The End of SaaS, or Its Evolution? — Strategies SaaS Companies Must Adopt in the Age of AI Agents

Before SaaS was proposed, software was purchased as packaged products. Individuals installed it on their PCs, and companies installed it on servers in their data centers. In other words, software usage was based on on-premises environments.

With the release of Windows 95 in 1995, the internet began to spread widely. Both individuals and companies started using content and services on the internet, and the late 1990s were marked by the enthusiasm of the dot-com boom.

At that time, all companies providing software assumed it would be used on-premises. However, as communication infrastructure developed and the internet spread rapidly, companies such as Amazon and eBay, which operated consumer-oriented businesses, began disrupting industries such as local bookstores and small retail shops with new internet-based business models. In contrast, nothing changed in the world of enterprise software.

Marc Benioff, the founder of Salesforce, believed that if software were provided not as a physical package but as a service delivered through the internet, large-scale servers would no longer be necessary, expensive upfront investments would disappear, and users could access the functions they needed through a monthly subscription fee.

In that case, software installation and configuration would no longer be required, and users could always access the latest features from anywhere. From the vendor's perspective, there would be no need to burn software onto CD-ROMs for distribution, and there would be no need to distribute new CD-ROMs each time the software was upgraded, significantly reducing costs.

Marc Benioff, who held this view, introduced the provocative catchphrase "No Software." With this message, he brought about a change in the way vendors deliver software, shifting the payment model from one-time purchases to monthly or annual subscription fees. This also led to a financial shift in which software was no longer treated as a capital asset but rather recorded as an expense, much like electricity or water.

Looking back now, this was a major turning point.

While it enabled anyone to use software that had previously been expensive at a much lower cost—something that could be described as the "democratization of software"—the technological foundation that made SaaS possible was the multi-tenant architecture, a design approach in which a single system infrastructure is shared by multiple customer organizations. For SaaS vendors, this dramatically reduced infrastructure operation costs, enabling them to offer services at even lower prices. As a result, multi-tenant architecture created a situation that benefited both vendors and customers.

On the other hand, many companies were concerned about storing their corporate data on the other side of the internet (at the time, the term "cloud" was not yet commonly used). They also wondered what would happen if internet communication stopped and services could no longer be accessed. There were questions such as: What would happen if the SaaS vendor went bankrupt? These concerns created resistance to relying on systems that companies could not fully control themselves.

Salesforce's CRM application (service) did not spread quickly at first. However, what Salesforce did was not simply present slides—it demonstrated an actual working CRM interface. Even companies that were skeptical for various reasons began adopting it, not through full-scale company-wide deployments, but through partial adoption or through formats such as free trials. Not only Salesforce but many other SaaS vendors were gradually accepted by enterprises using similar approaches.

Salesforce also held an event called "End of Software." By inviting industry stakeholders, analysts, media, and customers, the company expanded awareness of the SaaS category and its potential. In addition, when the dot-com bubble burst in 2000 and the market cooled, it was generally seen as a headwind for the broader market. Ironically, however, it became a tailwind for Salesforce and other SaaS vendors. While traditional software purchasing methods were expensive, the subscription model offered a lower-cost entry point, which accelerated the shift from conventional on-premises software purchases.

Since then, SaaS has gained recognition in the market, and services that were previously delivered as on-premises software packages have increasingly been provided through the internet.

In September 2025, Fuji Chimera Research Institute announced a summary of the report "Software Business New Market 2025 Edition," which analyzes the market size of 53 types of software provided to corporate users, including back-office solutions. The report forecasts that in fiscal year 2025, the SaaS/PaaS delivery model will grow by more than 10% compared to the previous year, and the market size is expected to exceed ¥3 trillion.

Software used by enterprises has shifted from purchasing packaged software for on-premises use to accessing software through the internet. As a result, over the past 20 years, SaaS vendors have been the central players in the IT industry, forming the foundation that supports corporate operations. SaaS has been an extremely convenient service, and by utilizing it, companies have been able to improve the efficiency of their work.

Just as SaaS pushed packaged software and on-premises usage methods out of the market, will AI agents push SaaS out of the market?

In conclusion, the risk is quite high.

In fact, the capital market has already begun to factor in this change. As of January 30, 2026, the stock prices of SaaS companies such as SAP and Adobe had declined by approximately 20% over the past six months. Meanwhile, Tesla, which is actively investing in AI, has risen by 32%, and Google (Alphabet) has increased by 73%. In addition, unlisted AI companies such as OpenAI and Anthropic have been valued at several trillion yen. The movement of capital clearly suggests the end of the SaaS era and the arrival of the AI era.

What Is an AI Agent? (User-Side Behavior in SaaS)

SaaS was a major turning point in the software business. However, from the perspective of users, there was actually no major change in usability between on-premises software and SaaS.

Certainly, there were certain operations and user interfaces that were unique to on-premises packaged software. However, today's SaaS applications can perform almost the same functions as on-premises packaged software. As a result, when users perform tasks such as expense reimbursement, CRM, SFA (Sales Force Automation), MA (Marketing Automation), or document creation, they open the SaaS interface, enter text, select items from pull-down menus, and press buttons.

As many of you know, most of what can be done in Microsoft Excel can also be done in Google Sheets.

However, both Microsoft Excel and Google Sheets still require human users to perform the operations. With AI agents, however, humans no longer need to perform those operations themselves.

Here, we will briefly explain the difference between generative AI, which many people are already familiar with, and AI agents.

Both generative AI and AI agents are based on the latest AI technologies, but there are significant differences in their roles and how they operate. Generative AI functions as an excellent creator or advisor. It generates content—such as text, slides, images, videos, or software code—according to the instructions given. However, it does not act autonomously and responds only when the user provides a request.

In contrast, an AI agent acts as an autonomous executor. It performs tasks in order to achieve a specific goal, such as searching, making reservations, or aggregating data. Because it operates autonomously, it can plan its own actions in order to achieve the goal and continue the process until the task is completed. In simple terms, generative AI only performs tasks that are explicitly instructed, while AI agents can produce results even from vague instructions, much like a capable human.

For example, if you give an instruction in natural language such as: "Based on the receipt images in this folder, register the expenses, create a departmental expense report for this quarter, attach the report to an email and send it to the department heads whose departments have exceeded the expense budget requesting countermeasures. If the department heads reply, evaluate their countermeasures and let me know."

The AI can access internal databases, data stored on your device, and the email server to check the necessary data and files, and complete the task within minutes.

This document does not go into detail about which specific AI agents would perform which actions in this process. However, if you consider the number of steps involved when giving such instructions to your assistant or to each department head, the process would normally be quite extensive. Until now, humans have handled these tasks using SaaS. AI is now evolving from a supporting role, like generative AI, into an executor in the form of AI agents.

UI-Centered Work Will Disappear

Executives, product marketing leaders, business leaders, and technology leaders may wonder: if tasks can be handled through natural language in this way, are the user interfaces (UI) and user experiences (UX)—which have been carefully developed to improve usability for users—still necessary? Based on our experience, we believe their necessity will rapidly diminish.

Until now, systems were designed on the assumption that humans would operate them. As a result, companies have invested significant effort in creating buttons that are easy to press, menus that are easy to view, intuitive operations, and input designs that are easy to use. However, if the assumption that humans will operate the system disappears, these efforts will no longer be necessary. For example, the interface of traditional screens and screens designed for AI agents will change as follows.

If humans no longer operate the system, visually appealing interfaces and easy-to-use button layouts will lose their value. Companies have paid for SaaS in order to improve employee productivity through the use of these services. However, if AI can easily produce the desired results, there will no longer be a need to pay for SaaS vendors or to train employees on how to use those systems.

The same can be said from the perspective of SaaS vendors. If it becomes common for customers to give instructions through natural language, the investment that has been made in visually appealing UI/UX will no longer be necessary. Vendors will only need to create the "interface designed for AI agents" mentioned above and equip it with AI capabilities.

When dealing with a single SaaS application, the interface may look as described above. However, in the future, screens for multiple systems or multiple SaaS applications will no longer be necessary. For example, it will look something like the following.

If work can proceed simply by accessing all internal systems or databases and giving instructions through a natural language interface, then people will no longer need to learn system menus or commands. Instead, systems will learn to understand human language.

To use another example, you might instruct an AI agent in the same way you would instruct a subordinate: "Please plan and execute next month's promotional campaign."

What would the AI agent actually do? It can be imagined that it would carry out processes such as the following:

1. Analyze data from past promotional campaign plans and their execution results

2. Focus the analysis on similar economic environments, similar customer segments, and similar promotional campaigns

3. Select the optimal target for the current campaign and extract the relevant customers from the customer database

4. Design the overall campaign plan, including the catch copy, lead copy, text, campaign content, and campaign period

5. A design AI agent then creates advertising banners and email newsletter content based on the overall plan

6. Prepare scheduled social media posts and email newsletter distribution in several stages

7. Enter the campaign budget and required revenue into the marketing budget management system and set the target ROI

At this point, you receive a report saying, "The promotional campaign is ready."

You then review the outputs from the various steps—such as the past analysis, overall campaign design, and advertisement design—and approve them if they are satisfactory.

In the processes from steps 1 through 7, the person who gave the instructions did not use any SaaS tools. Instead, an AI agent acting as a capable subordinate completed the work by collaborating with other AI agents.

This is only a prediction, but many companies will likely develop their own AI agents. Rather than thinking in terms of using SaaS or systems, companies will develop and utilize highly capable digital teams composed of AI agents.

Even highly capable employees may be able to handle various tasks flexibly, but in reality it is impossible for one person to be a professional in all areas such as sales, engineering, marketing, business development, branding, finance, and accounting. Therefore, work is organized and carried out as a team.

The same will likely happen with AI agents. Initially, AI agents will specialize in specific tasks. Over time, however, specialized AI agents will begin to collaborate autonomously. This will be similar to having a team composed entirely of geniuses in each area of work.

The Collapse of the Concept of Applications

An application is something developed for users to perform specific tasks.

This may seem so obvious that we rarely think about it, but when sending email we use an email application; when writing documents we use a word processor; when creating presentation slides we use presentation software; for spreadsheets we use spreadsheet software; for expense reimbursement we use expense management tools; for information sharing we use collaboration tools; and for project management we use project management applications. Each task has traditionally required launching a separate application.

However, as AI agents become widespread, even if the task a user wants to perform requires multiple applications, users may no longer be aware that they are selecting or using different applications.

For example, to carry out the earlier instruction, "Please plan and execute next month's promotional campaign," the following applications might be required: CRM, BI tools (Business Intelligence tools that support management decision-making), analytics applications, text editing software, presentation software, design tools, email newsletter tools, budget management systems, sales management systems, project management tools, and email software.

Yet to complete this instruction, the user does not need to be aware of which applications are being used. Applications will become services that operate behind AI agents and will no longer be visible to us.

This transformation is extremely significant and may be comparable in scale to the emergence of the internet.

Since the introduction of computers, humans have continuously studied how to operate them. Because we are now accustomed to it, we may not realize it, but we have spent considerable time learning how to type on keyboards, click a mouse, and more recently use gesture-based input on smartphones. In other words, we have invested a great deal of time learning how to interact with computers and devices.

However, when AI agents become mainstream, the need for such learning will decrease. People will simply give instructions in natural language—either through text or voice—just as they would when instructing another person. As a result, anyone, regardless of age or background, will be able to easily use computer resources.

AI agents will inevitably have a significant impact on the SaaS business model as well as on UI/UX. At the same time, from another perspective, this shift will increase the importance of human creativity. Rather than focusing on UI/UX design, system architecture, or database definitions, people will be able to focus directly on more fundamental questions such as "Why are we doing this?" and "What should we create?"

Does this mean a hopeless future for SaaS vendors?

We believe the answer is no. Instead, it marks the beginning of an era in which the creativity of employees working at SaaS vendors will become even more important.

Work and Organizations in the Age of AI Agents

If AI agents work autonomously and operating applications is no longer necessary, it may seem that there will be almost nothing left for humans to do. However, it is important to consider that AI agents currently have no will of their own.

If you instruct them, "Please plan and execute next month's promotional campaign," they will carry it out autonomously. If you instruct them, "Turn around the project whose progress is delayed," they will carry it out autonomously. If you instruct them, "Optimize our direct marketing," they will carry it out autonomously. If you instruct them, "Acquire new leads through social media," they will carry it out autonomously.

AI agents that operate autonomously based on such instructions can produce outputs extremely quickly and accurately. However, AI agents do not possess intent. They cannot hold an overarching intention such as "Why should this be done?"

What will be required of SaaS vendors—and more broadly of business professionals—going forward is not learning how to use generative AI. Nor is it memorizing Excel functions, writing programming code, or simply acquiring marketing knowledge.

What is required is the ability to define the question of "Why are we doing this?" and to present the overall vision and approach needed for multiple AI agents to carry out the work. In addition, humans must supervise the quality of the outputs produced by AI agents and take final responsibility for the results.

Roles That Only Humans Can Fulfill

Setting the Question: "Why Are We Doing This?"

As mentioned earlier, AI agents do not possess intent. This is because they lack subjective values. AI is ultimately a software program that produces optimal results based on program code and data. It does not possess fundamental human desires such as "because I am hungry" or "because I want to be loved." The outputs of AI agents are responses generated from externally provided inputs, and they do not (at least for now) possess the will to independently think, "I want to accomplish this."

While AI agents can execute tasks hundreds of times more efficiently than humans, determining why something should be done is something only humans can decide. Setting a question—defining the purpose behind an action—is essentially equivalent to taking responsibility for the result. AI agents cannot take responsibility for outcomes. This point will be explained further in the section on final responsibility.

Indicating the Overall Direction and Approach

AI agents can generate a large number of ideas and materials almost instantly, and they can work continuously 24 hours a day, 365 days a year. However, selecting which outputs best fit the intended context or which ones will resonate with people depends on human intuition and experience.

Companies and businesses do not operate purely on logic. It is necessary to interpret contexts that are difficult to quantify—such as human emotions, internal organizational dynamics, the social climate, and the atmosphere of the moment—and to determine whether AI-generated proposals align with those contexts. These are advanced abilities that only humans possess. Humans must also provide feedback to AI in human language and refine its outputs accordingly.

Final Responsibility

AI agents cannot take responsibility for the results they produce. Saying "I did it because the AI agent told me to" is not acceptable in the real world. In other words, AI agents cannot ultimately step into areas that require ethical or social judgment. This is because AI agents have no physical body or emotions, and they lack a sense of personal accountability—they are simply software programs (at least for now).

When interacting with AI, it may feel as if it understands context. However, AI outputs are fundamentally based on the probability of data patterns within software models. For example, if given the word "rain," AI can instantly generate around a hundred associated words. But interpreting the layered, narrative meaning that humans experience—such as the personal memories or experiences connected to rain—is far more difficult for AI.

Humans, on the other hand, can derive multidimensional meaning from events based on their experiences, emotions, and social context. Assigning meaning is the process of interpreting events, experiences, and environments by attributing subjective value, purpose, and significance to them. The ability to respond to questions such as "Why is this necessary?" through personal judgment shaped by one's aesthetic sense or philosophy is something that only self-aware human beings possess.

Therefore, setting a question—defining what should be done and why—is essentially equivalent to taking responsibility for the outcome. For this reason, humans must ultimately bear the final responsibility.

Where Should You Start in the Age of AI Agents?

Whether you want it or not, the era in which AI agents take center stage is coming. Therefore, the most important thing is simply to start something.

If you apply your imagination, AI can be incorporated—at least to some degree—into almost every type of task or operation. For that reason, it may be a good idea to begin by using AI platforms to interact with AI and AI agents, and explore what they can do and what they might be able to do. As you continue using them, you may discover new ideas and insights.

At the same time, it is important to strengthen individual human capabilities, as well as the overall strength of organizations composed of capable people. As AI agents evolve and become more widespread, the relative value of human capabilities will increase. The distinction between what AI can do or understand and what only humans can do will become clearer. For example, a deeper understanding of physical sensations and the five senses will become important, and qualities such as empathy, intuition, and sensitivity will become powerful capabilities.

Based on our own experience of transformation, our CEO has identified three key conditions required for companies to transition into AI-native organizations.

1. Leadership Must Drive the Transformation

In the era of big data and cloud computing, technological upgrades could be achieved by hiring specialists. However, adopting AI is an initiative that transforms organizational culture and the business model itself. We believe that the CEO and executive leadership must personally experience AI, understand both its potential and its limitations, and then define the direction of the transformation. A transformation of this scale cannot be achieved by leaving it solely to specialists.

2. Boldly Reexamine Existing Assumptions

To acquire something new, organizations must be prepared to reexamine their existing assumptions. Our company itself shifted direction—not by extending the security SaaS business we had built over the past seven years, but by moving toward a new direction as an AI platform business. While companies spend time carefully considering plans, AI continues to evolve and new competitors continue to emerge. From our experience, rather than spending too much time creating detailed plans, it is more effective to execute quickly on a small scale and rapidly repeat the cycle of learning from failure. This approach ultimately leads to more reliable progress.

3. Stay Alert and Actually Use the Technology

Every day, new LLM models, new AI products, and new technologies are introduced. Within our company, we operate an AI agent that collects, organizes, and translates the latest AI-related news every 30 minutes and distributes it internally through Slack. However, simply knowing information is not enough. We believe that actually using AI tools—experiencing what they can do and understanding how they might help solve problems within your organization—fundamentally improves the quality of decision-making.

QueryPie AI's Transformation — A Real Story of Overcoming the "Death of SaaS"

So far, we have discussed the impact that AI agents will have on SaaS and the changing roles expected of humans. From here, we will share the real story of how QueryPie AI itself transformed from a SaaS vendor into an AI-native company.

Our company was founded in Silicon Valley in 2017 and operated for seven years as a SaaS vendor providing enterprise security products. In the Korean market, we achieved a certain level of success, with our solutions adopted by 80% of unicorn startups. However, this business had structural limitations. Because selling security products requires dealing with conservative decision-makers, it was difficult to secure sufficient pricing relative to development costs, and from an ROI perspective, the growth ceiling of the business was becoming visible.

In November 2022, GPT-3.5 and ChatGPT were released, and just four months later GPT-4 was announced. This marked the moment when the boundaries that had existed between IaaS, PaaS, and SaaS began to collapse. Like many other companies, we initially considered enhancing the value of our existing products by adding AI capabilities. However, our CEO's decision soon shifted dramatically from that approach.

The Encounter with MCP — The Moment We Realized That "Improving Existing Products" Was Not Enough

The turning point came in November 2024, when Anthropic announced MCP (Model Context Protocol).

Before MCP appeared, LLMs were essentially "smart chatbots" that combined their trained knowledge with web search. However, with MCP, LLMs evolved into entities capable of directly connecting to internal databases, servers, infrastructure, and business applications through APIs and operating software.

Our CEO says that the moment he encountered MCP, he became convinced that improving existing products would not be enough—we would need to build an entirely new type of product. This realization became the starting point for our transformation into an AI-native company.

Building an AI-Native Development Organization

When developing the new product, the first step our CEO took was to establish a development organization optimized for AI utilization.

We formed a small, elite team centered on individuals with experience in machine learning, those capable of high-speed development, and those proactive in adopting new AI tools. In the early stage, the team consisted of about ten members.

In selecting AI tools, we also gained an important lesson from our own experience. We conducted parallel evaluations of multiple AI coding tools—such as Devin, Windsurf, and Cursor—by assigning different tools to different teams.

For example, we divided the teams so that ten people used Devin, ten used Windsurf, and ten used Cursor, enabling us to compare and understand the strengths and weaknesses of each tool.

Through this process, we discovered that tools used to build products by leveraging LLM APIs—such as Cursor and Windsurf—cannot keep pace with the development speed of tools provided by companies that develop and own their own LLMs.

Currently, all employees in our company use Claude Code by Anthropic. The decisive factor was the overwhelmingly fast pace of development in its built-in features and plugins.

One notable aspect of our approach is our policy of maintaining all AI tools through monthly subscriptions. While annual contracts could reduce costs by 10–20%, in the AI domain, "one year later" is almost an eternity. A better tool could appear tomorrow. We believe that the freedom to switch tools at any time is an asset far more valuable than the discount provided by long-term contracts.

100% AI-Generated Source Code — How the Role of Developers Has Changed

This elite team demonstrated the first prototype of an AI chat system connected with MCP in just two weeks.

Currently, across all QueryPie AI products, 100% of the source code is generated by AI. As a result, the role of developers has shifted from writing code to the following three responsibilities:

• Clarifying the specifications (Spec)

• Verifying the correctness of the test code

• Checking whether the results match expectations

Quality assurance (QA) has also changed. Instead of directly performing traditional white-box and black-box testing, the process now involves using Claude Code to review the developer's code and automating tests in integration with Playwright.

The design process has also fundamentally changed. In the past, designs were created in Figma, and front-end engineers implemented them in React. Now, branding guidelines are described in detail in a Claude.md file, and AI directly generates the UI. As a result, the role of designers has shifted from creating UI layouts to designing branding guidelines and producing icons or detailed images that are difficult for AI to generate.

As a result of this transformation, development speed has improved threefold. Tasks that previously required two to three people can now be handled by one person working together with AI.

From a Security Company to an AI Platform — A Strategic Shift

While considering the positioning of the new product, our CEO focused on the ecosystem-based business model represented by Salesforce. Instead of developing all functions in-house, the strategy was to provide a platform on which AI agents can be built, enabling customer companies to utilize it according to their own needs.

This platform implements core capabilities such as connections to multiple LLMs, MCP, Skills, and RAG.

In particular, with MCP, there is a common industry challenge: as the number of connected tools increases, it can place pressure on the LLM's context window. To address this issue, we developed our own mechanism called "Smart MCP Tool Discovery." In this approach, tools are first converted into vector embeddings, and only the necessary tools are dynamically loaded when required.

For RAG as well, performance varies significantly depending on the type of document, language, and file format. Therefore, we combine multiple models—including VLM, OCR, and AWS Bedrock—to achieve high accuracy across three languages: Japanese, English, and Korean.

For Skill execution, we dynamically generate Kubernetes (k8s) sandboxes, enabling tasks to run safely within isolated environments.

Winning Trust with a "Working Product" — Securing the First Customer

Even after developing the platform, sales were not easy at first. Building AI agents requires an understanding of LLM characteristics, data preparation, and careful consideration of security. Simply saying, "We have a platform—please use it," was not enough to change customer behavior.

The turning point came in July 2025, when we encountered a large BPO company. This company provides payroll services to numerous organizations—including global enterprises—and processes payroll for more than one million employees each month. However, it faced a challenge: "We want to transform our operations using AI, but we don't know where to start."

Our CEO personally traveled to the site and stayed there for about ten days, during which he developed a prototype. For parts that could not be implemented using our AI platform alone, we combined open-source solutions. The focus was on proving that the idea could actually work.

By demonstrating a working product, we were able to gain the customer's trust. About one month later, we signed our first contract for an AI product.

Conclusion

This white paper began with the striking topic of "the death of SaaS," and has explored the future brought about by AI agents, as well as the real story of our own transformation. The market capitalization of major U.S. SaaS companies has declined by ¥15 trillion, and this wave is certainly reaching the Japanese market as well. Even when looking at the recent stock prices of Japanese SaaS companies, a significant decline can be observed.

However, we see this not as the end for SaaS vendors, but as a major opportunity.

AI agents will certainly change the future of how computers are used and will have a significant impact on the business of SaaS vendors. However, SaaS itself once had a major impact on the traditional software package business. While the shift to SaaS led to the emergence of many new companies, there are also numerous examples of software package vendors successfully transforming their businesses into SaaS and continuing to thrive. In other words, this transformation represents a significant opportunity for SaaS vendors—an opportunity to move one step ahead of their competitors.

Innovation inevitably occurs in both business and technology. It is not always necessary to be the one who creates innovation, but it is essential to adapt to it. Rather than clinging to past business models or systems, what matters most is continuing to innovate, continuing to adapt to innovation, letting go of yesterday's assumptions, and approaching new technologies with curiosity and openness. At the very least, AI agents can become a powerful weapon for SaaS vendors.

We strongly encourage you to equip your services with this new capability—AI agents.