软件股「修复」迷思：反弹之后，AI Agent 到底是杀手，还是救星？

In the past two months, if there's one dominant theme in the US stock market, it's undoubtedly AI hardware, computing power, Nvidia, optical communications, electricity, and data centers.

However, another sector has also been quietly recovering — software stocks.

This isn't hindsight. In the article "Oil prices surge, interest rates resist cuts, the 'Magnificent Seven' stumble: Where to find Q2 US stock excess returns?", Maitong clearly pointed out the "valuation correction" in the software sector, emphasizing that not all SaaS is worth watching, but rather security software, leading enterprise platforms, and high-beta divergence pools. These include PANW.M, CRWD.M, NET.M, CRM.M, NOW.M, ZS.M, INTU.M, ADBE.M, MDB.M, SNOW.M, DDOG.M, TEAM.M, among others.

Looking back now, this judgment has been somewhat validated: Calculated from March 31 to May 22, the software ETF IGV rose from $80.05 to $94.01, an increase of approximately 17.4%. In other words, software stocks have experienced a clear period of low-level recovery.

But the bigger question is whether this rally is merely a technical rebound after a significant decline, or if AI Agents are causing the market to reprice some software companies.

1. Beyond the Rebound, Why Were Software Stocks Previously Shunned by the Market?

As shown below, let's first look at some data.

This data actually illustrates a very straightforward point: Software stocks aren't being shunned entirely. The market is already repricing certain sectors, especially companies in cloud monitoring, security, databases, and data clouds, whose rebound strength even significantly exceeds that of IGV itself.

But this doesn't mean "All SaaS is heading back to a bull market."

More accurately, this software rally is AI screening the winners. The market is beginning to distinguish which software will be replaced by AI Agents and which will become more important due to their implementation.

As we know, software stocks have been suppressed recently, not just because of worsening earnings, but because the market began to doubt the traditional SaaS business model.

Traditional SaaS is often charged per head, per seat. The more sales seats, customer service seats, or collaboration seats a company buys, the more subscription fees the software company receives. But with the advent of AI Agents, the market started asking a very sharp question: If one AI Agent can do the work of multiple employees in the future, will companies still need that many SaaS seats?

This is the core logic behind why software stocks had their valuations compressed.

AI Agents might automatically write emails, follow up with customers, generate contracts, analyze data, process tickets, and execute approvals. Once these tasks are automated, the traditional software logic of "more people = more seats = higher revenue" will be challenged.

Meanwhile, corporate AI budgets have recently flowed more towards GPUs, cloud computing power, data centers, and infrastructure. Software companies have been squeezed. More troubling, if AI only brings increased R&D and computing costs without improving profit margins, the valuation pressure on software companies will persist.

As we mentioned in our Q2 outlook, for enterprise software, platform-type tech, and cybersecurity companies, if AI only increases investment without improving profit margins, valuation pressure will continue to rise.

So, what the market truly dislikes about software stocks isn't just growth rates, but the certainty of the business model.

2. The Key Isn't "Having AI," But Whether AI Can Generate Revenue

The market is also realizing that AI Agents don't operate in a vacuum.

A truly implementable enterprise AI Agent needs at least four things:

• First, it needs an entry point, such as sales, customer service, marketing, IT tickets, and office collaboration – these are the real places where enterprise work happens;
• Second, it needs data. Without internal enterprise data, an AI Agent can only give generic answers and struggle to make real decisions;
• Third, it needs processes. Enterprises don't want AI to just chat; they want it to initiate approvals, update CRMs, process tickets, generate quotes, and drive business closure;
• Fourth, it needs permissions and security. In the future, it's not just humans who make mistakes; Agents might also operate incorrectly, access without authorization, or leak data. Therefore, identity, security, and auditing become even more critical;

This is why a repair logic for software stocks is emerging. AI Agents don't necessarily bypass software; they likely must operate on top of it.

In other words, while AI Agents will indeed impact parts of SaaS that "only sell seats and lack data and process barriers," for software companies that control customer entry points, enterprise data, workflows, and permission systems, it could become a new growth catalyst.

One of the most critical software earnings reports this week is from Salesforce.

Salesforce will report its FY2027 Q1 earnings after the market closes on May 27. The company has explicitly positioned itself as the "#1 AI CRM" and states that Salesforce helps enterprises become agentic enterprises, integrating humans, agents, apps, and data into a unified platform.

Market expectations currently place Salesforce's quarterly revenue at approximately $11.05 billion, a year-over-year increase of about 12%; adjusted EPS is expected to be around $3.11, up from $2.58 in the same period last year. The options market also indicates that traders expect CRM's stock price could swing nearly 9% after the earnings report.

But the real focus of this report is not whether traditional CRM revenue slightly exceeds expectations, but whether Agentforce can prove that AI Agents can be genuinely commercialized.

The market will look at several issues: Does Agentforce have real orders? Are AI features driving improvements in RPO and subscription revenue? Are customers willing to pay extra for the AI Agent, or do they just see it as an add-on feature to existing software? Can profit margins and buybacks continue to support the valuation of a mature software company?

CRM's answer is crucial. If Salesforce can prove that AI Agents enhance CRM rather than replace it, then the software stock rebound won't just be a valuation repair; it could lead to an AI-driven revaluation.

Another key earnings report is from Snowflake.

Snowflake will report its FY2027 Q1 earnings after the market closes on May 27. The company positions itself as the AI Data Cloud, emphasizing its platform helps enterprises derive value from data, applications, and AI.

The logic for SNOW.M and CRM.M is different.

CRM is more of a front-end entry point, while SNOW is more of a back-end data foundation. For an AI Agent to help an enterprise make decisions, it first needs to access, understand, and govern the company's internal data. No matter how powerful the model, it's difficult to truly implement without clean, unified, and callable data.

Before the earnings, Snowflake provided FY2027 Q1 product revenue guidance of $1.262 billion to $1.267 billion, corresponding to about 27% year-over-year growth. Zacks consensus estimates Q1 product revenue around $1.26 billion, about 26.9% growth. The market will also focus on million-dollar-plus customer counts, total customer numbers, net revenue retention rates, and the adoption of AI products like Snowflake Intelligence and Cortex Code.

So, the key for SNOW.M is not "whether it has an AI narrative," but whether that narrative can translate into more data consumption.

If Product Revenue exceeds the guidance range, customer consumption continues to recover, and RPO and large customer numbers continue to improve, it signals that enterprise AI is not just staying at the model layer but is starting to drive usage of the data platform.

Conversely, if the AI product story is exciting but consumption growth and guidance are weak, the market might interpret this round of rebound as a "valuation repair" rather than a "fundamental revaluation."

In a nutshell, CRM.M tests whether Agents can generate sales, while SNOW.M tests whether AI can drive data consumption.

3. How to Layer the Analysis of Software Stocks?

This software rebound shouldn't be simplistically viewed as all software rising together. A more reasonable approach is to layer it according to the capabilities required for AI Agent implementation.

• Layer One: Front-end Entry Points and Agent Monetization. Representative companies include CRM and NOW. They control sales, customer service, business processes, and enterprise workflow entry points. If AI Agents can genuinely integrate into workflows, these companies have the opportunity to monetize AI as a product.
• Layer Two: Data Foundations and AI Fuel. Representative companies include SNOW, MDB, and PLTR. For an AI Agent to understand an enterprise, it must tap into the company's internal data. The more complex the data, the more important governance becomes, and the higher the value of these platforms.
• Layer Three: Security, Identity, and Permissions. Representative companies include CRWD, ZS, OKTA, and NET. The more automated Agents become, the more enterprises need to manage permissions, audit actions, and prevent data leaks. Future security software won't just protect people; it must also protect the behavioral boundaries of Agents.
• Layer Four: High-Beta Divergence Pools. Examples include DDOG, TEAM, DOCU, and PATH. These companies have high elasticity but are also more dependent on earnings validation. If AI can increase usage frequency, customer stickiness, and revenue growth, they can continue to recover. If it's just a valuation bounce, their sustainability will be much weaker.

So, is the AI Agent a software killer or a software savior?

The answer is: it could be both.

For software companies lacking data moats, process barriers, and customer entry points, AI Agents might compress their value. But for platform-type software companies that control customer relationships, enterprise data, business processes, and security permissions, AI Agents could become a new growth engine.

This is also why software stocks have rebounded for five or six weeks, but the market hasn't delivered a final verdict.

The real answers will be hidden in the upcoming earnings reports: CRM needs to prove Agentforce can generate real orders; SNOW needs to show enterprise AI can drive data consumption; security software needs to demonstrate that in the age of AI automation, the demand for permissions and risk control will only strengthen; and high-beta software stocks need to prove the rebound isn't based on sentiment, but on revenue, profit margins, and guidance.

In short, true gold fears no fire. Ultimately, companies with entry points, data, processes, and permissions are the ones with the opportunity to transition from 'AI victims' to 'AI beneficiaries'.