-purpose agent to investigate 100 websites and write a report based on the survey. Probably it only has a 70% success rate, but it is enough to complete the report. Hu Yichuan, co-founder and CTO of Laiye Technology, gave an example.
But enterprise agents are completely different.
Since enterprise-level agents need to access business scenarios, it means that enterprise-level agents must achieve “zero mistakes”.Because for enterprises, whether it is financial reimbursement, supply chain management or customer relationship maintenance, it all affects the whole body. An incorrect order processing and an inaccurate customer information entry can cause direct financial losses and reputational crises.
Luo Yi, vice president of sales and easy products, believes that compared with the independent planning and divergent thinking of general-purpose agents, the requirements of enterprise-level agents are convergence and certainty. This convergent mindset requires the designer of enterprise-level agents to define the business path and make the execution results verifiable.
“When we make our own agent products, our thinking is infinitely convergent, we only provide specific capabilities in specific scenarios and output stable results, rather than providing users with divergent results like general agents.” Luo Yi said.
The second is the difference between general and vertical.
The goal of the universal agent is to become a universal platform that can handle a variety of tasks, and to demonstrate advantages in the fields of data, web pages, finance, etc. through the collaborative work of various tools. Therefore, this type of agent is more inclined to deep research.
However, toB is an area that requires vertical focus to generate business value and improve efficiency.
For example, only by deeply binding CRM business processes, data, and scenarios can we create true closed-loop value. SaleseSeller’s Sales Associate Agent can automatically sync meeting minutes to the CRM customer activity record after the meeting ends, and create follow-up tasks for sales. This seamless integration with business systems is difficult for general agents to achieve.
The last and most important point isEnterprise agents need to be able to truly implement in business scenarios. This is both the difference and the biggest challenge.
Based on the goal setting of the general platform, the “toolbox” that the general-purpose agent can call is open, mainly composed of browsers, virtual machines, search engines, code interpreters and other basic Internet tools. In addition, the MCP ecosystem is becoming more and more perfect, and the general-purpose agent can call various tools through MCP to output the final result for the user.
However, in an enterprise environment, agents do not only rely on general Internet tools to complete a large number of business-related tasks, but also need to integrate with the enterprise’s business systems, such as ERP systems, CRM systems, HRM systems, financial systems, etc.
“These systems may have been running for ten or even twenty years, many are desktop clients, and may not even have an API interface. It is difficult for general-purpose agents to truly break down the data silos of the system. And this is also the biggest difficulty in whether enterprise-level agents can land on the enterprise side. In the interview, Hu Yichuan and Luo Yi both expressed this view to Jiazi Lightyear.
Content summary: Tiangong AI, mapping: Jiazi Lightyear
3. Don’t be superstitious about MCP
Chen Guo, a senior expert in China’s management consulting industry, believes that if enterprise software wants to become a “rigid need”, it must meet two conditions:Manage core business data and integrate tightly with business processes.AI agents that are not integrated with business processes are dispensable.
In terms of how to connect with various business systems to achieve enterprise-level agent landing, MCP (Model Context Protocol) is considered to be the best solution to solve the problem of agents and external tools.
By providing a unified protocol, Agents can use various tools as easily as calling functions. Many general-purpose agents have shown users the great potential of MCP.
However, the practice of first-tier manufacturers has found that MCP is by no means omnipotent in complex enterprise-level scenarios.
Luo Yi, vice president of sales and easy products, said bluntly,MCPs are effective for single-functional, standardized tool-based applications (e.g., maps, payments, etc.) because they have a limited number of APIs and clear semantics. But for a complex business system like CRM, the situation is completely different.
“I probably have thousands of APIs in my product. When these APIs are encapsulated by the MCP server, which specific API is called, it is actually semantic. Therefore, we need to organize the semantics behind the API and put it in the MCP server to have better results. But this is actually a difficult task. Luo Yi revealed.
In short, the agent needs to know which API to call in what business scenario, for what purpose, and which API. This requires deep business knowledge, not just what a technical protocol can provide.
At the same time, the complexity of enterprise business further amplifies the limitations of MCP.
For example, transaction consistency and state management. Many operations in the enterprise are transactional and require multiple steps to succeed before they can be submitted, otherwise they need to be rolled back. For example, a complete order process may involve multiple API calls such as inventory checks, order creation, payment processing, logistics notifications, etc. However, MCP itself does not provide a transaction management mechanism. If an agent fails in the middle of a multi-step process, how to ensure data consistency is a huge technical challenge.
Another example is the permission and security issues that enterprise-level products will pay attention to. Enterprise systems have permission controls down to the field level. The MCP protocol itself does not contain complex permission management logic. When an Agent calls the API through MCP, how do I verify its permissions? How can I ensure it doesn’t access data beyond my authority? All of this requires a complex security and authentication system outside of the MCP.
There is also the issue of traceability and interpretability of decision-making. In high-compliance industries such as finance and healthcare, every decision of AI must be traceable and explainable. When an agent calls a series of tools through MCP to finally give a recommendation, the enterprise needs to be able to clearly audit the entire decision chain.
In the face of the limitations of MCP, toB manufacturers are also exploring and breaking the game according to their own advantages.
The first way is for business system vendors to open native interfaces and even release their own MCPs. This was a common reply received by Jiazi Lightyear when it visited manufacturers such as Sales Easy and Laiya.
When application vendors actively embrace the agent ecosystem and encapsulate their core capabilities into easy-to-call interfaces, the integration efficiency and reliability of agents will be greatly improved. More importantly, MCP can obviously bring new traffic to toB products in the AI era.
The second way is to use RPA as a “universal glue” to combine with the agent without a native interface. This is the advantage of typical RPA vendors. Through RPA, agents seem to have acquired “hands” and can perform stable and reliable operations on business systems.
Needless to say, MCP technology paints a blueprint for interconnection for the enterprise-level agent ecosystem, but it is not a cure. How to make agents play a role in business scenarios smoothly also needs to be explored by enterprises based on their own circumstances. The business of enterprise-level agents itself must also return to a deep understanding of the business and the ultimate pursuit of technical reliability.
4. Business models that are not easy to change
Technical feasibility is only the starting point, and whether a business can be established ultimately depends on whether its business model can be accepted by the market and generate sustainable profits.
The SaaS model has dominated the enterprise software market over the past decade. But the fundamental flaw of the SaaS model is that it sells “tools”, not “results”. Essentially, customers don’t buy the software itself, but expect it to solve real problems. When the economic environment tightens and the pursuit of ROI (return on investment) becomes more demanding, this “pay for tools” model seems inadequate.
And enterprise agents are changing all that.
Earlier this year, Sequoia Capital proposed in a closed-door meeting that enterprise-level software in the AI era will evolve from delivery tools to delivery results. “Jiazi Lightyear” found in the interview that this trend has been confirmed.
Hu Yichuan, co-founder and CTO of Laiye Technology, bluntly said that the process “has already begun”. For example, the key indicators that report to customers are “how many new leads have been mined” and “what is the email response rate”, which are clear results.
There are also intelligent customer service agents of Easy Sales, training agents of Beisen, etc., and the deliverables have changed from software functions to real work results.
The underlying logic of this transformation is that Agents can transform the “process” that was difficult to quantify in the past and relied on human experience and operations into measurable and deliverable “results”. This will undoubtedly be a huge change for the domestic market, which is generally not willing to pay.
However, it should be pointed out that the current business model shift from “delivering results” to “paying for results” is not easy.
For example, an agent can provide hundreds of leads, but how do you price these leads? “We can’t determine the value of each lead yet, so it’s hard to assess how much customers need to pay.” Hu Yichuan said bluntly.
Luo Yi, vice president of sales and easy products, also expressed a similar view on how to measure the value of results. “If there is a clear quantitative standard, asset valuation will be relatively easy.”
Behind this is a complex value evaluation system problem. The value of a sales lead depends not only on the lead itself, but also on various factors such as subsequent conversion rate and customer unit price. The value of a resume is also closely related to the importance of the position and the urgency of recruitment. Until these values are difficult to quantify accurately and fairly, pay-for-results business models are difficult to implement at scale.
According to the observation of “Jiazi Lightyear”, most toB manufacturers are currently adopting a transitional pricing method, some are priced according to the original SaaS model, and some will charge additional fees to agents, which has not completely separated from the charging framework of traditional software.
5. The “marathon” of enterprise-level agents has just begun
Back to the original question:Is an enterprise agent a good business?
The answer is yes. Several industry reports have painted an extremely optimistic growth curve for the AI Agent market.
According to a previous research report by the Head Leopard Research Institute, the market size of China’s Al Agent market will be 55.4 billion yuan in 2023 and is expected to reach 852 billion yuan by 2028, with an average annual compound growth rate of 72.7%. Among these 852 billion yuan, according to the data forecast of Guanyan Tianxia, the market size of B-end AI agents is expected to be 839 billion yuan in 2028, accounting for 98.5%.
Behind the astonishing numbers is the continuous investment in digital transformation and the urgent need for “new quality productivity”.
But this business is not easy to do. It’s not a short-term tuyere chase, but a marathon that tests technical depth, scenario insight, and business patience.
The essence of toB AI Agent is to use AI to reconstruct the production relationship of enterprises. It is not a patch of existing software, but a reengineering of business processes; Instead of more convenient tools, it delivers a more efficient workforce.
It is not a track that can be easily succeeded by relying on outlets and concepts. It requires practitioners to abandon the traffic thinking and entertainment mentality of the ToC market and truly sink into the “deep water area of business” of enterprises. It should deeply combine the general intelligence of large models with industry-specific knowledge, complex business processes, and strict security compliance requirements.
From the current point of view, everyone is “on the way”. For those players who are patient, wise, and have a deep understanding of the needs of enterprises, enterprise-level agents are undoubtedly a good business worth investing in for a long time. It’s not just about the success of toB business, it’s about shaping the future of work and productivity paradigms.
This change has just begun.
