Enterprise-level AI begins to be implemented, and whoever wins the scene wins the world

In the past two weeks, the enterprise-level intelligent agent development platforms have been in a state of rapid development.

Microsoft, Tencent, NetEase and other domestic and foreign giants have recently announced the launch of a new generation of intelligent agent development platforms. Compared with two years ago, the product logic of intelligent agent development has undergone earth-shaking changes, and the intelligent agents produced are more usable.

Why are enterprise-level intelligent agent development platforms so popular this year ?

On the one hand, the iterative upgrade of the large inference model brings better experience to the intelligent agent.

" It was often said in the past that ' each model can be the leader for hundreds of years ' , but in the field of AI , the time that large models can be the leader is very short. It can be said that ' each model can be the leader for three to five days, or even three to five hours ' . The speed of technological updates is extremely fast. " Ruan Liang, vice president of NetEase and general manager of NetEase Digital Intelligence, said with emotion.

Ruan Liang, Vice President of NetEase and General Manager of NetEase Digital Intelligence

It is precisely because of the continuous evolution of the reasoning big model that the capabilities of the intelligent agent are improved by one level every 7 months. This continuous evolution of capabilities has enabled " the intelligent agent to evolve from a tool to a partner, and can truly think while working. "

Therefore, as we enter 2025 , AI ToB manufacturers have launched or upgraded their intelligent development platforms in order to better achieve commercial results.

As a one-stop enterprise service provider under NetEase , NetEase Digital Intelligence naturally does not want to miss this wave of AI .

Since last year, NetEase has been integrating its ToB business architecture. By integrating the original NetEase Zhiqi and NetEase Computing businesses, it has upgraded to NetEase's new ToB brand NetEase Computing, and formed five major business segments including NetEase Cloud Communication, NetEase Cloud Commerce, NetEase EasyShield, NetEase Computing, and NetEase CodeWave , which report to Ruan Liang in a unified manner.

After the unified upgrade of the organizational structure, each segment of its business has also carried out iterative upgrades of the AI ​​technology of its products. The CoreAgent intelligent collaboration hub newly launched by NetEase Digital Intelligence is the result of technical collaboration among various segments.

The unification of the business architecture has enhanced the capabilities of each product and further strengthened the synergy of the organization.

Data shows that in the past three years, NetEase has invested more than 40 billion yuan in research and development, has four major AI laboratories, and has accumulated several years of experience in the field of AI .

After NetEase Digital Intelligence took on the heavy responsibility of NetEase's AI ToB and actively sought change in its tenth year, can it reach a new level?

Thanks to the advent of large inference models, the applications of intelligent agents are becoming more diverse.

From last year to now, a variety of intelligent applications have emerged, but these applications are often more oriented towards C- end consumers, while there are not many intelligent agents that are truly usable and easy to use for B- end enterprise applications.

According to Gartner data, the number of inquiries about intelligent bodies increased by 750% between Q2 and Q4 of 2024 alone , but the actual success rate of enterprise intelligent body deployment did not exceed 30% .

" The most important difference is the differentiated demands of enterprises and consumers for intelligent entities. " Ruan Liang said that, in comparison, enterprise-level intelligent entities have higher requirements for reliability, security, timeliness, and the ability to reduce costs and increase efficiency.

After all, enterprise-level applications involve multiple links, and once a problem occurs in one link, security, public opinion and other incidents may occur. Consumer-level intelligent agents are more fault-tolerant, and users allow them to make some mistakes and correct them by themselves.

Therefore, the current deployment of AI large models by enterprises is more distributed in a point-like manner, independent of the business, not integrated with the enterprise application architecture, and not entering the core business processes of the enterprise. Most of them are doing basic clerical work, and their actual value is relatively shallow.

So, how can we meet the high demand for intelligent entities in the implementation of enterprise-level AI and make enterprise-level AI truly usable and easy to use?

" From intelligent agents to enterprise-level AI applications, we need to open up the last mile of AI application implementation. " Ruan Liang said that the key lies in the " digital factory " for enterprise-level intelligent agent development , that is, the intelligent agent development platform.

The CoreAgent intelligent agent collaboration hub released by NetEase Digital Intelligence is a one-stop, implementable intelligent agent development and collaboration solution specifically provided for enterprises. It can support the development, evaluation, and hosting platforms of various forms of intelligent agents such as question and answer, workflow, and autonomous agents, and integrates enhanced knowledge base capabilities, security fence capabilities, and AI  asset sedimentation capabilities.

If the intelligent agents it provides are divided into two types:

One type is an intelligent agent that needs to be manually choreographed by programmers. It relies on workflow nodes given by humans to process tasks, such as input at the first node, judgment at the second node, execution at the third node, etc.

" This kind of solidified workflow orchestration has many benefits for enterprises. After all, enterprise business requires more certainty. " The product manager of NetEase CodeWave and CoreAgent told Guangcone Intelligence.

However, this type of intelligent agent is more dependent on programs given by humans. The pre-set workflow is like an assembly line, breaking down complex tasks into multiple simple steps. Each step has a clear goal and process, and follows certain rules to achieve specific goals.

This is also the model adopted by most current intelligent agent orchestration platforms. However, there may be bugs in the actual operation process , and the degree of implementation of some tasks cannot meet production needs. Even if there is an error in a small link, it may affect the efficiency of the entire process.

At this time, the CoreAgent intelligent agent platform can also provide a second form of intelligent agent, namely, autonomous intelligent agent. You only need to tell it what the task is, and it can independently plan what to do and how to complete the task.

" At present, well-configured autonomous intelligent agents can perform better in ' killing time ' scenarios, " said the above person in charge.

The so-called " killing time " scenario is a scenario that is tedious but has to be done, the workload is large and time-consuming, and the results are the same no matter who does it. For example, scenarios such as multi-person meeting invitations that are highly repetitive and require multiple rounds of confirmation, and HR interviews with multiple people, etc. , autonomous intelligent agents can already handle them well .

Compared with workflow-orchestrated intelligent agents, autonomous intelligent agents have stronger generalization capabilities, but they also bring more uncertainty in addition to independent thinking and autonomous planning. For enterprises , in order to better integrate them into the business and generate value, they must find a balance between the two, which is challenging.

In this regard, NetEase Digital Intelligence has provided two solutions:

First, enterprises can train autonomous intelligent agents based on their accumulated know-how , and the trained autonomous intelligent agents can complete tasks according to the ideas of business experts .

Second, the autonomous intelligent agent itself has strong autonomous learning and self-reflection capabilities. Therefore , after completing multiple tasks, with the help of manual adjustment, the autonomous intelligent agent can recognize which tasks are successful, summarize the successful experiences , and perform better the next time it performs similar work tasks.

To put it bluntly, companies can either fine-tune and train autonomous agents based on their own accumulated data knowledge, or allow autonomous agents to perform self -iteration and optimization through more task practices.

From the perspective of the entire industry, many ToB service companies are involved in the intelligent body development platform, and each company's products are showing a trend of homogenization.

So, what is the difference between NetEase Digital's CoreAgent intelligent platform and other similar products ?

On the one hand , Core Agent integrates the technical foundation accumulated by NetEase Digital Intelligence's various businesses over the past decade .

Specifically , the RAG technology used in the knowledge base is based on the algorithm technology accumulation of NetEase Cloud Commerce and Qiyu for many years. The safe and controllable operation of the autonomous intelligent body is based on the sandbox operation mechanism of the cloud native architecture of NetEase Cloud Information. The content security of the intelligent body is protected by NetEase Yidun AI content detection. If you want to connect the last mile between the intelligent body and the enterprise system, you can't do without the intelligent development capabilities of NetEase CodeWave . It can be said that Core Agent has been born with the longest board of each business since the beginning of product design.

On the other hand , in the past decade, NetEase Digital Intelligence has served more than one million external customers in various industries, and has penetrated into many bizarre and detailed scenarios; backed by the NetEase Group, NetEase Digital Intelligence has also accumulated a wealth of practical experience in intelligent agents in the process of polishing products with hundreds of millions of users within the group.

Therefore, the intelligent agents produced by Core Agent , whether they are manually choreographed agents or autonomous agents, can complete their tasks well after manual adjustment.

However, at this stage, enterprise-level intelligent agents are still in the early stages of implementation and have not yet been commercially implemented on a large scale. Truly usable and easy-to-use intelligent agents are still very scarce. So, how can enterprise-level intelligent agents really take root in enterprise business?

In addition to providing a general intelligent agent development platform that allows enterprises to deeply customize " tailor- made " intelligent agents , when faced with some highly standardized scenarios, a mature , plug-and-play intelligent agent can already become a qualified digital employee.

Starting from small scenarios and implementing them in enterprises in the form of digital employees can be said to be the optimal solution for the current application of intelligent entities .

At present, in specific and detailed scenarios such as online customer service, sales guides, and financial report generation and modification, digital employees have quietly taken up their posts without many people noticing. 

In this regard, Ruan Liang said: " AI Agent is the first digital employee of the enterprise, but if the enterprise wants to make good use of the intelligent body, it must clarify the capability boundaries of the big model, plan the capability boundaries of the digital employees, and play to their strengths and avoid their weaknesses. " 

This is like a business expert in an enterprise -  they don't have to take charge of the entire process, but they can become " irreplaceable " with their deep accumulation in a niche field. Just as financial experts focus on budget control and supply chain experts focus on inventory optimization, if intelligent agents can find precise " capability anchor points " in the enterprise business chain and become " digital experts " in a certain niche scenario  , they can also release maximum value in their areas of expertise. 

The construction of an enterprise-level knowledge base is the most critical link in achieving the goal of specialized talent.

Generally speaking, although the current large model capabilities are constantly evolving, there are still technical defects. If the scene is enlarged, it may cause hallucinations.

Take the digital employee for returns and exchanges as an example. On the surface, it is an intelligent customer service placed in front of the front desk, but it is able to connect to a series of processes from front-desk customer service to back-end product delivery, logistics transportation, etc. Once a problem occurs in any of the processes, it may cause customer dissatisfaction and even escalate into large-scale customer complaints, affecting the company's brand and corporate image.

Based on this, if you want digital employees to not make mistakes in specific scenarios, " you need to combine large and small models, new and old technologies, and most importantly, give AI the most precise knowledge range, including safety fences, etc. " Ruan Liang said.

" The digital employees released by NetEase Digital Intelligence all use the capabilities of the knowledge base . We have currently accumulated a lot of algorithms for knowledge segmentation and knowledge recall. " A relevant person in charge of NetEase Digital Intelligence said , " In the future, we will also launch a programmable knowledge base, allowing enterprises to integrate their own accumulated knowledge bases with other industry knowledge bases for visual programming , so that enterprises can freely process data resources. After all, the knowledge base is not universal. Only by flexibly adjusting the knowledge base according to the actual needs of the enterprise and combining the Know -How accumulation of its own industry can the capabilities of the intelligent body be more comprehensive and cope with more complex and changing scenarios ."

Ruan Liang also proposed another important principle for the implementation of enterprise-level AI : Don’t look for nails with a hammer, but look for AI based on needs .

To this end, NetEase Digital Intelligence has also released multiple industry intelligent entities, including digital employees for product shopping guides, digital employees for after-sales services, digital employees for private domain operations, digital employees for intelligent outbound calls, and digital employees for financial data analysis.

These digital employees are based on the needs of front-line customers and are intelligent entities formed by combining AI with NetEase Digital Intelligence's own product solutions and service capabilities. They are not imaginary.

"In fact, NetEase Digital Intelligence has held several  AI  agent creative competitions internally this year, and a large number of digital employee products for segmented scenarios have emerged. Our vision is that one day, professional agents in these small scenarios may have a clearer understanding of the specific business processes of a certain sector of the company than humans , effectively avoiding the problem of large model illusion. Once the agent's capabilities reach what we think is the qualified line, we will let the agent walk out of NetEase to help more customers solve problems. " Ruan Liang said, " In the future, corporate productivity will = (digital employees + human employees) ✖️ synergy index, and AI  agents will be the abstraction and solidification of the best time for corporate management in the AI ​​era. "

" The implementation of enterprise-level AI is like a marathon . It is a complex system project that needs to be done step by step, " said Arvind Krishna , chairman and CEO  of IBM  .

At present, for enterprises, it will become increasingly important to quickly integrate  AI  capabilities into IT  automation, business automation, and even reshape business processes.

However, there are still many difficulties for enterprises to implement AI .

The first is the coordination of intelligent agents. Although digital employees are now able to take root within enterprises, multiple AI  agents will ultimately be needed to work in series and in parallel in order to truly grow into towering trees.

In addition, the most core pain point is that " many companies currently have many old systems, which are often core business systems and make it difficult to implement AI development and transformation. " Ruan Liang said.

The next goal of NetEase Digital Intelligence is:

First, it connects and integrates the industry's big models, privately deployed intelligent entities, and enterprise-developed intelligent entities, and uses NetEase CodeWave for secondary development to help enterprises build high-quality AI application organizations and code assets.

Second, non-intrusive transformation of the enterprise's old systems. Through AI learning system essence and processes, adding AI interfaces such as floating layers and MCP interfaces , the core data, processes and functions of the old systems can be integrated into the new AI business processes, bringing new vitality.

This year marks the tenth anniversary of NetEase Digital Intelligence, and is also a key year for the transformation and upgrading of its To B business.

In the AI ​​era, NetEase Digital Intelligence is strengthening its AI ToB strategy, and its relationship with enterprises will also evolve from a " deliverer " of products and capabilities to a " value co-conspirator " in customer scenarios .

" We clearly see that the ultimate value of AI does not lie in the leading edge of a single technology, but in whether it can penetrate into business scenarios, solve real problems, and create actual value. In this process, the power of a single individual is far from enough. We need to unite more upstream and downstream partners to jointly build a healthy industrial ecological environment. In the era of artificial intelligence, there are no lone heroes, only companions. " said Ruan Liang.

Therefore, at the conference, NetEase Digital Intelligence announced the official establishment of the "AI  Future Ecosystem Alliance " , which will fully open up full-stack AI capabilities, allowing NetEase Digital Intelligence to become a " bridge " and " engine " connecting partners and empowering the ecosystem .

After all, only in the ToB track that pursues long-termism can Zhongxing go further.