Will AI Agents Replace SaaS? The Unbundling Thesis Explained

Jefferies traders coined a term in early 2026: "SaaSpocalypse." Microsoft's Satya Nadella told audiences that business applications as we know them could "collapse" in the agentic AI era. Marc Andreessen declared 2025 the year of AI agents and started talking about "unbundling the BPO."

The narrative is everywhere: AI agents will eat SaaS alive. Point solutions are toast. The future is a single intelligent layer that replaces your entire software stack with natural language.

It's a compelling story. It's also not quite right — and the truth is more interesting than the hot takes suggest.

The Unbundling Thesis, Explained

Jim Barksdale, the former Netscape CEO, once said: "There are only two ways to make money in business — bundling and unbundling." That observation has driven tech strategy for decades.

Here's the short version of what happened:

• 2000s–2010s: SaaS unbundled monolithic enterprise suites. Instead of one SAP installation handling everything, companies bought Salesforce for CRM, Workday for HR, Stripe for payments, and dozens of other point solutions.
• 2020s: AI agents are now rebundling all of it. Instead of switching between fifteen apps, you describe what you want in plain English and an agent orchestrates across all of them.

The thesis, popularized by VCs and industry analysts, goes like this: SaaS apps are fundamentally just CRUD databases with business logic and a UI. AI agents can absorb the business logic, skip the UI entirely, and operate directly on the data. Why pay for twenty separate interfaces when one agent layer can do the work?

This isn't purely theoretical. Nadella himself framed it this way — all the business logic moves into the AI tier, operating across multiple databases seamlessly rather than being hardcoded into individual applications. The app becomes optional. The agent becomes the interface.

Where Agents Are Already Winning

The money tells a story. AI agent startups raised $3.8 billion in 2024 — nearly triple the year before. The agentic AI market is projected to hit $199 billion by 2034. And specific verticals are already seeing real disruption.

Coding tools are the clearest example:

• Cursor hit $500M+ ARR with over a million users, doubling revenue every two months through 2025. Its parent company Anysphere raised $2.3 billion at a $29.3 billion valuation.
• Windsurf was acquired by OpenAI for $3 billion.
• GitHub Copilot has 1.8 million paid subscribers and 77,000+ enterprise customers.

These aren't autocomplete plugins anymore. They're autonomous agents that reason across entire codebases, write features end-to-end, and handle multi-file refactors.

Customer support is the second front:

• Intercom's Fin resolves up to 50% of support questions instantly with a 96% answer rate for multi-source queries.
• Zendesk's AI agents handle up to 80% of questions autonomously across 80+ languages.
• 75% of CX leaders expect the vast majority of interactions to be resolved without humans within a few years.

Legal might be the most dramatic. Harvey AI raised $760 million in 2025 alone across three funding rounds, jumping from a $3B to $8B valuation in a single year. It now serves 50+ of the AmLaw 100 firms and hit $190M ARR.

Then there are the vertical agent operating systems — startups building AI-native replacements for entire industry software stacks. Sandra AI runs dealership operations end-to-end. Paratus Health handles everything from front-desk calls to billing for outpatient clinics. Brickanta does bid analysis, cost estimation, and procurement for construction. These aren't workflows bolted onto existing tools — they're ground-up replacements.

The Three Things Agents Can't Replace (Yet)

Before you short every SaaS stock, consider what the unbundling narrative conveniently ignores.

Deterministic consistency

LLMs are probabilistic. They give you a great answer six out of ten times and a subtly wrong one the other four. That's fine for drafting emails. It's unacceptable for:

• Running payroll — you can't be "approximately correct" on tax withholdings
• Regulatory compliance — HIPAA, SOC 2, and GDPR demand deterministic audit trails
• Financial reporting — numbers need to match, every time, with zero hallucination

Critical business processes need guaranteed correctness, not statistical likelihood. As Bain's analysis notes, this is the fundamental constraint that keeps SaaS relevant.

Systems of record

Here's the part the "agents replace everything" crowd glosses over: agents need software to operate on. They need databases to query, APIs to call, and systems of record to read from and write to. An AI agent that handles customer support still needs a ticket system, a knowledge base, and a CRM underneath it.

As Fortune put it bluntly: "AI agents aren't eating SaaS — they're using it."

Decades of encoded complexity

Enterprise SaaS isn't just CRUD with a pretty UI. It's years of domain expertise baked into code — handling malformed EDI files from suppliers, managing multi-jurisdiction tax rules, supporting complex multi-party approval chains. An agent can't magically replicate what took Workday or SAP decades to build and battle-test.

It's highly unlikely that Fortune 500 companies will want to rebuild their own bespoke HR or supply chain software from scratch just because an LLM can generate a nice UI.

What's Actually Happening: The Stack Is Shifting

The real story isn't "agents kill SaaS." It's that the software stack is reorganizing into three layers:

Layer	What It Does	Examples
Systems of record	Store data, enforce constraints, maintain compliance	Salesforce, Workday, PostgreSQL
Agent orchestration	Understand intent, plan actions, call tools across systems	Claude, GPT, custom agent frameworks
Outcome interfaces	Natural language, proactive suggestions, autonomous execution	Chat UIs, Slack bots, voice agents

SaaS isn't disappearing. It's moving down the stack. The database and business logic layers remain essential — what's getting disrupted is the UI and workflow layer that sits on top.

This is why smart SaaS companies are racing to become agent-friendly infrastructure. Salesforce launched Agentforce. ServiceNow is building agent capabilities. Gartner predicts 40% of enterprise apps will feature task-specific AI agents by the end of 2026, up from less than 5% in 2025.

They're not fighting agents. They're becoming the plumbing agents run on.

So, Will Agents Replace SaaS?

Not exactly. But they'll hollow it out.

Here's the verdict in three lines:

• Agents will eat the UI and workflow layers. Nobody will manually navigate dashboards when they can describe what they need.
• SaaS becomes infrastructure. The companies that survive are the ones that make themselves indispensable to agents — great APIs, reliable data, robust tool-calling interfaces.
• The real losers are horizontal point solutions with thin moats. If your entire product is a nice interface on top of a simple workflow, an agent can replicate that overnight. If you own the data and the compliance layer, you're safe.

Deloitte's 2026 predictions frame this as a 5–10 year evolution, not a 2-year revolution. Gartner has warned that agentic AI supply already exceeds demand.

The SaaSpocalypse makes for a great headline. The reality is a slow, structural reshuffling — and the companies paying attention are already adapting. The ones pretending nothing has changed are the ones that should worry.