CHOSUNBIZ framed the change as a looming “SaaSpocalypse,” arguing that agents can perform tasks that previously required users to operate CRM, design, HR and collaboration software directly. The reporting identifies a real commercial pressure: if an employee can ask an assistant to file an expense, update a customer record, edit an image or assemble a report, the software vendor risks becoming invisible behind the AI provider’s interface. But the companies cited in the report are largely responding by making their own systems the action layer that agents call into—not by abandoning the SaaS model.
For Windows and enterprise IT teams, that distinction is consequential. An agent that can take action across Microsoft 365, Salesforce, Workday, Adobe or Atlassian is not simply a better search box. It is another privileged automation client touching identity, corporate data, approval chains and audit trails.
The seat license is no longer the whole transaction
Salesforce is the clearest example of the pricing transition. The company has positioned Agentforce as a way to perform CRM work rather than merely assist a licensed employee, and it has introduced Agentic Work Units as a way to measure completed agent work. Salesforce said in its fiscal 2027 first-quarter disclosures that Agentforce reached $1.2 billion in annual recurring revenue, up 205% year over year.
That is meaningful growth, although it should not be confused with evidence that Salesforce’s core business has been displaced by agents. Agentforce runs on the customer data, application logic, permissions and workflows that made Salesforce valuable before generative AI arrived. The agent produces revenue precisely because the underlying CRM remains the controlled place where customer information, business rules and sales processes live.
Salesforce’s own disclosures also demonstrate why task counts need careful reading. Its more recent figures refer to billions of Agentic Work Units across Agentforce and Slack, which are measurements designed by Salesforce rather than a generally comparable industry standard. A completed work unit might indicate a useful automation, but it does not tell an IT buyer how many tasks were approved without human intervention, how many required correction, or how much business risk was transferred to the model.
The commercial logic is still clear. Per-user SaaS pricing rises only when a company buys more seats. An agent can generate more consumption from the same workforce by handling routine work at all hours. That gives vendors a reason to sell execution capacity alongside subscriptions, and gives buyers a new variable cost to monitor.
For procurement teams, the practical implication is that an AI deployment can reduce friction for end users while adding a usage-based bill that grows outside the normal headcount model. License true-ups were already familiar; AI work-unit consumption, model calls and agent-run workflows are becoming a second meter.
Adobe is defending the application layer by entering ChatGPT
Adobe’s ChatGPT integration is more accurately described as distribution expansion than a retreat from creative software. Adobe had already put Photoshop, Acrobat and Adobe Express into ChatGPT in December 2025. Its newer unified Adobe plug-in exposes more than 70 tools spanning products including Photoshop, Premiere, Acrobat, Lightroom, Illustrator, InDesign, Firefly and Adobe Stock, according to Adobe and independent coverage from Tom’s Guide and Digital Camera World.
The point is not that a chatbot has suddenly replaced Photoshop. The point is that Adobe has accepted that many users will start a creative task in a conversational interface. Adobe’s answer is to let natural-language requests invoke its tools, then bring users into Adobe’s applications when they need review, adjustment or more advanced control.
That strategy matters in the Microsoft environment because Adobe is making the same push into Microsoft 365 Copilot, Claude, Gemini and Slack. Instead of requiring workers to launch a dedicated application first, Adobe is trying to place its capabilities wherever work begins. For a traditional desktop-software company, that is an attempt to retain the application, file-format and services relationship even if the front door changes.
There is a governance cost to that convenience. A prompt that edits a file or generates a document may look like an ordinary chat interaction, while the downstream action can involve proprietary images, contracts, customer materials or regulated records. Organizations enabling these connectors should treat them as application integrations, not as harmless consumer AI features.
At minimum, administrators need to establish which connector is approved, what Adobe account or enterprise tenancy it uses, whether uploaded content is retained or used for model-related processing, and whether the action history is captured in an audit system. If an organization has rules around Creative Cloud Libraries, SharePoint, OneDrive, Microsoft Purview or data-loss prevention, those rules should be evaluated against the actual connector path rather than assumed to follow the same controls as the standalone application.
Workday’s Copilot connection shows why the system of record still matters
The stronger rebuttal to the claim that AI will make enterprise applications irrelevant comes from Workday. In May, Workday announced that its Sana Self-Service Agent would be available in Microsoft 365 Copilot, allowing employees and managers to obtain HR and finance answers or complete routine tasks without leaving Microsoft 365.
Workday’s pitch is revealing: Sana can act through Workday’s existing security, permissions, business rules and approval mechanisms. That means Microsoft 365 Copilot may be the employee-facing interface, but Workday remains responsible for the authoritative employee, finance and workflow data. The agent is useful only if it is constrained by the system of record rather than allowed to improvise around it.
This is the model enterprise software vendors are converging on. The chatbot or assistant becomes a control surface; the SaaS product supplies the permissions, data model, transaction rules and auditability needed to safely complete an action. Workday has also tied Sana into Google’s Gemini Enterprise, underscoring that vendors are not betting on one front-end platform winning every customer.
For Microsoft administrators, this turns the question from “Should we allow AI agents?” into “Which agents can perform which actions under which identities?” An agent able to change an employee’s details, submit an expense, retrieve payroll information or trigger a finance workflow needs least-privilege access, strong conditional-access policies, scoped connector permissions and traceable approval boundaries. A conversational prompt is not a substitute for segregation of duties.
This also creates a support challenge. When an employee says that Copilot “made a change,” IT needs to determine whether Copilot interpreted the request, Workday’s agent selected the action, a workflow approved it, or an underlying enterprise application executed it. Logging has to preserve that chain.
Airtable’s sale does not prove a 90% SaaS collapse
CHOSUNBIZ uses Bending Spoons’ agreement to acquire Airtable as evidence that AI has devastated SaaS valuations. The deal is real: Bending Spoons announced on August 4 that it had agreed to acquire Airtable in an all-cash transaction. Axios independently reported the transaction at about $1.29 billion.
But the comparison in the submitted report is materially misleading. Bending Spoons said the deal assigns Airtable an enterprise value of $1.285 billion, while Airtable’s cash and cash equivalents imply an equity value of roughly $2.25 billion. Comparing the $1.285 billion enterprise-value figure with Airtable’s approximately $12 billion 2021 private-market valuation mixes two different measurements.
The correct like-for-like comparison is still steep: a $2.25 billion equity value is roughly 81% below a $12 billion valuation. Yet it is not the roughly 90% collapse suggested by comparing enterprise value with equity value. More importantly, neither figure isolates AI as the cause. Private-market valuations in 2021 reflected unusually cheap capital, aggressive growth expectations and a very different interest-rate environment. The record supports a major repricing of Airtable; it does not establish that AI agents alone caused it.
Bending Spoons’ announcement also complicates the extinction narrative. It said Airtable had annual recurring revenue of about $480 million as of June 2026, growing more than 20% year over year. A profitable acquirer paying for a company with hundreds of millions in recurring revenue is a sign that SaaS cash flows still have value, even when earlier venture valuations do not survive.
What enterprises should measure before calling it an AI transformation
The “SaaSpocalypse” label is useful as a warning to vendors with shallow products or weak customer data, but it obscures the operational reality. Software is becoming more exposed to AI intermediaries, and vendors that only provide a narrow interface may lose pricing power. Vendors that own trusted records, permissions, workflow state and deep integrations have a better chance to charge for agent-driven execution.
Enterprise customers should resist evaluating these products only by demo quality or the number of tasks an agent claims to complete. The more useful tests are concrete:
- Determine whether the agent acts through the application’s native permissions and approval workflow, or relies on a broad service account with excessive access.
- Identify the meter that will drive spending, including seats, AI credits, model usage, agent work units and third-party connector charges.
- Require logs that distinguish a user request, model interpretation, agent decision, application action and human approval.
- Test failure handling on routine but sensitive cases before allowing agents to alter customer, employee, financial or production data.
- Preserve an ordinary application workflow for exceptions, disputes and recovery when an agent makes the wrong call.
The near-term consequence is not the disappearance of Salesforce, Adobe, Workday or Microsoft-connected enterprise software. It is that their value will increasingly depend on whether they can let agents act safely on their data without surrendering the customer relationship—or the audit trail—to the chat interface.
Update: Additional details (August 17, 2026)
CHOSUNBIZ adds Gartner’s estimate that as much as $234 billion in enterprise application-software spending could face “agentic arbitrage” through 2030, as agents perform work across multiple systems and reduce reliance on separate human-facing interfaces. Gartner’s framing is that SaaS is being disaggregated and reshaped rather than eliminated.
The report also puts Salesforce’s delivered volume at 3.8 billion Agentic Work Units across Agentforce and Slack, and notes that Salesforce’s fiscal 2027 reporting now groups revenue under “Agentforce Apps” and “Data 360, Headless Platform, & Other.” Workday’s Sana Self-Service Agent reportedly launched with more than 300 skills, while Adobe’s Claude connector was described as offering more than 50 tools.
