Written by | Hao Xin

Edited by | Wu Xianzhi

The ascendancy of Agents in 2024 stands as both a testament to technological advancement and the dawn of demystifying large models.

The aura surrounding large models is gradually fading.

"In 2023, everyone generally felt the need to buy a large model, and many were eager to train their own. However, after training, they were at a loss regarding how to use it," essentially because large models present underlying capabilities akin to water, electricity, and coal, "not as products."

An industry insider defines "using large models" not as simply opening ChatGPT or Kimi to ask questions but as AI's ability to reduce costs and increase efficiency for enterprises.

If the theme of the previous year was "large model explosion," last year's theme was "large model implementation." People gradually realized that relying solely on powerful models is far from sufficient; how to use them is the "bottleneck" issue. Agents play the intermediary role of connecting large models with scenarios, especially with AI workflows truly showing the industry the path to implementation.

Some believe that "Agents primarily solve some long-standing problems of B-end delivery," while others evaluate that "model companies selling solutions externally are opting for Agents."

Agents enable large models in the cloud to begin "connecting with reality," with "large models + Agents" becoming the new norm in To B.

Su Chenxing, Co-founder & COO of LinkAI, has a strong perception of the heat in the Agent industry. "Around June 2023, when we just entered the market, there were only three domestic enterprises making Agent tools, including us. It wasn't until the second half of last year that they began to multiply, even showing a trend of supply exceeding demand."

According to Qichacha, there are over 200 newly established enterprises related to Agents within 1-3 years.

The phrase "chaotic beginnings" aptly describes the current state of the entire Agent market. Players come from diverse backgrounds, with large companies, AI startups, and entrepreneurial companies all entering the field. Without established standards, the industry's overall span is particularly vast, prominently reflected in customer order values, ranging from a few thousand yuan to tens of millions of yuan.

"At this stage, Agents are still a pseudo-concept, an intermediate product, and true Agentic AI is full of imagination."

Supply and Demand

At the end of the year, the head of a company's 2B business prepared to set a price for its upcoming Agent product, revealing that it would be "quite cheap," directly undercutting large companies. "With algorithms at 3,500 yuan and R&D at 2,500 yuan, both with a 20% discount."

Such a situation is extremely common in the current Agent market. Unlike the "clearly marked prices" of large models, the pricing system for Agents is very disorganized. According to Su Chenxing, the customer order value of Agents ranges from a few thousand yuan to tens of millions of yuan, with the spectrum corresponding to individual developers receiving sporadic orders to large government and enterprise project orders.

Starting from the supply and demand relationship on both ends, it may be easier to understand the current development status of domestic Agents. Here, we only discuss Agent delivery on the To B side, where demand drives the rapid implementation of Agents. There are two main lines here: top-down and bottom-up.

Su Chenxing believes that completely opposite decision-making chains determine the customer order value of Agents. Top-down generally corresponds to government and enterprise customers, whose logic is driven by intelligent transformation tasks, tending to procure first and then clarify which scenarios and businesses to implement. Because they need to complete rigid indicators, their budgets are relatively loose, often in the millions, which opens up the upper limit of the Agent market.

In the bottom-up decision-making chain, the key figures are those small and medium-sized enterprises that are "first movers." They have a certain accumulation of technical and practical experience, clearly knowing the effects of Agent implementation, and deciding whether to procure and which AI products to procure to solve problems in business scenarios. Because demand comes first, with a problem-conscious mindset to purchase, the budgets of this group are generally not very high, but due to the large volume, they still support most of the Agent market share.

Demand determines supply, which can be divided into four layers based on the ability to undertake demand. At the top of the pyramid are large companies such as Baidu, Alibaba, Volcano Engine, and Tencent, with the most complete cloud and model infrastructure and the largest contract amounts; below them are vendors like iFLYTEK and Wisdom Spectrum, with relatively complete infrastructure but slightly weaker in undertaking some ultra-large government and enterprise customers; the third layer is transforming or emerging Agent companies like Lanma, Shizai Intelligence, Dify, and LinkAI, focusing on serving small and medium-sized enterprise customers; the last layer is individual developers, who can meet some simple needs using tools.

Compared to the CV era of the Four AI Tigers, large models have made To B a bit easier, with the average customer order value of projects rising from hundreds of thousands to millions. In the view of business personnel, Agents solve the pain point of difficult B-end delivery in the CV era. In the past, countless small models were stacked to serve a business scenario, but now it's "large model + small model," a combination of generalization ability and accuracy, with Agents playing a connecting role, quickly building an application for customers to use in the form of low-code projects.

However, over time, so-called "large orders" are becoming less and less common. Industry insiders have reported to Photon Planet that in the early days, Wisdom Spectrum easily reported over ten million yuan for a To B project, roughly consisting of models plus fine-tuning, but today, such large orders are no longer feasible.

The "large orders" wrapped in the guise of Agents appear to be software applications but are actually a package of solutions, including cloud services, models, hardware, tool layers, industry knowledge, and after-sales service. This means that competition in the large order market has hidden thresholds. When customers "have money to spend," those with more complete infrastructure and mature service systems are more competitive, with the final winners basically locked in among large companies.

Among the Six AI Tigers, Wisdom Spectrum is the most aggressive in the B-end, with official data revealing that the total contract value for large models in 2023 was 350 million yuan. Its Vice President Chen Xuesong has rich To B business experience, having worked at Alibaba Cloud and Megvii. According to industry insiders, Wisdom Spectrum initially decided on the B-end direction, recruiting a large number of former Megvii employees to explore the large model B-end market, continuing the "software and hardware integration" approach of the previous era.

However, inherent weaknesses prevent Wisdom Spectrum from breaking through the barriers of the second tier. Many people have told Photon Planet that "Wisdom Spectrum has plans for an IPO in the future and faces heavy commercialization pressure." The above-mentioned person revealed that Wisdom Spectrum's latest focus has shifted to information technology application innovation, actively adapting to Huawei's H920B.

"The investment cost of information technology application innovation is not low, but other AI vendors have not yet done so. Wisdom Spectrum's choice to take orders with Huawei is a viable option."

Below the Large Companies

The positioning of large models is clear, with only a handful of vendors capable of providing model capabilities. Whether players in the first or second tier, their essence is to use Agents as a lever to develop the cloud and large model markets. Agents are the appetizer, while the cloud and large models are the main course.

Leaving aside the one or two hundred companies competing, China's remaining millions of small and medium-sized enterprises have not fully seen the demand for Agent applications. Su Chenxing believes that a large amount of future demand can still support the current Agent application and service enterprises in the third tier.

LinkAI is a startup that has risen alongside Agents, starting from an open-source project that connects large models to the WeChat ecosystem for dialogue, with its products evolving from initial conversation assistants to zero-code Agent construction SaaS products including multimodal large model aggregation services, RAG knowledge bases, Chat BI databases, plugin tools, Chat Bot, and workflow construction.

According to Su Chenxing, Agent products serving small and medium-sized enterprises can be divided into two categories: pure tool-type products and products with business scenario attributes. Pure tool-type products are highly versatile, requiring relatively light team investment and generally no sales team; while products with business scenario attributes require heavy delivery and time to accumulate industry experience in a snowball manner. After weighing the options, LinkAI made a trade-off, focusing on general SaaS products and leaving a small part for customized services in marketing, e-commerce, and other scenarios.

LinkAI is a microcosm of this wave of small but beautiful Agent startups, on the one hand quickly forming commercial revenue through standardized product delivery, and on the other hand leaving room for simultaneously connecting with large companies and B-end customers.

According to LinkAI, since commercialization began in early 2024, pure SaaS subscription revenue has exceeded 2 million ARR, with 70% of revenue coming from natural conversion of open-source projects and word-of-mouth promotion of PLG; confirmed revenue from To B projects has also exceeded one million yuan, with even more orders on the horizon.

Recently, LinkAI has initiated cooperation with Baidu. There are two ways to advance this cooperation: one is to list on Baidu Cloud's application market for Baidu's customers to use and also to direct traffic to its own products for conversion; the other is to cooperate with Baidu on To B projects, providing it with Agent tool layer and cross-channel capabilities.

Even though large companies have dedicated Agent construction platforms, they still cooperate with startups. The considerations are simple. It is more efficient and time-saving to directly introduce third-party cooperation than to coordinate across departments. Agents account for only a small part of large orders, with input and output not being proportional, but startups can achieve relatively good results with tens of thousands of yuan. The flexibility of startups across ecosystems is also one of the advantages valued by large companies.

For this reason, in the minds of Agent startups, there is a natural division of business and profit with large companies. However, things do not seem to be developing on the established track, with the biggest uncertainty coming from ByteDance's Volcano Engine.

At the last conference, Volcano Engine highlighted HiAgent, a product for enterprises to develop large model applications and Agents. Its definition of Agent application construction is roughly the same as that of startups on the market, making it a direct competitor.

According to an official explanation from Zhang Xin, Vice President of Volcano Engine, "If we compare the Doubao large model to Android, then HiAgent is the SDK (Software Development Kit) for enterprises to develop applications with system capabilities."

According to insiders close to Volcano Engine, their internal assessment indicators have changed. "In addition to whether the Doubao model is being used after deployment by customers, they also look at how many Agent scenarios have been implemented."

The once-lauded Doubao has been overshadowed, failing to generate revenue on the C-end. After the team was laid off and integrated into Volcano Engine's To B business line, no significant improvement was seen. This time, HiAgent directly targets enterprise customers, and its competitiveness remains to be tested.

AI SaaS

The current market definitions of Agents are similar. For example, Volcano Engine defines expert-level Agent applications as "private data + large models + Advanced RAG + Workflow".

When discussing the differences between Agents, many mention a keyword: industry attributes. Su Chenxing gave us an example: in the e-commerce customer service scenario, customers will first receive a general-purpose product. Based on this, they will make fine-tuning for different products and one-click copy relevant industry workflows based on templates. There is also relevant training at the business level, guiding customers to write prompts, build workflows, support them in importing relevant industry data, building knowledge bases, and more.

For now, Agents are still a pseudo-concept, "with the current product form not being the final state but an intermediate state of the entire industry."

Large models have unprecedentedly expanded the capability boundaries of Agents, but from a definitional perspective, current Agents lack self-reflection and self-planning capabilities, merely executing tasks according to choreographed processes. Although it is generally believed that in the next 1-2 years, "large models + Agents" will become the mainstream paradigm, at this stage, their essence is still low-code products.

"In the past, when promoting low-code, public awareness and acceptance were not high. Now, with the promotion and popularization of large models, people's desire to explore and acceptance of Agent products have significantly increased," said Su Chenxing.

The judgment of an intermediate state also aligns with the development direction of the entire technological path. The industry is transitioning from the "AI Agent" era to the "Agentic AI" era, emphasizing the shift from executing single scenario tasks to comprehensive capabilities of autonomous planning, decision-making, and task execution.

Exploring Agents from an alternate dimension has become a recurring theme in the SaaS industry. Currently, Agents primarily adopt a subscription-based business model, alongside project-oriented systems.

Internationally, Agents have infused new energy into the financing and commercialization of the SaaS market. This welcoming and collaborative environment has lured several Chinese Agent enterprises to expand overseas. It is noted that LinkAI is currently pursuing this strategy, with plans to introduce its overseas product in a phased rollout by the end of this month.

Within China, some traditional SaaS providers are embarking on a transformation journey, integrating AI capabilities into their existing products for iterative enhancements. However, this shift to AI SaaS presents a new challenge: while AI integration increases the unit price, it also diminishes market competitiveness against "AI-native" Agent products. For products offering similar functionalities, users invariably lean towards lower pricing.

Su Chenxing remarked that for the SaaS market, AI occupies a dual existence. Scenarios like data analysis and customer service are ideally suited for large models. This segment of the market may be gradually taken over by AI, yet the remaining realm of professional, traditional SaaS retains an indispensable role.

"Large models are intangible, yet Agents offer a platform for swift trial and error. For those aspects aligning with business needs, quick introduction and transformation are essential; for those that do not fit, adhering to the original system is crucial."