Starting last week, software stocks suddenly collapsed.

The trigger came from Anthropic. The company released legal tools powered by its Claude Co-Work AI agent. This new tool can perform multiple clerical tasks, including tracking compliance matters and reviewing legal documents. These functions are at the core of many legal software products.

Impacted by this news, the combined stock market value of the software, financial services, and asset management industries evaporated by approximately $285 billion on the same day. LegalZoom.com plummeted nearly 20%, and Thomson Reuters fell by 15%.

This was not an isolated incident.

Also last week, Google DeepMind released Project Genie.

Affected by Project Genie, industry giants in the gaming sector saw their stock prices collectively plummet: Take-Two fell by 7.93%, Roblox dropped by 13.17%, and Unity, once the dominant game engine, plummeted by a staggering 24.22%.

The combined market value of the three companies shrank by approximately $19.5 billion in a single trading day!

All these fluctuations stem from a harsh judgment: AI is set to completely disrupt SaaS.

The result of this concern is that investors are voting with their feet, and software stocks are bleeding heavily. According to Goldman Sachs, software has seen the highest net selling among all industries so far this year, with the industry's net exposure (as a percentage of the total U.S. net market value) hitting a historic low of 4.2%.

To date, Goldman Sachs' software sector (GSTMTSFT) has seen $2 trillion in market value evaporate, a drop of about 30%.

So, what exactly does it mean for AI to completely disrupt SaaS? Today, Silicon-based Insider will delve into this topic.

/ 01 /

The 'Premium' of Software Is Disappearing

The notion that 'software is dead' did not emerge recently.

Before it entered the public consciousness more broadly, Microsoft CEO Satya Nadella explicitly made this judgment in an interview in January last year:

The existing form of SaaS applications or business applications is likely to collapse in the age of intelligent agents.

From today's perspective, the direct impact of AI manifests mainly in two aspects:

First, AI-native companies are reshaping old workflows. Simply put, AI companies are reconstructing existing software workflows at a lower cost, creating price and model disruptions for traditional vendors.

This is particularly evident in the customer service outsourcing industry. In July 2025, Rogers Communications announced the termination of its call center contract with customer service outsourcing company Foundever, as Rogers would shift to AI chatbots.

This decision directly affected hundreds of jobs in Canada. In August of the same year, Foundever lowered its EBITDA forecast by about 10% due to lower-than-expected new and existing business volumes, as well as increased pricing pressure in the U.S. market.

Similarly, last year, two outsourcing companies, KronosNet and Foundever, faced difficulties amid AI-related concerns, with their debt trading prices once approaching the default range.

Second, the maturity of AI programming tools has led more companies to choose in-house development, reducing reliance on expensive standardized software. Functions that originally required external software and services are now being replaced by AI and internal engineering capabilities.

Under the traditional SaaS model, software cost structures are highly dependent on R&D investment. For a SaaS company, R&D expenses typically account for 25%–40% of annual recurring revenue, with the vast majority being programmer costs.

What programmers do is simple: they translate human language into computer language (code) or turn computer language (code) into human language.

The emergence of generative AI is systematically compressing these costs. On the one hand, developers can directly describe requirements in natural language, and AI generates runnable code. On the other hand, the complexity of the traditional 'technology stack' is being weakened.

Take the financial industry as an example. Traditional high-priced software like Reuters, FactSet, and Macrobond, once the initial data access and tool setup are complete, their marginal value rapidly declines. AI can handle a significant amount of analysis, organization, and modeling work that previously required 'professional services.'

In other words, as development barriers continue to lower, the marginal cost of software is rapidly decreasing, and the value of 'artificially packaged software services' is beginning to decline rapidly.

A study by Fiona Chen and James Stratton from Harvard University provides an intriguing indirect confirmation:

The productivity gains brought by AI are primarily reflected on the software supply side (within software companies) rather than among software users.

This means that AI is compressing software production costs more quickly but is not expanding software demand proportionally.

When AI commoditizes this high-level service, it also leads to a result: the high-gross-margin characteristic of the past software industry may no longer exist.

The reason is simple: when internal AI tools can accomplish 70%–80% of the functions, why would people still pay $15,000 to $20,000 for a license?

/ 02 /

The Business Logic Has Also Changed

Compared to short-term fluctuations in profitability, a more profound change is that the way software embodies business logic is being reshaped.

In Nadella's words, the essence of most software is a 'CRUD database with business logic.'

Data is stored in a database, and software manipulates the data—creating, reading, updating, and deleting it—through pre-written rules, driving processes forward.

The core reason enterprises pay for SaaS is not just to buy a data container but to purchase an entire set of business judgments solidify (solidified) into the system.

Products like Notion are essentially structured databases that help users store and update information and present it in a more user-friendly way. In the past, how data interacted and how processes were triggered were determined by the software itself.

But in the AI era, this premise is changing.

From data acquisition to data interaction, more and more steps are being directly completed by AI, and business logic is shifting from the software application layer to AI.

In this context, enterprises no longer rely on built-in software modules to organize processes. Instead, they delegate judgment, orchestration, and execution to intelligent agents, with software gradually degrading into a 'called capability and data module.'

This change is particularly evident in record systems (System of Record, SoR) centered on data collection.

Take CRM as an example. Traditional systems rely heavily on salespeople to manually enter information, and the completeness and timeliness of the data depend heavily on human cooperation, creating strong migration barriers.

Now, this landscape is being rewritten.

First, data collection methods are changing, shifting from manual input to automatic input.

For example, new-generation AI-native CRMs like Day.ai and Attio can directly access communication processes such as emails, videos, and messages to automatically collect valuable information, changing the past landscape of manual data collection.

Similar paths are emerging in customer service, recruitment, emergency response, and other scenarios.

AI often enters with a seemingly peripheral 'wedge function,' such as automatically answering calls, organizing candidate information, or generating summaries. But over time, these systems gradually accumulate the most valuable real-time behavioral data, giving them the potential to replace the original core systems.

Second, the efficiency gains brought by AI are not just 'faster' but 'able to do more.'

In the ERP field, Everest Systems is using AI to simplify financial analysis processes, turning work that previously required financial analysts into 'automatic summarization + automatic suggestions.'

In the legal services field, Tradespace's 'AI invention harvesting' tool can automatically identify invention clues within enterprises and generate application documents, directly cut into high-value service segments previously dominated by law firms.

More importantly, while intelligent agents perform tasks, they also accumulate decision trajectories, transforming them into contextual assets that connect 'data' and 'actions.'

Unlike traditional software, which only records 'results,' intelligent agents are situated within the execution path: they pull information from multiple systems, evaluate rules, resolve conflicts, and make action decisions. In this process, all inputs, judgment bases, exceptions, and 'why this was done' are completely frozen at the 'submission moment.'

These decision trajectories gradually form a contextual graph that connects entities, events, and causal relationships within the enterprise, becoming one of the most valuable single assets in the AI era.

This trend is also reflected in many AI-native applications.

For example, AI note-taking products like Granola and Abridge are no longer satisfied with just 'recording content' but have upgraded from recording tools to knowledge and decision-making assistants by understanding language, extracting structured information, and identifying intent.

Additionally, AI customer service companies like Decagon and Sierra AI are exploring highly customized intelligent agents to have them not only handle support functions but even become part of the product experience and revenue growth.

Overall, the value of AI does not lie in replacing a specific function but in its ability to understand business context.

Unlike traditional software, which can only operate along preset rules, AI can judge intent, weigh trade-offs, and decide the next action amid dynamic information.

It is this ability to understand the business itself that allows decisions and processes to no longer have to be solidified within the software but can be elevated to the AI intelligent agent layer.

Correspondingly, systems that originally carried business logic are beginning to recede into execution and storage tools, being partially or even entirely replaced by AI in key scenarios.

/ 03 /

Software Is Becoming 'Thinner,' Systems Are Becoming 'Thicker'

While the impact of AI on SaaS has become a consensus, the market is making a typical mistake: pricing all 'software' as if it were the same business.

In fact, the differences between software companies may be greater than those between software companies and manufacturing companies.

The judgment of Sequoia partner Konstantine Buhler provides another perspective: AI may not destroy SaaS but could accelerate enterprise-level consolidation, making the moats of leading companies even stronger.

He compared Freshworks and ServiceNow. The former has long positioned itself on lower engineering costs and faster product iteration but still struggles to shake the latter's dominance in the ITSM field.

The reason is simple: in enterprise decision-making systems, cost-effectiveness is not the only core factor. The certainty brought by high-level relationships, the certification and implementation ecosystem formed around the product, and the integration inertia of legacy systems in large enterprises can all influence the final decision.

At the same time, he admitted that product-led growth (PLG) may face greater challenges in the AI era.

PLG is essentially about low-barrier trials and high usability, but AI makes building good products easier, weakening their uniqueness. In contrast, the 'human moat' of enterprise sales-oriented approaches has advantages in trust, integration, and training that are harder to replace in the short term.

With this understanding, looking back at the collective decline of software stocks, instead of repeatedly discussing 'whether AI will kill SaaS,' it is better to ask: In the AI era, which software companies will survive?

Around this question, overseas investment analyst Daniel Pronk provides a highly valuable classification framework.

Simply put, Daniel Pronk divides software companies into three categories: horizontal software, vertical software, and generative software.

Horizontal software is easy to understand—it is usable by almost all companies, with Salesforce as a representative.

Much of their value lies in 'data visualization and orchestration,' building attractive dashboards, connecting a few workflows, and handling permission collaborations.

But in the AI era, these 'surface-level tasks' that once required expensive engineers to build are rapidly being compressed into a low-value commodity. AI excels at directly taking over these surface-level interactions through natural language.

When click-based dashboards are replaced by forms, horizontal software can easily shift from a 'hub' to a 'called underlying database,' losing pricing power.

Vertical software serves specific industries or niche scenarios and is nearly useless to other industries, such as utility billing, medical department processes, or local government regulatory systems.

The moat of such software lies in its depth of industry embedding. When AI enters these systems, it is more likely to first enhance efficiency incrementally rather than overthrow the foundation directly.

So, for such companies, AI is more like a productivity tool upgrade rather than a denial of their business model.

The market may worry that 'cheaper code will lead to more competition,' but in many vertical fields, the real constraints are: the market is too small, the responsibility is too heavy, and the customers are too conservative, filtering out most lightweight competition.

The risks for generative software are the most intuitive. Their value lies in 'generation.' When underlying models become stronger, generation capabilities will diffuse into more scenarios and even become default functions built into platforms.

A typical example is Adobe. In Daniel's view, Adobe must answer a severe (severe) question:

When core capabilities are dominated and generalized by underlying models, the tool itself loses its pricing power. If models advance to the point where they can generate and edit videos with one click, why would the entry point for creation remain in Photoshop?

Duolingo faces a similar situation. Although Daniel classifies it as vertical software because it serves a clear educational scenario, its value delivery method is closer to that of generative software, continuously producing and distributing learning content.

This also determines the type of risk it faces in the AI era. Duolingo's real challenge comes from the systemic lowering of content production costs. When large models can generate structurally complete, difficulty-adjustable, and instantly feedback-providing language learning content at extremely low costs, courses themselves will no longer be scarce.

In this case, entry barriers will significantly decrease, and the number of competitors may rapidly increase. Duolingo's advantage will shift from 'content and product capabilities' to brand, distribution efficiency, and user habit stickiness. However, whether these advantages can continue to translate into payment willingness in a highly homogeneous content environment remains uncertain.

For generative software to survive, it must transform itself into a critical work system: asset management, collaboration, permissions, versioning, compliance, workflows, and industry ecosystems. Only these parts tightly bound to organizational collaboration are harder to replace with free capabilities.

Almost certainly, a trend is that software is becoming 'thinner,' while systems are becoming 'thicker.'

This round of software stock collapses does not signify the end of the software industry but rather a cruel but necessary correction. It declares the end of an era where 'stacking features could exchange for high valuations.'

In the AI era, software companies that can truly survive must answer a more fundamental question: Beyond algorithms and computing power, do you still possess something irreplaceable—an understanding of industry rules, the ability to govern complex systems, and a long-term trusting relationship with clients?

Ultimately, what the market is re-evaluating is not the concept of 'software' itself, but the value behind it:

Are you merely selling functions, or are you offering a structural relationship that cannot be easily transferred?

Text/Lin Bai