When Salesforce reported a strong quarter in late 2024 and the stock still fell, analysts blamed macro. When Workday, ServiceNow, and Atlassian followed, they blamed sentiment. By the time the iShares Tech-Software ETF was down over 23% year-to-date in February 2026 — with Salesforce and Adobe each off more than 25% — it became harder to blame anything except the thing itself: the fundamental architecture of the SaaS business model may have been built on a premise that is quietly being withdrawn.

That premise is headcount growth.

The Hidden Dependency That Nobody Labeled

For two decades, enterprise software companies operated under a business model that looked like software economics but was, at its core, a bet on employment. Not their own employment — their customers’ employment. Every new hire at a Salesforce customer was a potential new seat. Every expanding sales team, every growing HR department, every company adding marketing headcount was, in aggregate, a growth engine for the entire SaaS layer above it.

Nobody called it that. The marketing language was “land and expand,” “net revenue retention,” “seat expansion.” But strip away the jargon and the structure was this: SaaS companies had externalized their growth function to the labor market. If enterprises kept hiring, SaaS companies kept growing. The two curves moved together so reliably for so long that the dependency became invisible.
What’s now becoming visible is what happens when the curves decouple.

Enterprise headcount is not growing the way it once did. A January 2026 CIO survey found IT budget growth decelerating to just 3.4% , but the more important signal isn’t the budget number — it’s where that budget is going. Funds are being diverted from application software toward AI infrastructure, which is absorbing roughly 75% of a $450 billion reallocation. Enterprises aren’t shrinking their ambitions. They’re executing them differently. Fewer humans, more capability. Which means: fewer seats.

Analysts have taken to calling this “seat compression.” A single AI agent can now perform tasks that previously required dozens of junior employees, making the incentive to maintain thousands of individual licenses for platforms like Salesforce or Adobe Creative Cloud evaporate. The math is not subtle: if you replace ten sales admins with one agent, you don’t just reduce headcount — you reduce the revenue of every piece of software those ten people were using.

Klarna as Case Study and Cautionary Tale

In 2024, Klarna’s CEO Sebastian Siemiatkowski made waves by announcing on an investor call that the company was shutting down Salesforce and Workday. The story that traveled through the tech press was that Klarna had replaced enterprise software with AI. The truth was more instructive than the myth.

Klarna confirmed to reporters that it had stopped using Workday and Salesforce’s CRM solutions — but instead of replacing them with AI, it was using Deel as its HR platform and pulling together other SaaS tools. The real story was that Klarna consolidated 1,200 SaaS applications into a unified in-house knowledge stack built on Neo4j, a graph database, as the basis for a custom AI layer. The saving wasn’t about replacing software with AI. It was about eliminating redundancy — 1,200 tools down to a coherent, AI-friendly architecture.

Siemiatkowski was later “tremendously embarrassed” by the public fallout from his comments, and by mid-2025 Klarna was quietly hiring humans back after discovering that the promise of faster resolution times and higher customer satisfaction through AI alone had fallen short.

Klarna didn’t replace SaaS. Klarna de-bloated SaaS. The distinction matters enormously, but the market isn’t making it.

The Klarna episode is actually a microcosm of something happening across enterprise IT broadly. The average number of SaaS applications per organization has dropped to 106 from a 2022 peak of 130 as businesses eliminate redundant tools. This is not an AI story. This is a hygiene story. Enterprises accumulated far too many point solutions during the frenzied SaaS expansion of the early 2020s, and they are now rationalizing. AI is enabling that rationalization — it’s easier to consolidate when a well-designed agent can bridge systems that previously each needed their own interface — but it is not the primary driver of tool reduction in most cases.

The mistake the market is making is conflating rationalization with elimination. The bearish thesis extrapolates Klarna’s cost-cutting exercise into an extinction event. That is likely wrong. But the bullish thesis — that enterprise software is structurally fine and this is just sentiment — is also wrong, for the same reason. The headcount dependency is real, and it is being disrupted. The question is how much, and at what pace.

“The mistake the market is making is conflating rationalization with elimination. The bearish thesis extrapolates Klarna’s cost-cutting exercise into an extinction event. The bullish thesis says enterprise software is structurally fine. Both are wrong.”

The Platform Gravity Trap

Here is the deeper problem for application software companies: they built gravity, not infrastructure.

There’s a useful distinction to draw here between two types of competitive moat. Infrastructure moats are network-effect and capability moats — the more you use them, the more powerful they become, and the more of your stack depends on them. Think AWS. Switching costs are immense but also somewhat rational: you’re not just leaving a product, you’re leaving a foundation.

Application moats are workflow and habit moats. They’re built on the fact that your employees use the product daily, that your data lives there, that your processes are organized around it. These are real, but they’re brittle in a way infrastructure moats aren’t. They depend on the humans who embedded the habit. When you remove the humans, the moat drains.

The SaaS companies at most risk are the horizontal providers that rely on high-volume seat counts — CRM, HCM, project management, collaboration tools — all facing a “Productivity Paradox”: their tools make employees so efficient that customers need fewer of them, which means fewer copies of the software. One analyst framing for this is “Headless SaaS” — a future where the value isn’t in the user interface at all, but in the underlying data and API logic, accessible by agents rather than humans.

SaaS Capital has observed that “merely being a SaaS company is no longer a ticket to premium ARR multiples” and that today’s environment is a “rich get richer” dynamic where the highs are higher and the lows are lower than previous cycles. What separates the rich from the poor in this framework isn’t growth rate or NRR in isolation — it’s whether the product’s value is tied to human workflow or to underlying data and logic that agents can access just as well.

Salesforce understands this, which is why Agentforce exists. Marc Benioff told investors he expects AI agents to deliver “3x, 4x the ability to multiply the monetization on customers” because those customers get 10x more value from the products. That is a vision of the future where Salesforce is not a seat-based software company but a platform charging for outcomes. Whether they can execute that transition, and how quickly, is the open question. Agentforce adoption in 2025 was largely pilots and seeded licenses rather than paid deals at scale, with enterprises hesitating before committing to broad seat-based agentic platforms — partly because deployment still requires specialized skills most organizations don’t have.

The Infrastructure-Application Divergence Is a Feature, Not a Bug

One of the cleaner signals in the current repricing is the divergence between infrastructure and application software valuations. In 2025, IaaS was the dominant growth driver, accounting for 77% of total XaaS contract value, fueled by large-scale AI infrastructure buildouts. Meanwhile, by November 2025, ERP, CRM, HCM, supply chain, and marketing technology were all still under pressure. The same pattern that emerged post-2022 is accelerating: infrastructure holds, application compresses.

This is a well-worn dynamic in technology history. When a new platform arrives, the platform layer temporarily captures most of the value, while the application layer gets commoditized or compressed until new applications emerge that are native to the new platform. The PC era saw this happen to mainframe applications. The internet era happened to desktop software. Mobile happened to web. The timing has varied but the direction has been consistent: new platforms extract value upward before distributing it back down.

According to Menlo Ventures’ 2025 State of Generative AI report, enterprise AI application spending reached $19 billion last year — more than half of all generative AI spend — with AI-native startups earning nearly $2 for every $1 earned by incumbent software vendors in the application layer. Incumbents aren’t losing to AI agents. They’re losing to AI-native startups who built without the architectural constraints that come from a decade of seat-based assumptions. This is the more accurate competitive threat: not “AI replacing Salesforce” but “AI-native CRM startups not needing to retrofit pricing and architecture the way Salesforce does.”

DevOps-adjacent software companies are commanding median EBITDA multiples of 36.5x, and data infrastructure companies are at 24.4x — not because these categories are immune to AI disruption, but because their products are inputs to AI workflows rather than workflows that AI is replacing. That’s the distinction the market is trying to price, imperfectly and all at once.

What Survival Actually Looks Like

The companies that will emerge from this transition strongest share one characteristic that has nothing to do with AI features: they own data that can’t be scraped, replicated, or synthesized.

Workday’s co-president made exactly this point in response to the Klarna episode. The defensible core of an HCM or ERP system is not the interface, not even the workflow logic — it’s the decades of transactional data that creates the organizational record, the compliance history, the financial audit trail. You can’t replace that with an LLM, as the Workday co-president argued: “The AI is not going to build the system for you.” An agent can run on top of Workday’s data. But you can’t train away the data itself.

This creates a mental model that might be useful for sorting winners from losers in the current repricing: the difference between being a container and being a conduit.

Containers hold data — they are the system of record. When AI agents need to act, they need containers. Containers retain value because moving data is expensive, risky, and complex. Conduits, by contrast, are the applications that move humans through workflows — the click here, type there, generate this report interfaces that mediated human activity. Conduits are what AI agents replace. When an agent can retrieve, analyze, and act on data directly, you don’t need a conduit for a human to do the same thing.

The SaaS companies most threatened are the ones that were primarily conduits: beautiful dashboards, workflow managers, collaboration tools built around the assumption that humans are the processing layer. The ones best positioned are either deep containers — owning unique, non-replicable data assets — or genuine platforms that can become the operating system for agents rather than humans.

The current selloff is indiscriminate. It will eventually become discriminating. When it does, the companies that understood what they actually owned — and built their AI transitions accordingly — will look very different from those that bolted AI onto conduits and called it transformation.

That distinction is worth more than any multiple compression we’ve seen so far.

Liat Ben-Zur is an AI strategy and governance advisor to Fortune 500 companies and technology boards. She led the first commercial GPT-4 launch at Microsoft and grew its consumer business from $6B to $13B. She serves on the boards of Talkspace, Compass Group PLC, and Splashtop, and advises Concord Music Group and GeneDx. She is the founder of LBZ Advisory LLC and is writing The Bias Advantage — a book about how leaders from underrepresented backgrounds can leverage their unique perspectives in the AI era. She writes about AI strategy, governance, and the business of intelligent systems at liatbenzur.com.