AI technology, the most prominent trend in today's capital and technology market, continues to evolve rapidly:

① In early 2024, the Sore video generation model was unveiled, marking a leap from text and image-based large models to video and other multimodal formats.

② The Suno model showcases initial success in AI-generated audio, such as music.

③ Throughout the year, large models like Llama-3, Gemini, Claude, and Doubao from various platforms emerged and iterated rapidly.

④ Tesla's FSD v13 signifies significant progress in AI's ability to 'understand the world through vision'.

⑤ ChatGPT 4.1~4.3 introduces an 'initially effective' AI large model with multimodal perception and logical reasoning capabilities, encompassing text, image, and voice.

⑥ Recently, 'AI agents' represented by Salesforce's Agentforce have taken another significant step towards truly replacing 'humans' with 'artificial intelligence'.

However, in the capital market, the upstream of the AI industry chain, dominated by Microsoft, NVIDIA, and TSMC (hardware chips and cloud computing infrastructure), has generally experienced volatility and flattened out since the second half of 2024, failing to sustain new highs despite continuous AI technological advancements. Conversely, the downstream of the AI industry chain, comprising software or SaaS service providers, has witnessed impressive gains. As illustrated below, Salesforce, one of the largest SaaS service providers, recorded an accumulated gain of over 30% in the second half of 2024, significantly outperforming the cumulative gains of less than 20% for the aforementioned three upstream leaders. Broadly speaking, the upward trajectory of the MSCI Software & Service Index since its 2022 low has also significantly outperformed the MSCI USA Index. Market voices frequently suggest that the downstream software segment of AI may outperform the upstream hardware and infrastructure in 2025.

Against this backdrop, Dolphin Research will use Salesforce (NYSE: CRM) as the research object and entry point. On the one hand, it seeks to explore the reasons and logic behind the market's optimism towards the software segment. One of the clearly identifiable reasons is Agentforce under the 'AI agent' concept mentioned earlier. Therefore, what Agentforce is and what impact it will have on the industry and Salesforce will also be discussed in this article.

Below is the main content:

I. What Has Salesforce Been Rising Since September?

1. A Brief Review of Stock Prices

Let's first briefly review the recent trend of Salesforce's stock price and the possible catalysts behind it. As seen in the figure below:

① The first key event was Salesforce's initial announcement of Agentforce on September 12th, immediately followed by a breakthrough in Salesforce's stock price, ending months of volatility and initiating an upward trend.

② On October 29th, Agentforce services were fully opened to users. A few days later (likely due to the market and customers assessing the usability of Agentforce), Salesforce's stock price once again broke through the volatility that had lasted for more than a dozen trading days, surging rapidly.

③ On December 4th, Salesforce disclosed its 3Q25 earnings, and Salesforce's stock price surged by more than 10% on the same day. However, objectively speaking, the current performance and guidance for the next quarter were not very strong. It did not dampen the current optimism but was also not sufficient to validate the market's optimistic expectations for Agentforce in the quarterly report. (More detailed discussion later).

④ On December 17th, Salesforce held the Agentforce 2.0 conference, introducing the company's outlook for additional features to be added to Agentforce in the future (primarily outlooks with relatively few actual implementations) and the effectiveness and user acceptance of Agentforce since its launch. The next day, Salesforce's stock price fell noticeably, but the Nasdaq index fell by nearly 3.6% on the same day, primarily dragged down by the macro market. According to our understanding, the market's reaction to Agentforce 2.0 was relatively positive.

2. Was Salesforce's 3Q25 Performance Good?

Briefly reviewing the 3Q25 performance, let's assess whether the quality of this performance justifies the stock price surge of over 10% the next day from the perspectives of the current quarter's performance and guidance for the next quarter.

In terms of the current quarter's performance for 3Q25, ① from an expectations perspective, the three key financial indicators of revenue, gross profit, and operating profit were only slightly higher than expectations by about 1% to 3%, just marginally better than anticipated, without any significant highlights that greatly exceeded expectations.

② From a trend perspective, the revenue growth rate in 3Q (both overall and core subscription revenue) continued to slow slightly by 0.1 percentage points compared to the previous quarter, indicating that the growth trend of Salesforce's existing business outside of Agentforce is continuing to decline.

Compared to the slowing revenue growth, Salesforce's improved profitability is relatively more noteworthy. Operating profit in 3Q25 increased by 26% year-on-year, and OPM increased by 280bps / 90bps to 20% compared to the same period and previous quarter, respectively. This was primarily due to the increase in profit margins brought about by the mere 5% growth in marketing expenses, a major expense (accounting for about 40% of revenue). However, as mentioned above, the magnitude of the OPM improvement was within market expectations.

③ The indicator with highlights from both the perspective of expectations and changing trends this quarter was cRPO (current remaining performance obligation, the amount of contracted but not yet recognized revenue), which increased by 10.5% year-on-year in 3Q, accelerating by about 50bps compared to the previous quarter and exceeding market expectations of 9%. Dolphin Research speculates that the market's interpretation of this may be that users have a high willingness to adopt Agentforce, which has indeed brought new contracts after its launch, driving the acceleration of cRPO growth. This may be a reasonable explanation for why Salesforce's stock price has been quite optimistic about 3Q performance under the current market narrative.

However, Salesforce's guidance for FY25 and the next quarter, i.e., 4Q25, is relatively negative. ① On the growth front, total revenue and core subscription revenue growth rates will further slow down. ② Under the Non-GAAP basis (excluding SBC), the notable improvement in operating profit margins in 3Q will decrease by 40bps quarter-on-quarter according to the guidance. ③ EPS and operating cash flow, which grew at double-digit rates in 3Q, will fall sharply to below 10% in 4Q. ④ The growth rate of cRPO, the biggest highlight of 3Q, will decline from 10.5% to ~9% in the fourth quarter.

In summary, there were no significant highlights in the 3Q performance apart from the accelerated growth of cRPO. The guidance for 4Q even showed marginal deterioration across all key indicators, which does not seem to justify the 10% post-earnings increase. Furthermore, the recent breakthrough points in Salesforce's stock price have precisely coincided with the launch and rollout of Agentforce. After the above brief review, Dolphin Research believes that the recent strong performance of Salesforce's stock price has little to do with recent fundamentals and is primarily due to the market's optimistic expectations and early reaction to the prospects of Agentforce, the earliest commercial example under the 'AI agent' concept.

Therefore, Dolphin Research's coverage of Salesforce this time will not start from the conventional perspectives of business models and barriers but will focus on Agentforce, which is currently of utmost concern to the market. It will attempt to answer what Agentforce and the so-called 'AI agent' are, whether Agentforce can truly bring incremental revenue to Salesforce that changes the investment logic, and how large the quantitative space is. The following text will focus on these issues.

II. Agentforce - Another New Technology Leading the Future?

1. What is an AI Agent?

The first question we need to understand is what exactly the 'AI agent' concept, to which Agentforce belongs, refers to and what are the essential similarities and differences between it and 'Chat bot' exemplified by ChatGPT and 'AI assistant' exemplified by Copilot. The following discussion will involve some 'difficult to understand' concepts. Dolphin Research will try to set aside the underlying technical details and briefly explain it from a perspective that is easy for us and the general public to understand, so that everyone can grasp what we are discussing.

In essence, the primary distinction between 'AI Agent' and previous types of 'Chatbot' or 'AI assistant' lies in the different levels of evolution of AI from 'instrumentality' to 'subjectivity (or autonomy)'. According to OpenAI's vision, the development of AI technology towards true AGI (Artificial General Intelligence) can be divided into five stages. Among them, the first stage is a chatbot with natural language interaction capabilities; the second stage of AI has certain reasoning and problem-solving abilities; the third stage is the 'AI agent', which differs from the second stage of AI technology in that 'AI agent' can not only provide solutions but also has the ability to autonomously execute them.

To put it in a more colloquial and analogical way:

① Earlier appearances such as ChatGPT and Copilot are primarily still 'tool-type' AI that assists in completing certain tasks under human guidance or is 'task-oriented'. Essentially, this type of AI technology is not qualitatively different from the 'computers' and 'Office suites' we used before and is still just a tool.

② However, an 'AI Agent' (in an ideally mature state of technology) is capable of being 'goal-oriented'. An AI agent can independently collect necessary information, determine and break down the steps required to achieve the goal, and take actions to implement them. Humans only need to set the goals or results that need to be achieved for the AI agent and provide the necessary resources and supervision. In other words, an AI agent can be analogized to a 'digital' employee under human leadership (i.e., 'Digital labor' mentioned multiple times by management), rather than just a tool.

In fact, the evolution path of AI agents compared to ChatGPT and the like is very similar to that of another mainstream application direction of AI technology - autonomous driving. As we may be more familiar with, the levels of autonomous driving technology can be divided into L1~L5. ChatGPT and Copilot can be analogized to Levels 2~3 of autonomous driving, capable of assisting drivers with lane changes, automatic braking, and other auxiliary tasks or achieving travel from A to B under relatively frequent human supervision. In contrast, AI agents can be analogized to Level 4 autonomous driving, which can autonomously travel from A to B with little or no human intervention.

From this, we can also faintly glimpse that although the technological development paths of AI are diverse, they share a sense of 'converging towards the same goal'. Large models, autonomous driving, robots, and other technologies combined together may one day truly give birth to entities that possess both 'intelligence' and 'physicality', capable of nearly completely replacing human labor.

2. How Far Are We from AI Agents? The above discussion of AI agents is a concept and outlook in a mature and ideal state. Whether and when ideal AI agents can be realized is still an unknown question. Again, discussing the possibility and timing of realizing AI agents from the perspective of underlying technology is not within Dolphin Research's capability. We will only briefly discuss the key components and technologies needed to realize mature AI agents from a perspective that the general public can understand, so that everyone can assess for themselves how far AI agents are from reality.

As mentioned earlier, a mature AI agent has the ability to independently complete information collection, analysis and decision-making, and implementation. Therefore, a mature AI agent needs to have three major modules:

① Analysis and decision-making module (brain): Such as various LLM-based AI models. To Dolphin Research's understanding, the current large models already possess mature natural language interaction capabilities and certain reasoning and analysis abilities. However, they are still some way from being able to conduct long-chain reasoning, analysis, and judgment with a high degree of 'correctness'. Based on our understanding, the current large AI models still require some time to develop.

② Perception Module (Five Senses): Hardware and corresponding models capable of perceiving and analyzing various types of information, such as text, vision, and hearing. In terms of hardware, the perception end should be unrestrained, with cameras, microphones, and various sensors already being quite mature. Currently, large multi-modal models capable of understanding images, videos, and languages have already achieved "initial success". For example, the recently released GPT-4o multi-modal model and Tesla's pure vision autonomous driving technology have both verified that current large models already possess a certain ability to understand visual information. As for language and text recognition technology, it is even more mature.

③ Execution Module: The pivotal distinction between the L2 and L3 stages of AGI lies in the execution capabilities inherent to L3. Dolphin Investment Research highlights that one of the primary challenges in the effective implementation of AI agents pertains to the execution module. A notable issue is that, while various AI models currently possess rudimentary abilities to generate text, PPTs, speech, and even some simple videos and 3D models, their capacity to execute tasks by invoking relevant APIs is neither "universal" nor comprehensive, necessitating pre-embedded APIs which are challenging to fully encompass.

However, taking computer operations as an illustration, the development of AI's "universal" operational capabilities is already underway. Metaphorically, AI with "universal operational capabilities" can acquire the requisite information by scanning the display screen (simulating human eyes) and manipulate various software by mimicking mouse and keyboard usage, without reliance on APIs.

④ In summary, the three essential modules required for AI agents currently possess at least preliminary technical capabilities. According to Dolphin Investment Research's current understanding, the primary technical hurdle presently lies in the ability of large models to conduct reliable reasoning and judgment, accurately perceive the current situation (whether physical, work-related, or interpersonal) through video, speech, and other information, and in the final execution phase.

3. Who exactly is Agentforce?

The above discussion has largely focused on the vision of AI agents in an ideal state, where "digital employees" have already "arrived at the future," poised to replace human labor on a large scale. So, how does the Agentforce actually rolled out by Salesforce compare? Does it already possess considerable "autonomous working" capabilities as envisioned?

To encapsulate at a high level with the above image as a reference: Agentforce represents Salesforce's integration and fusion of its underlying technology platform (PaaS), years of accumulated data (Data), and decades of expertise in SaaS technology and industry knowledge as a CRM leader (so-called Industry know-how), combined with current AI technology, to create various Agents capable of handling diverse tasks, assisting in work that includes but is not limited to sales, customer service, marketing, and data analysis.

However, from the perspective of ordinary users and investors, sophisticated technical capabilities and industry knowledge often remain a "black box" that is difficult to comprehend. We can gain insight into how Agentforce operates by examining a more concrete example - constructing an Agent responsible for managing expense reimbursement applications through Agentforce:

① Initially, define the Agent's role, work content, or objectives; ② Specify the various scenarios (Topics) where the Agent needs to intervene, such as receiving employee reimbursement applications or employees inquiring about reimbursement rules and regulations; ③ Define and standardize in detail the actions (Actions) the Agent should take in different scenarios; ④ Establish when to trigger the Agent's intervention and execution in the workflow, along with possible processing outcomes; ⑤ After the above settings, we obtain an Agent dedicated to expense reimbursement, and the final screenshot depicts an example of feedback from this Agent.

It is evident that the current Agentforce is still far from the ideal state where "AI Agents" can independently analyze and decompose task goals, make rational judgments and operations, and achieve the desired objectives. It still necessitates specific and precise settings for roles, scenarios, actions, processes, etc., which may still resemble robots operating under predefined rules in the pre-AI era.

However, the core difference lies in the fact that this setting process does not involve code programming but rather uses natural language to describe corresponding situations, rules, operations, etc. From this perspective, Agentforce can essentially be compared to a "code-free" programming tool. While Agentforce currently clearly still requires more and more precise "guidance" compared to humans, its potential greatest value lies in enabling the general public (without programming abilities) to construct their own "digital assistant employees" relatively simply and conveniently to handle some relatively straightforward yet tedious and time-consuming tasks.

4. How is Agentforce being implemented?

Based on the above example, it is apparent that Agentforce is currently suited for relatively simple and repetitive tasks. According to the company's disclosure, since the release of Agentforce 1.0, the fastest-growing segment has been customer service (service agent). Given that customer service generally does not involve decision-making and mostly only involves text communication (low technical difficulty), and the use of robots to assist customer service was already commonplace before the AI era, it is not surprising that the service agent is the fastest-growing segment. As an example, Salesforce has also launched Agentforce for customer service on its official website. Below is a conversation between Dolphin Investment Research and Agentforce, which you can use to personally assess the strengths and weaknesses of Agentforce compared to other customer service robots or ChatGPT.

In Dolphin Investment Research's subjective opinion, we did not observe a significant difference in Agentforce's language understanding and communication capabilities compared to mainstream LLM large models like ChatGPT. However, in terms of wording norms, preventing "hallucinations" or "nonsense," and refusing to answer irrelevant questions, Agentforce has higher standards for the "lower limit" of answer quality compared to C-end products like ChatGPT.

On December 17, 2024, Salesforce held a presentation for Agentforce 2.0, with the main information including:

① Firstly, some achievements since the release of Agentforce were highlighted. For instance, the fastest-growing service agent currently handles 32,000 customer inquiries per week, of which 83% can be independently managed by Agentforce, reducing cases that previously required escalation for manual handling by 50%.

② The usage scenarios supported by Agentforce will expand from the initially launched customer service and sales to more industries, scenarios, and roles. Such as personal shopping agents, human resource agents responsible for recruitment, agents assisting in healthcare, tax payment, education, and more.

③ The deployment and usage scope of Agentforce will be extended to third-party platforms outside of Salesforce, enabling users to create Agentforce Agents that can access data on the SAP platform or execute related ERP operation processes on the SAP platform.

④ The aforementioned features for Agentforce 2.0 are scheduled to be launched in February this year, and the next evolution - the Agentforce 3.0 press conference - is expected to be held around May this year.

Summarizing this 2.0 press conference, it is evident that the 1.0 service agent has achieved certain results. Combined with Dolphin Investment Research's understanding from surveys, users' views on the service agent are relatively positive (although the penetration rate is not high). Regarding the quality of the management's vision for the subsequent development of Agentforce upon its full implementation, Dolphin Investment Research cannot speculate before the actual product launch. However, from the evolution pace of Agentforce being updated every 2-3 months, it is almost certain that Agentforce and the "AI Agent" technology it represents will likely experience rapid iterations and developments in the future.

III. The dream is ambitious, but what is Agentforce's actual potential?

Above, we have preliminarily clarified what Agentforce is from a conceptual perspective. Next, we will attempt to analyze from a quantitative angle: ① How much revenue or cost savings Agentforce may bring to users; ② What is the potential market size for Agentforce; ③ How much net incremental revenue is Agentforce expected to bring to Salesforce in the short to medium term?

1. Taking the Service Agent as an example, what is the potential market size for Agentforce?

Taking the currently smoothest-implemented service agent of Agentforce as an example, Salesforce currently prices the service agent at $2 per conversation (actual discounts may apply). In comparison, according to industry research, the average cost for a human customer service representative to respond to an inquiry (conversation) is approximately $2.7 to $5.6. As a cross-verification, we conducted our own calculations: ① According to inquiries, the average annual salary of a customer service employee in the United States is approximately $35,000 to $70,000; ② The average weekly working hours of a single employee are 40-50 hours (possibly higher); ③ Assuming an average communication time of 10 minutes per interaction (including idle waiting time). Based on the above assumptions, Dolphin Investment Research's own calculations for the single-time communication cost of human customer service are approximately $2.8 to $4.5, which aligns with market research data.

From this perspective, Agentforce's pricing of $2 per service agent conversation is approximately 45% lower than the average labor cost. In other words, ideally, if enterprise users adopt service agents to replace human customer service representatives, they can save about half of their labor costs. However, it should also be considered that the current service agent does not yet possess the ability to fully match human customer service representatives (as demonstrated in the demo). Therefore, we believe that the $2 pricing, which is not significantly different from the lower limit of labor costs, may not be very attractive to enterprise users to adopt Agentforce, and there is indeed a need to provide discounts on the nominal pricing.

Through the above analysis, it is evident that the "Digital labor" provided by Agentforce can indeed help enterprise users save considerable labor costs under ideal conditions (assuming that the working capabilities of Agentforce are close to those of human employees), thereby potentially attracting enterprise users to adopt Agentforce to replace human employees. The next question then arises: taking American customer service as an example, what is the theoretical potential market size for Agentforce?

According to research, there are currently about 3 million human customer service positions in the United States. Referring to our previous calculations, a single human customer service representative handles 13,000 customer inquiries per year. Based on penetration rates and pricing per conversation under conservative and optimistic scenarios, Dolphin Investment Research estimates that the market size that Agentforce is expected to achieve in the customer service market could range from $2 billion to $39 billion. Compared to Salesforce's approximately $9 billion in Service cloud revenue in FY25, the $2 billion incremental market size under conservative scenarios is not substantial (not to mention competition from other potential competitors who may also launch similar services). In an ideal state, AI agents need to achieve a significant replacement rate for humans (such as at least 30-50%), and charge higher prices (such as $2 per conversation) for capabilities close to human employees, in order to bring incremental space several times the current revenue scale.

Theoretically, as Agentforce expands into various industries such as sales, education, law, and finance, the total theoretical TAM (Total Addressable Market) of Agentforce can increase several to tens of times compared to the single customer service industry, potentially reaching a scale of hundreds of billions or even trillions of dollars. In the long run, if "Digital labor" can indeed replace human labor in general situations, its TAM space is vast, potentially "all-encompassing." However, these jobs have significantly higher capability requirements and complexity compared to customer service jobs, and Salesforce itself has not yet determined the pricing method for Agentforce in industries other than customer service. We will not "forcefully guess" and quantitatively calculate the TAM size of Agentforce in other industries here. Simply put, qualitatively, the total TAM imagination space for AI agents is undoubtedly immense, with ten trillion dollars not being an upper limit, but its premise - that "artificial intelligence" can replace human work with guaranteed quality - is still somewhat distant.

2. How much incremental revenue can Agentforce bring to Salesforce?

The above calculations are based on the potential market size of Agentforce from a medium to long-term perspective. From a short to medium-term perspective, such as within 3 years, what impact could Agentforce have on Salesforce's performance?

First, it is crucial to clarify that, as mentioned earlier, Agentforce adopts a pay-per-use model for customer service rather than the traditional subscription-based model with fixed service fees per seat. As existing users adopt Agentforce to replace the original service cloud, it will generate new revenue while also leading to a decline in original subscription revenue.

Given that the company does not disclose subscription user numbers or average revenue per user for its services, we must rely on hypothetical scenarios for our calculations. For users of varying tiers and pricing within Service Cloud, let's assume Agentforce charges $1 per interaction and replaces 20% of original subscription seats. In this scenario, Agentforce could generate incremental revenue ranging from 37% to 101% for users across three different pricing tiers. For the two higher-tier users, who contribute a larger proportion of revenue and thus hold greater significance, Agentforce, under a 20% penetration rate, could yield incremental revenue of 37% to 41%.

On average, we consider mid-tier "Unlimited" users to be representative of the Salesforce user base. Under conservative and neutral scenarios, assuming a 5% and 10% adoption rate of Agentforce in customer service within two years, respectively, this could lead to 10% and 20% incremental revenue for Salesforce Service Cloud. This contribution is noteworthy. However, if Agentforce's success is limited to Cloud services within two years, its impact on Salesforce's total revenue would be minimal, ranging from 3% to 6%.

In summary, the concept of "AI agents" that can think and work independently, and the prospect of large-scale replacement of human labor by "Digital labor," represents one of the most exciting and imaginative technological advancements since the AI revolution. Various technologies required for realizing "AI agents" have already achieved preliminary success and are rapidly evolving. Personally, I believe there is a significant likelihood that "AI agents" will become a reality in the long run.

However, the current reality is that Agentforce is still an auxiliary tool that requires humans to predefine detailed rules and processes. Currently, Agentforce may be more valuable as a "codeless programming tool" for office automation rather than a digital employee capable of "independent thinking and work."

Quantitatively, if Agentforce could attain near-human working capabilities, its potential market size (TAM) would be substantial. For instance, in the U.S. customer service industry, a 50% penetration rate could yield a market size of nearly $40 billion. Across various industries, a TAM in the hundreds of billions or even trillions of dollars is plausible.

Currently, however, Agentforce does not possess near-human working capabilities. In a more realistic and immediate context, focusing on customer service (Service Cloud), Agentforce could bring about 10% to 20% incremental revenue within two years. While this may not seem substantial, for Salesforce, a company with a central growth rate of total revenue below 10%, it represents a meaningful improvement in revenue growth.

Therefore, as an "Agentforce" that has been released for only a few months and carries more conceptual significance than actual performance, it is premature to expect "x-fold" growth for Salesforce. Its impact on sentiment and valuation is currently more pronounced than on fundamental performance.

Of course, Salesforce's value extends beyond Agentforce. In our next article, we will explore from various angles whether Salesforce has other notable highlights worth considering.

- END -