AI Disruption Triggers $1 Trillion Selloff as Enterprise Software Faces Existential Threat

Salesforce and Friends Deserve This AI Squeeze

Product releases from Anthropic PBC last week triggered a nearly $1 trillion selloff of enterprise software stocks, extending a decline that's been on course for some time. Salesforce Inc., which sells customer relationship management software, and Workday Inc., which provides financial management software, are both down more than 40% over the past 12 months. On Wednesday, the market rout spread to wealth management firms and, it seemed, any company that appeared to be in the crosshairs of artificial intelligence. This sell-first, ask-questions-later moment is overdone. In the case of enterprise software makers, investors are missing a crucial technical distinction between vendors like Salesforce, and AI tools such as Anthropic's Claude Cowork that suddenly look like a threat. What AI can do increasingly well is the higher-level knowledge work that many software-as-a-service (SaaS) applications were built to facilitate. That part of the business of Salesforce and its peers is indeed under threat. But AI can't yet compete with their underlying, systems-of-record offerings, which process proprietary data like billing, compliance and audit trails for a corporate customer. "These are precisely our areas of strength," Madhav Thattai, an executive at Salesforce, tells me. He's right. Agents, the hot new AI tools that carry out tasks like booking an appointment instead of just answering questions, can't replicate thousands of bespoke business rules built up over years, areas where firms like Salesforce, SAP SE, Oracle Corp. and Epic Systems Corp., used by hospitals, are entrenched. Other SaaS executives have pushed back on the bears. SAP Chief Executive Officer Christian Klein argued on an earnings call in January that clever generative AI models couldn't work with the critical business data and workflows that are his company's bread and butter. As independent analyst and former venture capitalist Benedict Evans has noted, successful SaaS products are sparked by someone identifying a unique, organizational problem, mapping out a workflow and then coding it into software. That's how you get the complex, often niche workflows that become the plumbing infrastructure for banks, schools, retailers, hospitals and police departments. It's when SaaS firms also act as plumbers that things go downhill. The applications built on top of all that database infrastructure have long been terrible: clunky, unintuitive and overpriced, and sometimes insecure. To make matters worse, customers are often stuck using these systems because moving from one provider to another is a lengthy and expensive process. Sign up for the Bloomberg Opinion bundle Sign up for the Bloomberg Opinion bundle Sign up for the Bloomberg Opinion bundle Get Matt Levine's Money Stuff, John Authers' Points of Return and Jessica Karl's Opinion Today. Get Matt Levine's Money Stuff, John Authers' Points of Return and Jessica Karl's Opinion Today. Get Matt Levine's Money Stuff, John Authers' Points of Return and Jessica Karl's Opinion Today. Bloomberg may send me offers and promotions. Plus Signed UpPlus Sign UpPlus Sign Up By submitting my information, I agree to the Privacy Policy and Terms of Service. "Software industry models in the U.S. are shaped around monopolization, offering low quality and bad security for high prices," writes Matt Stoller, director of research at the American Economic Liberties Project, a nonprofit that campaigns against monopolistic practices. Stoller, in his latest newsletter, describes a 2016 meeting with community bankers in which they derided their niche software vendors as "expensive" and "terrible." Swedish fintech company Klarna Group Plc stopped using software from Salesforce and Workday in 2024, replacing the incumbent vendors with tools from smaller SaaS companies with names like Deel and Neo4j, then using an AI coding tool called Cursor to build a more modern application layer on top. In other words, Salesforce customers like Klarna aren't just replacing their old SaaS software with AI tools. They're also using AI to build their own applications to better serve their needs, squeezing out the expensive interface layer while keeping the underlying data intact. The rather boring data management and compliance systems sold by big SaaS companies aren't what's under threat, but their apps are. And Salesforce sits right on the fault line of what's safe and what's vulnerable, being partly a system of record and partly a knowledge interface that AI tools are trumping. Last year Salesforce boldly tried to stave off the threat, becoming the first large tech company to sell AI agents with a program it called Agentforce. CEO Marc Benioff said the new platform was core to what Salesforce did, even suggesting the company could change its name to "Agentforce." But Agentforce's performance has been lackluster. Christine Marshall, a Bristol-based Salesforce trainer and one of the company's most recognized outside experts, wrote a surprisingly tepid review of the platform last year. After testing Agentforce on six common tasks, she found it only handled two of them well, and fumbled helping a user reset their password. Salesforce's Thattai wouldn't address the review when I asked him about it, saying that Agentforce behaved in a consistent and reliable way, even when paired with a large language model. Agentforce, ironically, uses AI models from OpenAI and Anthropic, the companies that now appear to be threatening its business. A market correction was needed for the inelegant applications that so many businesses are locked into. But SaaS companies have enjoyed high valuation multiples because they control both the infrastructure and interface. If technology from Anthropic and OpenAI can sit on top of those systems of record, they'll start to erode the pricing power of the SaaS companies that run them. That means that for enterprise software's bloated application layer, the age of easy margins may be over. More from Bloomberg Opinion: * Claude AI May Make Analyst Groupthink Even Worse: Parmy Olson * Who's On the Other Side of the Big AI Selloff?: Chris Hughes * Wall Street's Doom-Mongering on Software Is Bizarre: Dave Lee Want more Bloomberg Opinion? Terminal readers head to OPIN <GO> . Or subscribe to our daily newsletter .

Software Apocalypse: It's the model not the sector.

Anthropic's release of Claude Cowork industry plug-ins on January 30 sent a message that Wall Street heard loud and clear: frontier AI labs are no longer sticking to building tools for developers. They are expanding to build the replacement for enterprise software itself. Within 48 hours, Thomson Reuters fell 16%, RELX dropped 14%, Wolters Kluwer lost 13%, and Monday.com was down over 20%. Jefferies trader Jeff Favuzza called it a "SaaSpocalypse." However, the market probably overreacted and threw the baby out with the bathwater by hitting sell on every software stock. For the past week, we have been looking at every corner of the market to properly measure who will suffer the most and the least from this SaaSpocalypse. We found that there are three types of SaaS that will each have a different fate, and decided to build a quadrant to reflect this reality. Feel free to play around with this chart via this link. To really understand who the winners and losers of this new evolution of SaaS companies will be, we will first have to look how each of them make money as it is the best litmus test to judge a company's health. These are companies whose core asset is proprietary, curated data that cannot be reproduced by scraping the internet or training a model on publicly available sources. Think of Thomson Reuters' Westlaw (that has decades of attorney-curated case law), Intuit's TurboTax (100 million Americans' tax returns and financial records), Experian's credit bureau (data on 1.4 billion people and 200 million businesses), or Wolters Kluwer's UpToDate (clinical decision data used by 3 million clinicians globally). Their business model is licensing access to irreplaceable datasets. In an AI world, their data becomes more valuable, not less, because AI models need high-quality, domain-specific data to produce trustworthy outputs in fields where getting it wrong has legal, medical, or financial consequences. As investor Elad Gil put it in a recent McKinsey interview: "Data as some useful input to do something better for your business is incredibly valuable. Data as a core differentiator of your business is rarer. And data as a primary competitive advantage applies to a very, very small number of companies." These are the companies most people think of when they hear "SaaS": per-seat subscription tools that automate business processes for human users. Salesforce for sales teams. Workday for HR departments. Monday.com for project managers. Adobe for designers. ServiceNow for IT workflows. All software company CEOs are now scrambling to acquire AI companies, therefore calling Thoma Bravo, a Private Equity specialist in acquiring distressed tech startups among others. Their revenue is a direct function of how many humans sit in front of screens clicking buttons. The business model is essentially a gym membership for the office: you pay per person, per month. Just like gyms, SaaS companies make money only when you're not really using the service. The problem is obvious. When one employee with an AI agent does the work of five, you do not need five licences anymore. You need one. And the gym's revenue just dropped 80%, not because the equipment got worse, but because the customers disappeared. Databricks CEO Ali Ghodsi made a sharper observation recently: AI is not killing SaaS by replacing enterprise systems. It is killing them by making their interfaces irrelevant. For decades, the moat these companies relied on was UI complexity: millions of people trained on Salesforce, SAP, or internal dashboards. When workers can instead ask questions and take actions through natural language, that moat evaporates. Software becomes plumbing, not destination. These are companies building the connective tissue between raw AI models and enterprise operations: the "operating system" layer that neither the AI labs nor the workflow tools own. Palantir is the purest expression of this "platform". Snowflake and Databricks occupy adjacent territory on the data infrastructure side. ServiceNow is attempting to pivot here from its workflow origins. Their business model is fundamentally different from per-seat SaaS. They charge for outcomes, data throughput, or platform access rather than counting human users. As Goldman Sachs CIO Marco Argenti described in his 2026 AI outlook: companies will shift from deploying human-centric staff to deploying human-orchestrated fleets of specialised multi-agent teams. These hybrid teams of humans and machines will charge clients by tokens consumed (the units of data used by AI models), not seats occupied. This is the "agent-as-a-service" economy. Palantir's Ontology platform is the clearest example of what this looks like in practice. It functions as a digital twin of an entire organisation: not just tables and databases, but real-world objects (employees, aircraft, purchase orders, threat actors) with their relationships, properties, and business logic mapped as a living model. AI agents operate on Ontology with full enterprise context. The platform is model-agnostic (it works with GPT-4, Claude, Gemini, Llama, or any custom model), meaning better and cheaper AI models make the platform more powerful without threatening its position. The selloff unfolded in three discrete shocks over ten days. Anthropic released eleven industry-specific plug-ins for its Claude Cowork autonomous agent, spanning across legal, finance, sales, marketing, customer support, data analysis, and enterprise search, each capable of automating end-to-end professional workflows. The blog post didn't mash words: "Tell Claude how you like work done, which tools and data to pull from, and how to handle critical workflows." Wall Street heard it as a direct assault on every per-seat SaaS vendor in existence and hit sell on every SaaS stock they could find. As a consequence, Thomson Reuters cratered 16% in 48 hours. RELX dropped 14%. Wolters Kluwer fell 13%. LegalZoom lost 20%. Some had it coming, others not. Goldman Sachs' basket of US software stocks sank 6% in a single session, its worst day since the April tariff shock. Total single-day US software losses hit $300B. The IGV iShares tech-software ETF hit its most oversold level relative to the S&P 500 in its 25-year history. Jefferies called it the largest relative oversold reading they had ever recorded. Anthropic released Claude Opus 4.6 (1-million-token context, "agent teams" coordinating multiple AI workers in parallel, native Excel and PowerPoint integration, outperforming GPT-5.2 by roughly 144 Elo points). The same day, OpenAI launched Frontier, an end-to-end enterprise platform for building and deploying AI agents across business systems, with early customers including Uber, State Farm, Intuit, and Oracle. Two frontier AI labs, on the same day, both saying: we are coming for your enterprise software contracts. By February 6, the S&P 500 Software & Services Index sat more than 20% below its October 2025 peak. Total market cap destroyed: more than $1 trillion since January 28. The AI labs are raising and spending at a pace that dwarfs anything the software industry has ever seen. Consensus capital expenditure estimates for the four largest hyperscalers (Amazon, Alphabet, Meta, Microsoft) in 2026 now exceeds $600B, revised upward repeatedly since October 2024. Anthropic's revenue trajectory highlights the demand for these services. From $1B annualised revenue at the start of 2025 to $9B by year-end (9x in twelve months), the company is now targeting $18B for 2026 and closing a $20B+ funding round at a $350B valuation. Claude Code alone hit $1B in ARR within six months of launch, described as the "fastest-growing product of all time." As a comparison, Salesforce took 17 years to reach $18B in revenue. Anthropic is on track to do it in two. The initial selloff targeted the obvious victims: horizontal SaaS companies with per-seat pricing models. But the fear is spreading to corners of the software market that most investors assumed were protected. Dassault Systemes, the French industrial software giant, crashed 22% on Wednesday, 11 Feb: its largest single-day decline on record, wiping out roughly €6B ($7.1B) in market value. Dassault creates "virtual twins" of complex machines using simulation software. Its full-year 2025 revenue was flat at €6.24B, and 2026 guidance of 3-5% growth badly missed analyst expectations of 5.9%. Software revenue dropped 5% in Q4. Licence sales fell 7%. What scared the investors was not only the weak growth. It was the AI narrative behind them. Nvidia CEO Jensen Huang recently described Dassault as being at the "epicentre of the next frontier of artificial intelligence." But that frontier, "world models" (AI systems that simulate and navigate the physical world), is exactly what threatens to commoditise Dassault's industrial digital twin offering. Former Meta Chief AI Scientist Yann Le Cun has started fundraising for his stealth startup called AMI Labs, focussed on building these very 'world models' that could endanger Dassault Systèmes' business. Additionally, Nvidia's own Omniverse platform, combined with open-source physics simulation models, could replicate much of what Dassault's 3DEXPERIENCE platform does, at a fraction of the cost. Dassault Systèmes' P/E has compressed from 35x to 20x. The lesson here is: the cheapest models will eat the low hanging fruit first. The tech-savvy firms, the cost-cutters, the SMBs and nimble startups that are already comfortable with AI tools will adapt to this new reality the quickest. They will cancel their Monday.com subscriptions, build internal tools with Claude Code, and replace junior staff with AI agents. This cohort of companies is already moving: SaaStr reports it runs an eight-figure business with single-digit headcount and 20+ AI agents, down from 20+ full time employees. This is a deadly blow to "workflow providers" as they get cannibalised by internal tools and orchestration agents. The carnage is concentrated but not contained. Declines from 52-week highs as of early February: For years, giants such as HubSpot and Atlassian flourished by taxing human productivity through a monthly fee for every 'seat' a company filled. However, as AI moved from simple assistance to autonomous agency, the market realised what was going on: the workforce of the future would be digital rather than human-only. Investors panicked as they realised that an AI agent does not need a Figma login or a Workday account to function. The resulting crash in these stocks represented the death of the 'per-seat' economy, as capital rotated aggressively from the tools that humans use to the intelligence that replaced them. Every single name here is in negative territory on a one-year basis except MongoDB (+25%) (who profits massively from the agentic deployment in enterprises, and Zoom (+11%) (who investors are using as a proxy for Anthropic exposure via Zoom's participation in its Series C round). SaaS companies like Adobe, Salesforce, Workday, and ServiceNow were treated as "growth" stocks throughout the pandemic era. COVID forced digitalisation compressed a decade of enterprise software adoption into 18 months. Suddenly every company needed cloud-based collaboration, CRM, HR, and creative tools. SaaS stocks re-rated to 30-50x forward earnings on the assumption that this growth trajectory was permanent. But this dream scenario didn't last long, as growth slowed, but the valuations remained sky high. It was propped up by two temporary tailwinds: a low interest rate environment that made future cash flows worth more, and the first leg of the AI era (2023-2025) where large language models were new and promised increased efficiency for employees and existing workflows. The narrative was: "AI makes your Salesforce reps more productive, so you buy more seats, not fewer." That was the narrative that led to the SaaSpocalypse on Jan 30th. The enterprise AI use cases demonstrated by Claude Cowork and OpenAI Frontier go far beyond what anyone imagined during the ChatGPT honeymoon phase. When execs realised that these are not productivity boosters only for human workers but instead were autonomous agents that replaced the human workers entirely, their whole worldview changed almost overnight. They learned to make the distinction between an existential threat disguised as a temporary tailwind. Better late than never? Companies that were trading at 30-50x forward earnings on the promise of everlasting 20%+ revenue growth are now trading at 12-20x as the market asks the question on whether their competitive advantage is really as deep as they previously thought. To add fuel to the fire, the codebases at these companies took decades to build at a cost of billions of dollars. That accumulated software was capitalised as intangible assets on SaaS balance sheets, representing a good chunk of their enterprise value. But with Claude Code or Cursor, what took senior engineers months can now be done in 48 hours, and junior engineers usually turn up to product meetings with the product entirely built using Claude Code in a matter of hours. The technical moat these SaaS players had is vanishing, making SaaS a distribution layer instead of their previous product excellence that drove the sky high valuations. There are some concerns about AI that keep the SaaS companies from losing everything overnight. Firstly, AI is massively loss making, as OpenAI ended 2025 with $20B+ in Annual Recurring Revenue (ARR) but faces projected 2026 losses of roughly $14B and cumulative losses through 2029 are estimated at $115B. Profitability is not expected until 2030. Anthropic targets breakeven by 2028. The cost of training frontier models continues to rise even as the cost of per unit inference continues to fall. We spoke about this shift to inference in a previous essay. This gives much-needed time to the "data owner" SaaS companies to play catch-up and monetise their positions (like the Intuit-OpenAI deal). Secondly, AI is largely unregulated. There are no seat minimums, no licensing requirements, no professional standards for deploying an AI agent that handles customer support or writes legal briefs. The barriers to adoption are essentially zero for any company willing to experiment, which accelerates the cannibalisation of SaaS seats. While the lack of regulation here will accelerate the "workflow substitution", it won't be able to touch the US Food & Drug Authority-validated systems, the credit bureau infrastructure or the lethal side of military use. Finally, and most importantly, AI does not have a settled business model of its own. Despite the size of these frontier labs, they are still in a "startup" phase without a fixed business model, without a fixed customer base, and with skyrocketing costs. To address that, AI labs are trying to be infrastructure providers (like Cisco during the dotcom boom), platform companies (like Palantir), and application vendors (like the best-in-class SaaS companies) all at the same time. They are competing with their own customers (building CRM-like features while selling API access to companies that build CRMs). This instability creates opportunities for incumbents on the "data owner" & "platform" side that adapt quickly, but it also means the disruption is chaotic and unpredictable. What we believe will happen over the next 2-3 years is the very painful transition away from the per-seat business model that SaaS vendors have cherished for the past two decades. Per-seat pricing in today's environment is becoming extortionate. When one AI-augmented employee does the work of five, charging $150 per seat per month for each of those five seats is a tax on inefficiency that CFOs will not tolerate. SaaStr's analysis shows that at Salesforce, price increases account for roughly 72% of forward growth (6.3 percentage points of price hikes against 8.7% total ARR growth). Interestingly, Salesforce is trying to turn "AI that works" into a business model shift: moving beyond charging only for human seats, it is pricing agentic output as usage by taking Agentforce from a simple $2 per conversation to Flex Credits priced per agent action (20 credits per action, positioned as $0.10 per action). The clever part is the bridge between old and new spend, because the Flex Agreement lets customers convert user licences into Flex Credits and back again, so budget can flow from human licences to digital labour as workflows automate. We are still very early to what the next pricing models look like, as Salesforce is just one example. Here are some hypotheses on where the next business model for software companies could look like in the age of agents: You pay for the result, not the process. Think of it as similar to the success fee model in M&A advisory or executive search: the software vendor earns a percentage of the value created, not a monthly retainer regardless of output. This is high-margin but hard to standardise. You pay for tokens consumed (millions of tokens of AI processing) or "units of cognition," as Elad Gil described to McKinsey: "Instead of buying customer success seats, you're buying customer support queries that are answered." Salesforce's Flex Credits and similar models are early versions of this. Companies like Palantir negotiate custom contracts based on the specific enterprise use case, data volumes, and operational requirements. They get this information through their "Forward Deployed Engineers" who are technical consultants who spend weeks in the prospective enterprises, making an extremely detailed graph of how the company works and deciphers their operational pain points. Anthropic has started trialling this business model where they have their FDEs deployed at Goldman Sachs for 6 months, to co-develop specific tools that their organisation will require. This is the highest-value approach but requires deep integration and long sales cycles. Palantir's average contract value is rising rapidly (180 deals above $1M in Q4, 61 above $10M), and net dollar retention hit 139%. The transition will be as painful as the shift from on-premise software to cloud in 2010-2015. Adobe is the canonical example: when it moved from perpetual Creative Suite licences ($2,599 one-time) to Creative Cloud subscriptions ($55 per month) in 2013, the stock dropped 16% over the following year as revenue temporarily collapsed before recovering to all-time highs. The margin compression and growth slowdown will be similar, but the recovery will be selective: only companies that own proprietary data or occupy the orchestration layer will rerate. Lastly, the question that might haunt these AI companies for years to come: headcount. The deployment of Claude Cowork and OpenAI Frontier across Fortune 500 companies will reduce knowledge-worker headcount. McKinsey estimates current AI technologies could automate roughly 57% of US work hours. Goldman Sachs projects 300 million full-time jobs globally are exposed. The pace of AI adoption is racing ahead of the readiness level of these enterprises, with SaaS seat counts caught in the crossfire. (our selection of winners and losers subject to the AI-driven business model repricing) To dig further into smaller companies bound to benefit from this AI-driven business model repricing, create your own StockScreener like we did: To find out more about ETFs available to index this AI-driven business model repricing, make your own ETF Screener like we did: The transition from per-seat to fluid, outcome-based pricing will be painful, and it will take years. This structural repricing is similar in magnitude to Adobe's shift from perpetual licences to subscriptions in 2013, but with a harder edge: the business model is going to be more fluid and outcome driven to accommodate an agentic workforce, rather than the static per-seat subscription model. What propped up these stock prices until now was a unique mixture of tailwinds that has now reversed: COVID-driven forced digitalisation created an illusion of permanent demand acceleration. Low interest rates inflated the present value of future cash flows. Incumbents were massively underprepared for this AI expansion on their businesses, giving them a false sense of security. And the first wave of AI (2023-2025) was perceived as additive (more productivity per seat) rather than substitutive (fewer seats needed). That view is coming crashing down as of January 30 and February 6. The market is now repricing the entire software sector through the lens of a simple question: does your business model survive in a world where AI agents replace human workers? Companies that own the proprietary data AI needs (Thomson Reuters, RELX, Wolters Kluwer, Intuit, Experian) will emerge as winners. Companies that orchestrate AI within enterprises (Palantir, Databricks, Snowflake) will build the next generation of enterprise infrastructure. But companies that merely provide per-seat workflow tools, with no proprietary data moat and no platform lock-in, face a prolonged and painful transition that the market is only beginning to price. As Jefferies' Ron Eliasek wrote yesterday: AI will be transformational, but software companies with deep moats, trusted brands, and top talent are still positioned to thrive. We agree with the first half. The second half depends entirely on which type of software company you are talking about. It is the model (both the business model and the AI model), not the sector.