Salesforce’s stock dropped noticeably after their last earnings call — not because their numbers were bad, but because analysts started asking uncomfortable questions about what happens to CRM software when an AI agent can just do the CRM work instead of organizing it. That question is now being asked about nearly every SaaS category. The threat isn’t that AI companies are building competing software. It’s that the entire model of selling a seat license to a tool that helps humans do a job is getting undercut by systems that can do the job directly. This isn’t a future concern. It’s happening now, and the companies that built the last decade of enterprise software are having to reckon with it in real time.
The Core Problem: SaaS Was Built Around Human Limitations
Think about why SaaS tools exist in the first place. Salesforce exists because salespeople need somewhere to track their deals, log their calls, and forecast their pipeline — tasks that require human attention and coordination. Zendesk exists because support tickets pile up and someone has to manage the queue. Asana and Notion exist because teams struggle to keep track of who’s doing what. Every major SaaS product is essentially a well-designed interface for managing the friction created by humans working together.
The implicit assumption baked into every one of those products is that there’s a human on the other end who needs a good UI, good notifications, and good reporting to stay on top of their work. Pull that assumption out, and the entire product architecture starts to look different. You don’t need a kanban board if an agent is handling the tasks. You don’t need a ticket routing system if an agent is triaging, responding, and escalating without anyone touching a dropdown menu.
This is why Andreessen Horowitz and other major funds have been openly discussing what they’re calling the “per-seat apocalypse” — the structural risk to software businesses that charge by the number of humans using a product. When agents replace workers or dramatically reduce the number of humans involved in a workflow, the seat count shrinks. Revenue shrinks with it.
Where the Disruption Is Already Visible
The clearest early signal came from customer support. Companies like Intercom were already moving fast to build AI into their products, but startups like Sierra (founded by former Salesforce co-CEO Bret Taylor) are building AI customer service agents from scratch — not AI features bolted onto a help desk, but agents designed to resolve issues end-to-end without human escalation. The pitch isn’t “this will help your support team work faster.” It’s “you’ll need a smaller support team.”
Sales development is another category getting hollowed out quickly. The entire SDR software stack — tools like Outreach, Salesloft, and Apollo — was built to help humans send better cold emails, sequence follow-ups, and manage outbound pipelines. AI agents can now do all of that autonomously. Startups like 11x are building “AI SDRs” that prospect, research, personalize, and follow up without a human in the loop. The question for Outreach isn’t whether their product is good. It’s whether there’s a job left to augment.
Legal work is another front. Clio, the dominant practice management platform for law firms, built a strong business helping lawyers manage cases, time tracking, and billing. But as tools like Harvey AI (which now has serious traction with large law firms) handle the substantive legal work — drafting, research, document review — the practice management layer becomes thinner. Lawyers still need billing software, but you need fewer lawyers, so you need fewer seats.
Code is perhaps the most visible example. GitHub Copilot was the first major signal, but it was still fundamentally an autocomplete tool requiring a developer to steer it. What’s happened since — with tools like Cursor becoming the default IDE for a large chunk of developers, and agentic coding systems like Devin and its successors handling multi-file, multi-step engineering tasks — suggests that the number of developer seats needed to ship software is actually declining. Not dramatically yet, but directionally, clearly.
The Three Patterns of Disruption
Not every SaaS category is being disrupted the same way. It’s worth being precise about the different patterns, because they have different timelines and different implications for incumbents.
- Full task displacement: The agent doesn’t augment the worker — it replaces the workflow entirely. Customer support is the clearest example. An agent that can handle 80% of tier-1 support tickets end-to-end doesn’t make the support team more productive; it shrinks it. The SaaS tool built to help that team disappears with the headcount.
- Workflow compression: The job still exists, but it takes far fewer people. A marketing team that used to need five people to run campaigns — one for copy, one for design coordination, one for analytics, one for email ops, one for reporting — might run the same output with two people using AI tools. The per-seat SaaS revenue from that team drops 60%, even if the surviving two users are paying for premium tiers.
- Interface elimination: The agent interacts directly with underlying systems via APIs, bypassing the SaaS UI entirely. If an AI agent can write directly to your CRM’s database, generate reports from raw data, and trigger workflows through API calls, the human-facing interface becomes optional. This is the most subtle threat and potentially the most dangerous for incumbents, because it doesn’t require replacing the software — just routing around the parts that require human interaction.
How Incumbents Are Responding (And What’s Actually Working)
The incumbent response has ranged from genuinely smart to transparently defensive. Salesforce’s Agentforce is the most prominent attempt by a legacy SaaS company to get ahead of this. Rather than waiting for agents to eat their lunch, they’re building the agent layer themselves — trying to position Salesforce not as a tool humans use, but as a platform agents run on. The pitch is that enterprise customers already trust Salesforce with their data, their workflows, and their integrations, so they should trust Salesforce-native agents to act on that data. It’s a coherent strategy, and Marc Benioff has been aggressive about pushing it publicly.
Whether it works is a different question. The challenge is that Salesforce’s existing product architecture wasn’t designed for agents. The UI-first assumptions are baked deep. Building an agent layer on top of a system designed for human navigation is harder than building an agent-native system from scratch, and startups aren’t carrying that legacy weight.
ServiceNow has made a similar bet, leaning hard into agentic automation for IT and HR workflows where they already have strong data and process ownership. Their advantage is that enterprise workflows in IT service management are well-defined enough that agents can execute them reliably — the tasks are structured, the outcomes are measurable, and the blast radius of errors is contained. It’s a smarter beachhead than trying to build general-purpose agents.
Microsoft is in the most interesting position. They’re both an incumbent SaaS player (Office 365, Teams, Dynamics) and a major AI infrastructure provider through their OpenAI relationship. Copilot for Microsoft 365 started as an augmentation play, but the direction is clearly toward agents that take actions across the Microsoft ecosystem. If that succeeds,