Anthropic recently introduced a set of 11 plugins for its AI agent Claude Cowork. These tools are designed to automate financial data analysis, legal document preparation, customer service, research writing, sales support and a host of other tasks. Almost immediately after the announcement, the shares of application software developers and data providers went down.
Against this backdrop, the stocks of the largest and most stable corporate software players fell sharply in a single trading day. Shares of Salesforce, ServiceNow, Adobe and Workday fell by about 7 percent, while Intuit fell by almost 11 percent. Over the same period, valuation multiples across the sector declined rapidly: The average forward earnings multiple for software companies fell from approximately 39X to 21X within a few months.
A familiar argument has returned to the market: if an AI agent can do the work itself, why does anyone need SaaS as a separate product?
Headlines heralding the arrival of the SaaSpocalypse are spreading rapidly. It’s a convenient media label that easily explains any fluctuations in the valuations of public SaaS companies. However, it is now shaping investor discussions and raising questions about who will be replaced first and how to determine which products can still be protected.
Below is a practical map of what will break first, what will survive and why.
5 Types of SaaS Defensibility During the SaaSpocalypse
- Integration hubs.
- Systems of record with a high cost of error.
- Data moats and feedback loops.
- Narrow domain expertise and regulation.
- High switching costs.
Why Is This Happening Now?
Artificial intelligence could not remain indefinitely as mere infrastructure in the form of a cloud or a single platform. Sooner or later, it had to take shape in the form of applied tools that directly automate work. This happened when the AI agent interface became simpler than the usual set of clicks it replaced.
The key shift is that reasoning models and the agent layer transform a complex task from “going through 20 screens” to formulating an intention. Now the user describes the exact goal, not the path to it:
- “Prepare a weekly report on the status of the project, identify risks and assign responsibilities.”
- “Reconcile accounts with payments and note discrepancies.”
- “Prepare a contract and indicate potentially risky clauses.”
This is a fundamentally new threshold for user experience. And when the interface changes so radically, it becomes clear that in many cases, SaaS was primarily a shell for getting work done.
Analysts’ estimates also support this. According to Goldman Sachs, by the end of the decade, AI agents will not only significantly expand the software market itself but also capture a disproportionately large share of profits. In this logic, agents cease to be an add-on to applications and gradually become the main interface for work. By 2030, more than 60 percent of the economic value of software could pass through agent systems rather than traditional SaaS licenses.
This forecast reflects structural changes: Agentic AI does not simply generate results; it interprets intentions and executes workflows across multiple systems. As economic value is concentrated at the level that makes decisions, directs and executes work, rather than simply storing data, a growing share of software revenue may shift to the agent level, even if the underlying SaaS infrastructure remains in place.
What Will Break First: SaaS as UI and Simple Logic
SaaS products whose value boils down to their interface and simple operations are the most vulnerable. These primarily include forms on top of spreadsheets; lists, statuses, and trackers; simple calculators; if-then automation and data transfer between systems.
For businesses, this means a direct impact on their employment structure. If a process can be reassembled as “subject matter expert plus an agent,” one role becomes sufficient where previously you needed five to 10.
This shift is already visible. Large companies directly link office role reductions to the introduction of AI: International law firm Clifford Chance, for example, cut 10 percent of its back-office staff in London, citing increased use of AI as one of the factors.
Politicians describe office-based professions, from lawyers and accountants to consultants and marketers, as among the most vulnerable. Governments are already preparing for a big retraining effort. According to London mayor Sadiq Khan and technology secretary Liz Kendall, “some jobs will inevitably disappear,” and by 2030, they plan to train up to 10 million workers in basic AI skills.
You Can Automate Routine, Not Responsibility
There is a point where action and responsibility diverge, however. An agent can prepare a plan, collect updates, generate a report and even initiate actions in systems. But responsibility remains with the person.
That is why roles that appear to involve routine aggregation (a classic example is project management) do not disappear. As task execution shifts to agents, these roles lose their operational load but remain critical where there is responsibility for deadlines and risks. This is especially noticeable in areas where the cost of error is high, as in aviation, where the autopilot performs most of the flight, but the pilot and airline are still legally responsible for errors.
Large SaaS Won’t Disappear; It Will Embed Agents
The most likely scenario for existing market players to survive is to embed AI agent capabilities directly into their products.
There are clear structural reasons for this. Existing platforms already have the critical necessities for agents to operate in a real environment: access rights, audit logs and action logging. In addition, they control the system context without which an agent cannot operate safely and predictably. Therefore, it is easier for such companies to add their own agent interface than to risk users taking their work to an external agent layer.
The downside is a strict economic constraint. Each agent request has its own price. The more internal agents you promise, the greater the pressure on your monetization model. Corporate clients rarely want to switch overnight from per-seat/subscription pricing to significantly higher pay-per-transaction pricing.
Therefore, many suppliers will try to “hide” the capabilities of agents within the framework of their usual pricing policy for as long as possible. At least until the economics of unit production force a clearer compromise via the underlying cost structure of operating agent systems. Each interaction with an agent involves variable costs for inference, computation and coordination. As usage scales, these costs become material, and providers can no longer absorb them indefinitely within a fixed subscription fee. At that point, suppliers will be forced to make the trade-off explicit: introduce usage-based pricing, limit agent functionality or restructure their monetization model.
Who Survives and Why: 5 Types of SaaS Defensibility
The logic behind this protection seems quite pragmatic. SaaS is least vulnerable to the “SaaSpocalypse” when it performs one of the basic infrastructure functions.
1. Integration Hubs
The product acts as an integration hub and orchestration layer: it supports reliable connectors, manages dependencies between dozens of external systems, access rights, restrictions and SLAs. An AI agent can simplify working with this complexity, but it is much more difficult for it to completely replace such a layer.
2. Systems of Record With High Cost of Error
Finance, HR, security, healthcare, law and billing are areas where complete traceability of actions is critical. In such systems, AI agents typically serve as a convenient interface on top of the product, while the core accounting and control functions remain within the platform itself. Too much here is tied to responsibility and the consequences of errors to completely remove control outside the system.
3. Data Moats Plus Feedback Loops
If a product relies on unique data or history that improves results over time, simply adding an LLM will not reproduce this quality. Without this layer, the application quickly turns into a simple search.
A good example is Sprouty. At first glance, it may look like just another app for parents. Its key value, however, lies in the long-term home context that users themselves create. This includes a personal daily routine and all the changes that occur with the child. Even home pediatricians do not have this level of detail.
Therefore, a universal agent without access to this context can only give general advice; essentially, it becomes a more convenient form of search. A product that works with accumulated history and built-in feedback mechanisms provides a more accurate understanding of the situation and possible scenarios. At the same time, it remains within the specified limits of responsibility. Sprouty, for example, is not a medical device and does not provide diagnostics.
4. Narrow Domain Expertise and Regulation
We are not just talking about a CRM, but, for example, a dental CRM with clinical processes, integration with insurance companies, regulatory requirements and industry logic. In such niches, a superficial agent without a deep subject model will quickly run into reality. Here, value is created by a combination of expertise, processes and regulatory compliance. An agent can only complement this layer, not replace it.
5. High Switching Costs
Systems that are deeply customized for specific processes and have accumulated historical data do not disappear just because a more convenient interface has appeared. The cost of transition simply changes form. Even if an agent simplifies the work, abandoning an existing solution still requires time and organizational effort, which for many companies is critical.
The ‘SaaSpocalypse’ Is a Catchy but Inaccurate Frame
In reality, there is no need to panic. The reality we currently find ourselves in is quite pragmatic. The agent layer tends to filter out products that are essentially complex business processes disguised as a convenient interface.
Certain verticals may indeed be displaced. But in most cases, agents act differently: They narrow the market, leaving those who are able to adapt and whose products are built around a sustainable core rather than a set of clicks.
