Unlike large models, Agents present enterprises with a "turnkey solution," complete with instructions and even a calculated return on investment (ROI), which businesses highly value.

Author | Si Hang

Editor | Pi Ye

Produced by | Industry Home

"This year, we're developing large models tailored for specific scenarios, and we'll delve deeper next year. Currently, we're in discussions with cloud service partners about creating digital employees," a unicorn logistics fulfillment platform shared with Industry Home.

If AGI (Artificial General Intelligence) is the ultimate aspiration of large models, then Agent intelligent agents represent a pivotal step towards this goal. They also signify a key indicator of the "qualitative leap" in large language models. In essence, Agent intelligent agents can truly flourish only when large language models attain a certain level of maturity.

In July 2024, OpenAI outlined five levels leading to AGI (see image below), positioning Agent intelligent agents at L3. OpenAI asserted that we are on the verge of reaching L2, the level of human-like reasoning capable of solving intricate problems.

A few months later, Zhipu.AI, hailed by foreign media as "the most likely domestic contender to become the next OpenAI," unveiled the intelligent agent AutoGLM from a distinct perspective. According to Zhipu, large models have already attained L3 status, capable of utilizing tools and executing actions, albeit lacking self-learning capabilities.

However, from a market demand perspective, the shift towards enterprise-level Agents is evident. Statistics predict that the global Agent market will reach $285 billion by 2028.

For enterprises, the genuine value of AI lies in cost reduction and efficiency enhancement, which current chatbots cannot fully deliver. Thus, in 2024, the year of large model commercialization, numerous large state-owned and private enterprises embarked on upgrading their internal IT infrastructure or addressing specific challenges through the development of industry-specific large models.

Nonetheless, not all enterprises possess the resources to build industry-specific large models or engage in related development efforts. Furthermore, after a year of exploring large model commercialization, it became apparent that enterprises were still uncertain about how to effectively construct and leverage these models.

They needed a "turnkey solution."

Unlike large models, Agents provide this "turnkey solution," presented directly to enterprises, complete with instructions and a calculated ROI.

"The Agent War" has already commenced.

According to foreign media outlet Medium, by the end of 2024, there will be 500 million Agents across various industries, escalating to 50 to 100 billion by 2025.

Overseas AI companies have already joined this Agent fray.

Firstly, AI giant OpenAI, with over $13 billion in funding, has entered the Agent arena. Secondly, Anthropic, with over $7.3 billion in funding, has also joined. AI Agent companies like Adept have invested $413 million in Agent development, while Imbue has invested $220 million, and Magic AI has spent $145 million on Agent research and development.

In China, cloud providers, large model vendors, operators, software vendors, and others have commenced exploring Agents.

Since the beginning of 2024, internet giants like Baidu, Tencent, and Alibaba have launched their Agent development platforms, leveraging their respective large model platforms. These platforms offer low-code and no-code Agent development options, complemented by comprehensive computing power and model layer services.

These Agent development platforms aim to expand the model ecosystem and compete for users in the era of AI large models. However, strictly speaking, the Agents built through these platforms do not align with the "Agents" described by OpenAI, which are more action-oriented, such as serving as "digital employees" in enterprises to achieve cost reduction and efficiency enhancement.

For truly action-oriented Agents, their application is currently confined to large enterprises. According to previous Industry Home statistics on large model bidding projects, Agents emerged as a trend by 2024, with telecommunications operators ranking among the top three in procurements, primarily for intelligent customer service Agents.

Apart from cloud providers and large model vendors, some software vendors are also endeavoring to create Agents through a SaaS+AI approach.

A notable success story is Salesforce, the overseas SaaS giant, which introduced SDR (Sales Development Representatives) and Einstein Bot. These tools aid enterprises in screening sales leads, scheduling meetings, and providing video avatars resembling potential customers to assist salespeople in rehearsing pitches through role-playing.

Why will Agents become the primary narrative in 2025?

As large model technology, products, and commercialization mature, customers are increasingly concerned with who can provide a standard answer and seamlessly address their pain points, rather than large model rankings, new technologies, or model architectures. The answer undoubtedly lies in Agents.

From Technology to Implementation: Agents as the First Step in AI

According to 2024 large model bidding projects, winning bids primarily fall into three categories: computing power, industry-specific large models, and Agents.

Typically, only governments, large state-owned enterprises, or industries with significant GPU consumption, such as autonomous driving companies and operators, have the demand to purchase computing power.

For industry-specific large models, buyers are usually large enterprises. Developing these models necessitates robust IT expertise and involves reorganizing enterprise knowledge and bridging barriers between various IT systems, increasing development complexity.

Moreover, from the past year's exploration of large model commercialization, it is evident that enterprises are still unclear about how to construct and effectively utilize these models. Consequently, industry-specific large models may not always be the optimal choice under specific circumstances.

Agents, however, present a distinct advantage. They offer AI solutions tailored to specific scenarios in the era of large models, akin to a "turnkey solution."

For instance, intelligent customer service, one of the most widely utilized applications, demonstrates significant value. The head of a large customer service model project informed Industry Home that previously, the resolution rate for intelligent customer service in the industry hovered around 70%, with a 30% transfer rate to human agents. Post-adoption of large model customer service, the resolution rate surged to over 90%, saving enterprises thousands of dollars within just ten days, truly reflecting cost reduction and efficiency enhancement.

Certainly, Agents are not yet fully mature. From an industry standpoint, intelligent customer service and AI coding assistants are the most prevalent Agents. Industry Home understands that within many internet companies, intelligent customer service was the initial enterprise-level Agent project they attempted.

From a customer type perspective, large enterprises currently exhibit the strongest willingness to purchase. In the era of large models, a notable change on the demand side is the shift from CPU to GPU consumption, necessitating more resources and significant investments. Presently, only large enterprises and state-owned enterprises can afford large models and Agents.

The head of Baidu Intelligent Cloud Keyue revealed to Industry Home that over the past two years, the most prominent change in POC projects has been an increased proportion of large enterprise customers.

It's crucial to note that the inception of any novelty necessitates exploration and innovation. Similarly, in the software industry, when enterprises embark on new projects, they typically undergo a crucial and time-consuming step known as POC (Proof of Concept).

It is understood that some large companies commenced collaborating on Agent-related POC projects with state-owned enterprises as early as mid-to-late 2023. As the Agent ecosystem matures, these trends will extend to small and medium-sized enterprises in the future.

Who Can Obtain an Agent Entry Ticket?

Many Agents previously existed in SaaS form but are now becoming the preferred choice for enterprises in the era of large models.

The transition from SaaS to Agents in the era of large models also signifies a disruption in the underlying architecture. While SaaS was previously based on IaaS+PaaS, the current architecture is founded on large models, particularly the computing power layer + MaaS/model layer.

Under this disrupted architecture, not all enterprises can secure an Agent entry ticket.

Since Agents are developed based on large or small models, Agent companies must possess modeling capabilities or collaborate with large model vendors. For example, RealAI, a traditional software vendor primarily offering RPA solutions, began releasing its self-developed large models and transitioning to Agents in 2023.

Internet companies like Baidu and Tencent also rely on large model capabilities. Both have introduced their large model customer service robots, leveraging underlying model capabilities grounded in ERNIE Bot and Wenxin Yiyan, respectively, which are then fine-tuned for specific applications.

Similar to the cloud computing era, the large model era also witnesses more standardized Agent versions, existing as more standardized SaaS offerings.

The influx of these standardized versions into the market heralds the commencement of the Agent narrative in 2025.

What distinguishes Agents in the era of large models from traditional software, apart from the underlying architecture?

One significant difference lies in Agents being self-learning software. Although large models have not yet evolved to the point where Agents can fully self-learn and evolve, as described by OpenAI at L3, large model vendors can encapsulate their industry expertise into SOP processes and feed them to Agents, enabling semi-autonomous evolution.

In the future, as large model capabilities advance, Agents will attain a genuine stage of self-learning, attracting more small and medium-sized enterprises to the Agent narrative.

However, regarding enterprise-level Agents or intelligent agents, a standardized paradigm has yet to emerge, encompassing product roadmaps, internal AI construction, and business and service models.

Taking the business model as an example, in the past cloud computing era, SaaS software subscription models were primarily divided into subscription fees and customized development. However, as we transition from cloud computing to the large model era and SaaS transforms into Agents, more diversified payment models are emerging.

Currently, there are three main categories:

1) Traditional SaaS subscription-based billing

2) Token-based payment, a novel business model arising in the era of large models, where payment is on-demand and based on Agent invocation capabilities

3) Revenue sharing through ecological cooperation, such as based on sales growth, efficiency improvements, or through system integrator collaboration, integrating Agents into their products or services to achieve profit through sales sharing, cooperation promotion fees, etc.

It is evident that the narrative surrounding Agents is unfolding.