"Will 'artificial intelligence' really replace 'human labor'? How much benefit can Salesforce gain from it?

As the biggest theme in the current capital and technology markets, the development of AI on the technology front is changing rapidly:

① The release of the Sore video generation model in early 2024 represents the evolution of large models from text and images to video and other multimodal forms;

② The Suno model indicates that AI has begun to show results in music and other audio generation;

③ By mid-year, large models from different platforms such as Llama-3, Gemini, Claude, and Doubao have emerged and iterated like mushrooms after rain;

④ AI, represented by Tesla FSD v13, has made significant progress in "understanding the world through vision";

⑤ ChatGPT 4o1~4o3 has shown us an AI large model with "initial results" that combines multimodal perception and logical reasoning capabilities in text, video, and voice;

⑥ Recently, the "AI agent" represented by $Salesforce(CRM.US) with the release of Agentforce has taken a step towards "artificial intelligence" truly replacing "human labor."

However, in the capital market, the upstream AI industry chain represented by Microsoft, NVIDIA, and TSMC, namely hardware chips and cloud computing infrastructure, has generally been fluctuating and flat since the second half of 2024, without continuing to reach new highs alongside the continuous evolution of AI technology. On the contrary, the downstream of the AI industry chain, namely software or SaaS service providers, has seen significant gains. As shown in the chart below, one of the largest SaaS service providers, Salesforce, has accumulated a rise of over 30% in 2H24, significantly outperforming the less than 20% cumulative rise of the aforementioned three upstream leaders. From a broader perspective, the MSCI Software & Service index has also significantly outperformed the MSCI All America Index since its low point in 2022. There are not infrequent voices in the market suggesting that the downstream software sector of AI may outperform the upstream hardware and infrastructure by 2025.

Against this backdrop, Dolphin Investment Research will focus on Salesforce (NYSE: Salesforce) as the research subject and entry point, attempting to explore the reasons and logic behind the market's bullish outlook on the software sector. One of the clearly identifiable reasons is the previously mentioned concept of "AI agent" under Agentforce. Therefore, what exactly is Agentforce, and what impact will it have on the industry and Salesforce? This will also be a topic of exploration in this article.

The following is the main content: 1. What has Salesforce been rising since September?

1. Simple Review of Stock Price

At the beginning of the article, let's briefly review the recent stock price trends of Salesforce and the possible catalytic reasons behind it. As seen in the chart below:

① The first key event, Salesforce announced Agentforce for the first time on September 12, after which Salesforce's stock price immediately broke through the months-long consolidation trend and began to rise trendily;

② On October 29, the Agentforce service was fully opened to users, and a few days later (the delay of a few days should be due to the market and customers evaluating the usability of Agentforce), Salesforce's stock price broke through again after several trading days of consolidation, rising rapidly;

③ On December 4, Salesforce disclosed its Q3 FY25 performance, and on that day, Salesforce's stock price rose over 10%. However, objectively speaking, the current performance and guidance for the next quarter are not very strong, which does not "douse cold water" on the current optimistic sentiment, but it is also not enough to validate the market's optimistic expectations for Agentforce's quarterly report. (This will be discussed in more detail later);

④ On December 17, Salesforce held the Agentforce 2.0 launch conference, introducing the company's outlook on the additional features that will be added to Agentforce (mainly an outlook, with few actual implementations), and the effects and user acceptance of Agentforce since its launch. The following day, Salesforce's stock price clearly fell, but the Nasdaq overall dropped nearly 3.6% that day, mainly dragged down by the macro market. According to our understanding, the market's reaction to Agentforce 2.0 is relatively positive.

2. Is Salesforce's Q3 FY25 performance good?

A brief review of the Q3 FY25 performance, looking at the quality of this performance from the perspectives of current results and guidance for the next quarter to see if it deserves the over 10% surge in stock price the day after the announcement.

In terms of Q3 FY25 performance, ① from the perspective of expected differences, the three key financial indicators of revenue, gross profit, and operating profit are only slightly above expectations by about 1% to 3%, which is just slightly better than expected, with no significant highlights that greatly exceed expectations.

② From the trend perspective, the revenue growth rate in Q3 (whether overall or core subscription revenue) continued to slightly slow down 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 continuously slowing revenue growth, Salesforce's improvement in profitability is relatively more noteworthy, with operating profit in Q3 FY25 increasing by 26% year-on-year, and OPM increasing by 280bps / 90bps to 20% The main contributor to the profit margin improvement is the marketing expense, which accounts for about 40% of revenue, only growing by 5%. However, as mentioned above, the extent of the OPM improvement is within market expectations.

③ This quarter, from both the perspective of expected differences and changing trends, a bright spot indicator is cRPO (current remaining performance obligation, which refers to the amount of signed contracts that have not yet been recognized as revenue) growing by 10.5% year-on-year in Q3, accelerating by about 50bps compared to the previous quarter, and exceeding the market expectation of 9%. Dolphin Investment Research speculates that the market's interpretation may be that users are highly willing to adopt Agentforce, which indeed brought in new contracts after its release, driving the acceleration of cPRO growth. This may be a reasonable explanation for why Salesforce's stock price reacted quite optimistically to the Q3 performance under the current market narrative.

However, Salesforce's guidance for FY25 and the next quarter, i.e., Q4F25, is relatively negative. ① On the growth side, total revenue and core subscription revenue growth will further slow down. ② Under Non-GAAP standards (excluding SBC), the notable improvement in operating profit margin in Q3 is expected to decline by 40bps in Q4. ③ The double-digit growth in EPS and operating cash flow in Q3 is expected to drop significantly to below 10% in Q4. ④ The biggest highlight in Q3 - the growth rate of cPRO is expected to decline from 10.5% to around 9% in Q4.

In summary, the Q3 performance, aside from the accelerated growth of cPRO, does not have significant highlights, and the guidance for Q4 shows marginal deterioration across all key indicators, which does not seem justified for the 10% increase in performance. Moreover, the recent breakthrough points in Salesforce's stock price coincide precisely with the launch and rollout of Agentforce. After this brief review, Dolphin Investment Research believes that the recent strong performance of Salesforce's stock price is not closely related to the recent fundamental performance, but mainly due to the market's optimistic expectations for the prospects of Agentforce - the earliest commercial instance under the "AI agent" concept, which has been reflected in advance.

Therefore, Dolphin Investment Research's coverage study on Salesforce does not start from the conventional perspectives of business models and barriers, but focuses on the current market's attention on Agentforce, attempting to answer what Agentforce and the so-called "AI agent" really are. Can Agentforce truly bring incremental revenue that changes the investment logic for Salesforce? From a quantitative perspective, how big is the space? The following text will revolve around these questions.

II. Agentforce -- Is it another new technology leading the future?

1. What is an AI agent?

The first question we need to understand is, what does the concept of “AI agent” under Agentforce specifically refer to? What are the essential similarities and differences compared to “Chat bot” exemplified by ChatGPT and “AI assistant” exemplified by Copilot? The following discussion may involve some “difficult to understand” concepts, but Dolphin Investment Research will try to set aside the underlying technical details and start from a perspective that is easy for us and the general public to understand, so that everyone can grasp what we are actually discussing.

In a highly summarized way, the difference of “AI Agent” compared to previous types of “Chatbot” or “AI assistant” mainly lies in the different degrees of evolution from “tool-based” to “subjective (or autonomous)” AI. According to OpenAI's vision, the development of AI technology towards true AGI (Artificial General Intelligence) can be divided into five stages. The first stage is chatbots with natural language interaction capabilities; the second stage AI has certain reasoning and problem-solving abilities; the third stage is “AI agent”, the essential difference from the second stage AI technology is that “AI agent” can not only provide solutions but also has the ability to autonomously execute those solutions.

To put it in a more colloquial and analogous way:

① Earlier versions like ChatGPT and Copilot are primarily “tool-based” AI that assist in completing certain tasks under human guidance, or “task-oriented”. Essentially, this type of AI technology is not qualitatively different from the “computers” and “Office suites” we used before; it is still just a tool.

② In contrast, “AI Agent” (in an ideally mature technological state) can “be goal-oriented.” The AI agent can autonomously collect necessary information, determine the steps needed to achieve the goal, and execute the final actions. Humans only need to set the goals or outcomes for the AI agent and provide the necessary resources and supervision. In other words, the AI agent can be likened to a “digital version” of an employee under human leadership (the “Digital labour” frequently mentioned by management), rather than just a tool.

In fact, the evolutionary path of AI agents is very similar to that of another major 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 to L5. ChatGPT and Copilot can be compared to L2 to L3 levels of autonomous driving, which can assist drivers in tasks such as lane changing and automatic braking, or achieve driving from A to B under relatively frequent human supervision. In contrast, AI agents can be compared to L4 autonomous driving, meaning they can autonomously achieve driving from A to B with little or no human intervention.

From this, we can vaguely glimpse that although the technological development paths of AI are different, they have a sense of "different paths leading to the same goal." Technologies such as large models, autonomous driving, and robotics may one day combine to create entities that possess both "intelligence" and "physical presence," potentially replacing human labor almost entirely.

2. How far are AI agents from us?

The AI agent discussed above is a concept and outlook in a mature ideal state. Whether and when an ideal state of AI agents can be realized is still an unknown question. Again, exploring the possibility and timeline of achieving AI agents from a bottom-up technological perspective is not within the capabilities of Dolphin Investment Research. We will 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 gauge how far AI agents are from reality.

As mentioned earlier, mature AI agents have the capability to independently complete information collection, analysis, decision-making, and execution. Therefore, mature AI agents need to have three major modules:

① Analysis and Decision-Making Module (Brain): Such as various AI models based on LLMs. According to Dolphin Investment Research's understanding, current large models have already developed mature natural language interaction capabilities and certain reasoning and analytical abilities. However, in terms of achieving long-chain reasoning, analysis, and judgment capabilities with a high degree of "correctness," our understanding is that current AI large models still require some time for development.

② Perception Module (Five Senses): Hardware and corresponding models that can perceive and analyze various types of information such as text, visual, and auditory. In terms of hardware, there should be no constraints on the perception side; cameras, microphones, and various sensors are already quite mature. Currently, there are already "initially effective" models that can understand multimodal information, including images, videos, and language. For example, the recently released GPT-4o multimodal model and Tesla's pure vision autonomous driving technology have both validated that current large models possess a certain ability to understand visual information As for language and text recognition technology, it has become even more mature.

③ Execution Module: Just as the core difference between L2 and L3 stages of AGI is that L3 possesses execution capabilities, Dolphin Investment Research believes that one of the main challenges for AI agents to mature and land is also in the execution module. One major issue is that, while various AI models currently have the preliminary ability to output text, PPTs, voice, and even some simple videos and 3D models, this capability, which requires calling relevant APIs for execution, is not "universal" and requires pre-embedded APIs, which are difficult to exhaust.

However, taking computer operation as an example, the current "universal" operational capabilities of AI are also under development. To illustrate, an AI with "universal operational capabilities" can obtain the required information by scanning the display screen (simulating human eyes) and can use various software by simulating mouse and keyboard operations, without relying on APIs.

④ In summary, the three major modules required for AI agents currently at least possess preliminary technical capabilities. According to Dolphin Investment Research's current understanding, the main technical challenge remains in the ability of large models to perform reliable reasoning and judgment, accurately recognize the current situation (whether physical, work-related, or interpersonal) through video, voice, and other information, as well as the final execution end.

3. What exactly is Agentforce?

The above is more of an ideal vision for AI agents, in which "digital employees" widely replace human labor, indicating that "the future has arrived." So how does Salesforce's released Agentforce actually perform? Does it possess considerable "autonomous working" capabilities as envisioned?

From a high-level overview, Agentforce is Salesforce's integration of its underlying technology platform (PaaS), years of accumulated data, and decades of industry knowledge as a leader in CRM in SaaS technology (the so-called Industry know-how) through current AI technology, molding it into various agents responsible for different tasks, helping to execute work including but not limited to sales, customer service, marketing, and data analysis.

However, from the perspective of ordinary users and investors, advanced technical capabilities and industry knowledge are more of a difficult-to-understand "black box." We can look at a more concrete example—building an agent responsible for handling expense reimbursement applications through Agentforce—to see how Agentforce actually operates:

① First step, roughly define the role of the agent, the work content it is responsible for, or the goals of the work; ② Define the various scenarios (Topics) that require the agent's intervention, such as receiving employee reimbursement applications, employee inquiries about reimbursement regulations, etc.; ③ Detailed definitions and specifications of the actions that Agents should take in different scenarios; ④ Setting when to trigger Agent intervention in the workflow, and what possible outcomes there are; ⑤ With the above settings, we have created an Agent responsible for expense reimbursement, and the last screenshot is a case of feedback from this Agent.

It can be seen that the current Agentforce is still far from the ideal state where an "AI Agent" can independently analyze and break down task objectives, make reasonable judgments and actions, and deliver expected goals. It still requires specific and accurate prior settings of roles, scenarios, behaviors, processes, etc., which may still be similar to robots operating under preset rules in the non-"AI era."

But the truly core difference is that the above settings do not involve code programming; instead, they can be described using natural language to outline the corresponding situations, rules, operations, etc. From this perspective, Agentforce can essentially be likened to a "code-free" programming tool. Although Agentforce currently still requires more and more precise "guidance" compared to humans, the potential greatest value of Agentforce is that it provides the general public (without programming skills) with a relatively simpler and more convenient way to build their own "digital assistant employees" to handle some relatively simple but complex and time-consuming tasks.

4. How is the implementation of Agentforce?

From the above example, it can be seen that Agentforce is currently still suited for relatively simple and repetitive tasks. According to the company's disclosure, the fastest implementation direction since the release of Agentforce 1.0 has been customer service (service agent). Since customer service generally does not involve decision-making and mostly involves text communication (low technical difficulty), and using robots to assist customer service had already been a common practice before the AI era, it is not surprising that service agents are the fastest implementation direction. As an example, Salesforce has also launched Agentforce on its official customer service website, the following is a communication between Dolphin Investment Research and Agentforce, which you can use to feel the advantages and disadvantages of Agentforce compared to other customer service robots or ChatGPT.

Dolphin Investment Research's subjective view is that we did not perceive a significant difference in language understanding and communication ability between Agentforce and mainstream LLMs like ChatGPT. However, in terms of word regulation, preventing "hallucinations" or "nonsense," and refusing to answer irrelevant questions, Agentforce has higher "lower limit" requirements for answer quality compared to ChatGPT and other consumer products.

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

① First, it mentioned some achievements of Agentforce since its launch. For example, the fastest deployed service agent currently handles 32,000 customer inquiries per week, with 83% of customer inquiries being independently managed by Agentforce, reducing the cases that previously required manual reporting by 50%.

② The usage scenarios supported by Agentforce will expand beyond the initial customer service and sales to more industries, more scenarios, and more roles. This includes personal shopping agents, recruitment agents, and agents assisting in healthcare, tax, education, and other areas.

③ The deployment and usage scope of Agentforce will extend to third-party platforms outside of Salesforce, such as allowing users to create Agentforce Agents that can call user data on the SAP platform or execute related ERP operational processes on the SAP platform.

④ The features mentioned in Agentforce 2.0 are scheduled to go live in February this year, and the next evolution - the Agentforce 3.0 launch event is expected to be held around May this year.

In summary of this 2.0 launch event, it can be seen that the 1.0 service agent seems to have achieved certain results. According to research from Dolphin Investment Research, users have a relatively positive view of the service agent (though the penetration rate is not high). As for the quality of the management's vision for the future development of Agentforce when it is realized, Dolphin Investment Research cannot speculate without concrete evidence before the actual product launch. However, given that Agentforce updates its iterations every 2 to 3 months, it can be almost certain that the "AI Agent" technology represented by Agentforce will likely see very rapid iterations and developments.

III. Big Dreams, How Much Contribution Can Agentforce Actually Make?

Having clarified the conceptual understanding of what Agentforce is, we will now attempt to analyze from a quantitative perspective: ① How much revenue or cost savings Agentforce may bring to users; ② How large the potential market space for Agentforce is; ③ How much net incremental revenue Agentforce is expected to bring to Salesforce in the short to medium term? 1. Taking Service Agent as an example, how large is the potential market for Agentforce?

Taking the service agent, where Agentforce has been implemented most smoothly, as an example, Salesforce currently prices service agents at $2 per conversation ( discounts may actually be offered). In comparison, according to industry research, the average cost for human customer service to respond to an inquiry (conversation) is about $2.7 to $5.6. As a cross-validation, we conducted our own calculations: ① According to research, the average annual salary for a customer service employee in the U.S. is about $35,000 to $70,000; ② Assuming an average weekly working time of 40 to 50 hours per employee (it could be higher); ③ Assuming the average time spent on each communication is 10 minutes (including idle waiting time). Based on these assumptions, the estimated cost of a single communication for human customer service is about $2.8 to $4.5, which is close to the market research data.

From this perspective, Agentforce's pricing of $2 per communication for service agents is approximately 45% lower than the average human cost. In other words, ideally, if enterprise users replace human customer service with service agents, they could save about half of their labor costs. However, it should also be considered that service agents currently do not possess the ability to fully match human customer service (as demonstrated in the demo). Therefore, we believe that the $2 pricing, which is not significantly different from the lower limit of human costs, may not have a strong appeal for enterprise users to adopt Agentforce, and there is indeed a need to offer discounts in nominal pricing.

From the above analysis, it can be seen that the "Digital labour" provided by Agentforce, in an ideal scenario ( assuming Agentforce's capabilities are close to those of human employees ), can indeed help enterprise users save a considerable amount of labor costs, thus having the potential to attract enterprise users to adopt Agentforce to replace human employees. The next question is, what is the theoretical potential market for Agentforce in the U.S. customer service sector?

According to research, there are currently about 3 million active human customer service positions in the U.S. Based on our previous calculations, a single human customer service representative handles about 13,000 customer inquiries per year. Therefore, based on conservative and optimistic scenarios for penetration rates and pricing per communication, Dolphin Investment Research estimates that Agentforce has the potential market space in the customer service market to reach between $2 billion and $39 billion. Compared to Salesforce's projected $9 billion Service Cloud revenue in FY25, a conservative scenario of a $2 billion incremental market size is not considered substantial (especially since this does not take into account that other competitors may also launch similar services) In an ideal scenario, the AI agent needs to achieve a considerable replacement rate for human labor (for example, at least 30% to 50%) and charge a higher price with capabilities close to human employees (such as $2 per communication), to hope to bring several times the current revenue scale of incremental space.

Of course, theoretically, as Agentforce expands into various industries such as sales, education, law, and finance, the theoretical total TAM of Agentforce can multiply several times to dozens of times compared to the single customer service industry, reaching hundreds of billions or even trillions of dollars in industry scale. Looking further into the long term, if "Digital labor" can indeed replace human labor in general situations, its TAM space could be said to be "all-encompassing." However, the requirements for capabilities and complexity in the aforementioned jobs are significantly higher than those in customer service, and Salesforce itself has not yet figured out the pricing model for Agentforce in industries outside of customer service. Therefore, we will not "forcefully guess" the quantitative measurement of Agentforce's TAM in other industries. Simply put, the total TAM imagination space for AI agents is obviously enormous, with a trillion dollars not being the upper limit, but the premise assumption—that "artificial intelligence" can ensure quality in replacing human jobs—is still somewhat distant.

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

The above is an estimation of the potential market size of Agentforce from a medium to long-term perspective, so from a short to medium-term perspective, for example, how much impact could Agentforce have on Salesforce's performance within three years?

First, it is important to clarify that, as mentioned earlier, Agentforce adopts a pay-per-use model in customer service, rather than the traditional subscription model that charges a fixed service fee per seat. As existing users adopt Agentforce to replace the original service cloud, it will lead to a decline in original subscription revenue while generating new income.

Since the company does not disclose the number of subscription users for its services and average customer unit price data, we can only provide examples for estimation. For users of Service Cloud at different tiers and pricing, assuming Agentforce charges $1 per communication and replaces 20% of the original subscription seats, Agentforce can contribute 37% to 101% incremental revenue for users at three different pricing tiers. For the larger user segments with a significant revenue share, under the assumption of a 20% penetration rate, the incremental revenue brought by Agentforce is approximately 37% to 41%. **

Overall, assuming that mid-tier "Unlimited" level users represent the average situation of all Salesforce users, we mainly look at expectations under conservative and neutral scenarios. If we assume that within two years, the adoption rate of Agentforce in customer service reaches 5% and 10% respectively, it could bring an incremental revenue of 10% and 20% to Salesforce Service Cloud. The contribution is quite considerable. However, if Agentforce only successfully lands in Cloud service within two years, then in the above two scenarios, the contribution of Agentforce to Salesforce's total revenue would only be 3% to 6%, which is not significant.

3. Summary

Looking back at the discussion above, the concept of an "AI agent" that can think and work independently, and the future "Digital labor" that could widely replace human labor, can be said to be the most exciting and the most imaginative technological development direction since the wave of AI technology. Moreover, the various technologies required to realize the "AI agent" have indeed made preliminary achievements and are undergoing rapid iteration and development. From my personal perspective, in a longer-term view, there is a considerable probability that "AI agents" could be realized.

However, beyond the imagination, the reality is that the current Agentforce is still an auxiliary tool that requires humans to pre-specify relatively detailed rules and processes. Currently, the more significant meaning of Agentforce may be that it provides a "programming tool" for office automation that does not require coding, rather than a digital employee that can "think and work independently."

Quantitatively, if Agentforce can indeed achieve work capabilities close to that of humans, then its potential market space (TAM) is very considerable. Just in the future of the U.S. customer service industry, a 50% penetration could yield a market size of nearly $40 billion. If it can achieve impressive penetration across various industries, then a TAM of hundreds of billions or even trillions is not impossible.

However, it is clear that Agentforce currently does not possess work capabilities close to that of humans. From a more realistic and immediate perspective, taking customer service (Service Cloud) as an example, within two years, Agentforce may bring about 10% to 20% incremental revenue to Service Cloud. This is not a lot, but for Salesforce—a company where total revenue growth is less than 10%—it does have some marginal effect on improving revenue growth Therefore, as an "Agentforce" that has been released for only a few months, with its conceptual significance outweighing its actual performance, it is currently clearly insufficient to bring "x-fold" growth to Salesforce. The impact on sentiment and valuation is stronger than the impact on fundamental performance.

Of course, Salesforce has more than just Agentforce. In the next article, we will explore from more perspectives whether the current Salesforce has other highlights worth betting on.