Eight Theories of How Agents Strangle SaaS
speculations on suffocation
On January 28, 2026, ServiceNow reported fourth-quarter earnings that beat Wall Street expectations for the ninth consecutive quarter. Subscription revenue was up 21% year-over-year. Guidance was raised. The stock dropped 11%.
The same day Microsoft posted $81.3 billion in quarterly revenue, beating both top and bottom line estimates. Cloud revenue crossed $50 billion for the first time. The stock fell 10%, erasing $357 billion in market value in a single day. It was the largest single-day dollar loss in Microsoft’s history.
Within a week, Anthropic launched Claude Cowork and a legal AI plugin. Software stocks entered a bear market. The iShares Expanded Tech-Software Sector ETF dropped 22% from its highs. Roughly $285 billion in market value vanished. Jefferies traders coined a term for it: the “SaaSpocalypse”.
Something strange is happening. SaaS companies are hitting their numbers. They are growing revenue, beating estimates, raising guidance. And the market is punishing them anyway.
The new thing turns out to be a set of converging forces reshaping the $273 billion SaaS industry from the outside in. Underneath the selloff are at least eight distinct theories about how AI disrupts SaaS.
Here is the map.
Theory 1: The Superagent Eats the Interface
The most ambitious theory goes like this: AI agents become the primary way humans interact with software. Not copilots sitting inside apps but central orchestrators that sit above all your apps and talk to their APIs on your behalf.
Think about how a marketer runs a Q1 campaign today. They open Google Analytics to pull traffic data. Switch to HubSpot for lead scores. Jump to Hootsuite for social scheduling. Fire up Mailchimp for email sequences. Each tool has its own login and its own dashboard and its own logic for what “performance” means.
Now imagine telling a single agent: “Plan my Q1 campaign. Analyze last quarter’s performance across all channels, generate content variations, schedule posts, set up A/B tests, and optimize in real time.” The agent talks to each tool’s API and pulls the data and does the work and reports back. You never log in to anything.
In this world SaaS tools become what the telecom industry calls “dumb pipes.” They store data and expose APIs but the value creation from the analysis and the decisions and the workflows moves to the agent layer. The agent builds network effects through accumulated history and cross-app insights and personalized knowledge graphs that no single SaaS copilot can match. The more you use the agent the better it gets and the harder it is to switch.
This creates a brutal dynamic for SaaS companies. If the agent is the interface then the software underneath is interchangeable. Pricing wars follow. Innovation margins shrink. The moat that took twenty years to build becomes irrelevant because users never see your product anymore.
Satya Nadella described this shift with unusual candor on the Bg2 Pod: “Business applications are essentially CRUD databases with a bunch of business logic. The business logic is all going to these agents. They’re going to update multiple databases, and all the logic will be in the AI tier.”
Coming from the CEO of the company that owns both Dynamics 365 is a remarkable admission.
Theory 2: The Great Unbundling
Every successful SaaS product is at its core a bundle. You pay for a CRM but you really only use contact management and pipeline tracking. You pay for a project management suite but you mostly need a shared task list. The bundle worked because building custom software was expensive and hard.
AI changes the economics of building.
People using AI chatbots are learning the atomic units of AI-based software: prompting, context windows and few-shot learning. They are developing opinions about how workflows could be better and easier and cheaper. That creates the opportunity to unbundle those workflows into separate purpose-built apps.
But the unbundling is also happening at the individual level. A sales team paying $30,000 a year for Gong’s call analysis and CRM integration can now replicate the core feature with Claude Code in a weekend. The full product has multi-language support and compliance controls and a slick UI. But many teams do not need all of that. They need the one feature that solves their specific problem.
Thanks to AI agents and AI-assisted coding getting good enough you can rebuild most enterprise SaaS functionality and host it cheaply and get 90% of the capability. That last 10% is real -- security and scale and compliance and support -- but for thousands of smaller teams 90% is more than enough.
This is the classic Christensen disruption pattern. The incumbents serve the high end of the market with comprehensive products. The disruptors sneak in at the bottom with simpler and cheaper alternatives. By the time the incumbents notice the disruptors have improved enough to compete everywhere.
Let see an example. LegalZoom built a billion-dollar business on legal templates: fill in your name, pick your state, pay $99, get an LLC. For years, this was genuinely valuable because legal language felt intimidating and hiring a lawyer felt expensive. But templates are just structured text generation with conditional logic, and that is exactly what large language models do natively.
When Anthropic released its legal plugin for Claude Cowork, LegalZoom’s stock cratered 20% in a single session. The stock had already been bleeding for years, down 81% from its IPO price, but that single Tuesday made the thesis explicit: why would anyone pay for a pre-built template when they can describe their situation in plain English and get a custom legal document generated in seconds? LegalZoom doesn’t hold your corporate records. It doesn’t integrate with your bank. It doesn’t file your quarterly compliance reports. It generates documents. And document generation is now a commodity that costs fractions of a cent per API call.
Theory 3: The Uncertainty Tax
Here is a theory that is less about technology and more about investor psychology.
SaaS companies have enjoyed premium valuations for decades because of their predictability. Recurring revenue. High margins. Low churn. These are the qualities that made SaaS the greatest business model in software. But AI introduces genuine uncertainty about the durability of those economics. And uncertainty is the one thing financial markets hate most.
Consider Salesforce. It has been around for 25 years. It is battle-tested and deeply embedded in enterprise workflows and still growing. But investors are now asking: what happens when AI agents reduce the number of humans who need CRM seats? If ten AI agents can do the work of a hundred sales reps you do not need a hundred Salesforce licenses anymore. You need ten. That is a 90% reduction in seat revenue for the same work output.
It’s already happening. A customer recently terminated a $350,000/year Salesforce contract and replaced it with a custom AI-powered solution. One data point does not make a trend but multiply it across thousands of enterprise customers and you start to understand why the market panicked.
The uncertainty tax means that even SaaS companies with strong fundamentals get repriced downward. The market cannot confidently model what their business looks like in five years.
The result is a strange divergence: SaaS companies are trading at price-to-earnings ratios near ten-year lows while their current fundamentals remain strong. The market is pricing in a future that has not arrived yet.
Theory 4: The $0 MVP
Traditional SaaS was expensive to build. An MVP cost $500,000 to $1 million. You needed designers and engineers and QA testers and DevOps and months of development time. This cost structure created a natural barrier to entry. If you had a working product with paying customers it was hard for competitors to catch up because they had to spend the same money and time you did.
AI collapses this barrier.
New startups can adopt coding agents fully -- Claude Code and Cursor and Codex -- to go from idea to working prototype in days instead of months. The cost of an MVP approaches zero. A single developer with good taste and an AI coding agent can build what used to require a team of ten.
At Every Dan Shipper has described how they run five software products each primarily built and run by a single person shipping to thousands of daily users. A single developer doing the work of five is already reality for teams that have adopted compound engineering where each bug fix and code review and feature becomes a learning loop that makes the next feature easier to build.
For incumbents this is terrifying. Any successful feature you have can be replicated by a hundred competitors overnight. The high-margin window that used to last years now lasts months. And many large companies are slow to adopt AI coding tools because of bureaucracy and security reviews and institutional inertia.
The irony is cruel: the cost-cutting decisions that enterprises made to stay profitable in the short term(reducing headcount and deferring R&D) are exactly the decisions that make them slower to adopt the tools that could save them.
Theory 5: The Incumbent’s Dilemma
Large enterprise software companies have something that startups do not: decades of customer relationships and regulatory expertise and compliance certifications and institutional trust. These are real moats. A hospital is not going to rip out its EHR system because a startup demos a cool AI agent.
But these moats can also become prisons.
The incumbent’s dilemma is that enterprises integrate AI too slowly and too cautiously and in ways that preserve existing business models rather than embracing new ones. A company like Workday has deep expertise in HR compliance across thousands of legal jurisdictions. That expertise is genuinely hard to replicate. But if Workday treats AI as a feature to bolt onto its existing product rather than a fundamental rethinking of how HR software works it risks losing relevance to AI-native challengers who start from scratch.
The “slow roll” death looks like this: gradual feature obsolescence without bold reinvention. Customers do not leave overnight. They just stop expanding their contracts. They consolidate seats. They start building small internal tools for workflows they used to do in the SaaS product. By the time revenue starts declining the window for reinvention has closed.
Workday’s own trajectory is instructive here. In February 2025 the HR software giant laid off 1,750 employees -- 8.5% of its workforce -- to “prioritize innovation investments like AI.” Then in early 2026 it cut another 400 support roles. A company that sells workforce management software is shrinking its own workforce because of AI.
Meanwhile the AI-native companies are moving fast and loud. Shopify CEO Tobi Lutke told all employees in a leaked memo that teams must now prove AI cannot do the work before they are allowed to hire. Duolingo CEO Luis von Ahn went further and announced the company would stop using contractors for any work AI can handle. His team then shipped 148 new language courses in under a year -- work that previously took a decade. These companies are using AI to disrupt the economics of their own operations before someone else does.
Theory 6: The Invisible App
This theory takes the agent story from Theory 1 and pushes it to its logical conclusion: AI agents do not just sit above SaaS interfaces. They bypass them entirely.
Today when you use a CRM you interact with a dashboard. You click buttons and fill in fields and read reports. The interface is not just a nicety -- it is the product. SaaS companies employ hundreds of designers and product managers to build and refine these interfaces. The entire value proposition is: “We made a complex workflow feel manageable through good design.”
But agents do not need good design. They need good APIs.
An AI agent reviewing contracts does not need a beautiful document review interface. It needs to read the contract and compare it against precedent and flag deviations and route decisions to humans. It can do all of this through APIs without any visual interface at all.
This is what Tina He described in Every as the world of “boring businesses” that will dominate the AI era -- companies that own the essential infrastructure between AI decisions and real-world consequences. The ones that own what AI must flow through but cannot replace: data sources and regulatory compliance and physical-world integrations.
If this theory is right then the entire UI/UX layer that makes SaaS products feel premium becomes worthless. What remains valuable is the data and the integrations and the trust layer. SaaS does not die -- it becomes invisible.
Theory 7: The Probabilistic Divide
Not all SaaS is created equal. This theory argues that AI’s impact depends on what kind of work the software does.
On one end of the spectrum you have “probabilistic” SaaS for tasks where errors are tolerable and creativity is valued. Content creation tools and marketing automation and copywriting assistants. These categories get eaten alive by AI because the AI can do the work directly. Why pay $200/month for a content optimization tool when Claude can write and edit and optimize content in a single conversation?
On the other end you have “deterministic” SaaS -- systems of record where accuracy is non-negotiable. HR databases that calculate payroll. Financial systems that handle regulatory reporting. ERP systems that manage supply chains. These systems survive because “good enough” is not acceptable. A payroll system that hallucinates is a lawsuit.
The divide creates a winner-takes-more dynamic. Companies with deep data moats in deterministic categories like Workday for HR compliance or Epic for healthcare records become more valuable because AI agents need reliable data sources. Meanwhile thin workflow layers in probabilistic categories get demolished.
Zendesk is a useful case study. Its value was always tied to the number of support agents using the platform. Fewer agents means fewer seats. ServiceNow revealed after its $2.85 billion acquisition of Moveworks that AI agents now resolve 90% of IT and 89% of customer support requests autonomously inside its own operations. If that performance generalizes across the industry the seat-based support model collapses. The data from customer interactions, resolution patterns and institutional knowledge remains valuable. But the interface and the per-seat pricing do not.
Jasper AI -- the AI content writing tool that raised $125M at a $1.5B valuation in October 2022. Revenue peaked around $90M ARR in 2023. Then ChatGPT arrived and commoditized its entire product. Revenue dropped to roughly $55-88M in 2024. The company cut its internal valuation by 20%. Both co-founders stepped down. Layoffs followed. A thin wrapper on top of GPT that charged $49/month for something ChatGPT could do natively. They have recovered but the damage is real.
Theory 8: The End of the Seat
For decades SaaS pricing was simple: charge per user per month. More employees meant more seats meant more revenue. This model was elegant because it scaled with customer success. As companies grew their software bills grew proportionally.
But AI breaks this alignment. If AI agents do the work of ten humans companies do not need ten seats. They need one seat and nine agents. The agent does not have a login. It does not need a dashboard. It should not cost the same as a human user.
Gartner predicted that over 30% of enterprise SaaS solutions will incorporate outcome-based pricing components by 2025. Salesforce has already started experimenting with what it calls “Agentic Enterprise License Agreements” -- flat-fee structures designed for companies deploying AI agents at scale.
This is a profound shift. Instead of paying for access to software companies pay for outcomes delivered. Instead of counting heads they count results. Instead of selling seats vendors sell capabilities.
Palantir CEO Alex Karp crystallized this on an earnings call that triggered a $300 billion market cap decline across SaaS stocks: AI is now good enough at writing and managing enterprise software that many SaaS companies risk becoming irrelevant. The old contract -- “we charge you for the privilege of using our software” -- is giving way to a new one: “we charge you for the work our software actually does.”
So Is SaaS Dead?
No. But it is transforming into something that looks very different from the industry we have known.
The best analogy might be what happened to telecommunications. Phone companies did not disappear when the internet arrived. But they went from being the most prestigious companies in the world to being utilities -- essential and profitable but no longer commanding premium valuations or capturing the most talented people. Data still flows through their pipes. They just do not control what happens with it.
SaaS companies that adapt by building agent-friendly APIs and shifting to outcome-based pricing and owning valuable data in deterministic categories will thrive. Companies that cling to seat-based models and human-centric interfaces will gradually lose relevance.
The eight theories above are not a menu where only one turns out to be right. They are all happening simultaneously at different speeds in different market segments. The unbundling is already underway. The repricing is happening in real time. The agent layer is being built as you read this.
Jensen Huang pushed back on the panic: “This notion that the software industry is in decline and being replaced by AI” is wrong he said. He might be right that AI augments software rather than replacing it. But augmentation that reduces the number of humans who interact with your product is still a form of disruption for the company selling per-seat licenses.
Is the market over-reacting? If yes, this won't be the first time.
Theory 7 (probabilistic vs deterministic) is the most underrated one here. Workflow tools are vulnerable because agents can replicate the orchestration. Data systems survive because agents need something to orchestrate.
The "API over interface" theory is playing out right now. I use maybe 3 web dashboards directly. Everything else goes through API calls my agent makes. The UI became optional faster than I expected.
Cost collapse for MVPs is real but slightly overstated. The build is cheap. The iteration and management around it still costs time and attention.
Interesting article, and food for thought. One more trend I've noticed: companies providing data infrastructure are now using their data moat to add capabilities that were earlier offered by their partners, essentially cannibalizing the ecosystem itself. I see convergence and consolidation into a few large companies that can serve customers of any scale. Unless you're a system of record, which giants already dominate, it seems almost impossible for new or emerging companies to become really big tech. What's your take and guidance for upcoming startups?