When AI Agents Replace SaaS

The most-discussed AI prediction of 2025 was Sam Altman's offhand comment that "agents will eat SaaS." Whether or not Altman meant it as a serious forecast, the prediction has shaped a wave of agent-shaped startups and a corresponding wave of SaaS company anxiety. The honest version is more nuanced. Some categories of SaaS are genuinely vulnerable. Others are largely safe. The work is in identifying which is which.

This piece is a structured assessment for builders, founders, and SaaS operators trying to figure out where they sit.

The core claim

The argument for "agents replace SaaS" runs roughly:

A SaaS product is, abstractly, a structured interface to a database (its CRUD layer) plus some business logic plus some integrations plus a UI. An LLM agent with tools can, in principle:

• Query and update databases via SQL or API tools
• Execute business logic via code execution
• Integrate via tool calls to external APIs
• Present output as text/voice/UI in a conversational interface

If all of that works reliably, then much of what a SaaS product offers, the structured interface, becomes redundant. The user just talks to an agent that understands the underlying data and the desired action, and the agent handles the rest.

This argument has been made about every "no-code" tool wave since the 1990s, and historically the wave has died out for one reason: most users actually want structured interfaces. Conversational interfaces are slower, less precise, and harder to scale across an organization. The 2025-2026 generation of agents is genuinely better than previous waves, but the structural argument hasn't fundamentally changed.

What's true is that some of SaaS is replaceable. The question is which categories.

Categories that are genuinely vulnerable

Three SaaS categories where agents will eat meaningful market share through 2027.

Internal data-aggregation tools

Mid-market SaaS that aggregates data across other systems for internal reporting, basic BI dashboards, internal status pages, "show me this metric across these systems" tools, are genuinely vulnerable. The agent shape is well-suited: ask a natural-language question, the agent queries the underlying systems, returns a structured answer or a generated dashboard.

This category includes a lot of internal tools that companies build themselves. The marginal cost of "build an agent that does this" is dropping fast enough that custom internal tooling will increasingly be agent-shaped rather than dashboard-shaped.

Examples of vulnerable products: internal-only BI dashboards, sales-rep performance reporting, custom HR analytics. Less vulnerable: external-facing analytics, products with strong vertical-specific features.

Form-and-workflow SaaS

A lot of SaaS is, structurally, a form: enter data, route it through some approval process, output a structured result. Expense reports, time-off requests, simple ticketing systems, basic CRM data entry, all of these are mostly forms with workflow logic.

Agents can handle this category natively. "I need to expense $50 for client lunch" → agent extracts the data, validates against policy, routes for approval, updates the system. The structured form interface becomes an option, not a requirement.

Examples of vulnerable products: simple expense management, basic ticketing, simple CRM data entry, simple project management. Less vulnerable: complex multi-stakeholder workflow tools where the structure of the form is the value.

Integration glue

Zapier, Make, n8n, Workato, products whose value is moving data between systems based on triggers. This category is squarely in the agent-shaped path: agents naturally bridge systems through tool calls, and the imperative "integration glue" pattern is exactly what agents do well.

Zapier's response has been to add AI features and become an agent platform itself. Whether that pivot succeeds depends on execution, but the structural pressure is real.

Examples of vulnerable products: integration-as-a-service platforms, simple workflow automation. Less vulnerable: deep vertical-specific integration tools (e.g., healthcare-specific HL7 integration).

Categories that are mostly safe

Three categories where SaaS retains structural advantages over agents.

Complex collaborative interfaces

Figma, Notion, Linear, Slack, Google Docs, products where the value is many people interacting simultaneously through a shared, structured interface. Agents can supplement these (and increasingly do) but can't replace the underlying collaboration substrate. Many people typing in the same document need the document's structure; an agent that mediates everyone's edits would be slow and lossy.

These products will incorporate AI features extensively. They won't be replaced by agents.

Domain-specific specialized workflows

Medical records, legal case management, banking core systems, ERP, products where the domain-specific structure encodes years of regulatory, legal, or operational requirements. The structured interface isn't just convenience; it's compliance. An agent that "interprets" what should be a structured field is a liability.

These products will adopt AI for specific assistive features (clinical decision support, contract review). They won't be replaced.

High-throughput operational tools

Salesforce-as-CRM-of-record, ServiceNow-as-ticket-of-record, Workday-as-HRIS-of-record. These products are the systems-of-record for huge enterprises. The value is the durable structured representation of the business, not the interface. Even if agents become the primary user interface, the underlying SaaS product remains the data backbone.

These products will increasingly expose agent-friendly APIs and be queried by agents. They won't be replaced.

The middle ground

A lot of SaaS sits in a middle ground, partially vulnerable, partially safe. Three patterns:

SaaS with strong moats around proprietary data is mostly safe. If your product owns data that the agent can't get elsewhere (a marketplace's transaction data, a network's relationship graph), agents need your product as a tool, not as a replacement.

SaaS where the workflow is the value is mostly safe. If users specifically want to follow a structured process (compliance workflows, audit trails, multi-stakeholder approvals), the agent interface is a less-good fit than the structured one.

Long-tail vertical SaaS is mixed. Many small vertical SaaS products will be replaced by general-purpose agents fine-tuned to the vertical. But the leaders in any given vertical typically have moats (relationships, data, workflows) that the agent doesn't.

What this means for SaaS founders

Three planning implications.

Audit your value proposition. Is your product's value the structured interface, the data, the workflow logic, or the integrations? Products whose primary value is "we provide a nice form for entering this data" are vulnerable. Products whose primary value is data, workflow rigor, or system-of-record status are safer.

Become AI-native, not AI-decorated. Adding a chatbot to a SaaS dashboard is not a moat. Re-architecting your product around AI primitives (semantic search, agent-friendly APIs, natural-language operations) is. The companies that survive the agent wave are the ones who treat agents as the primary user interface their product has to support, not a feature.

Think about distribution. Even if your product survives technically, the distribution layer is shifting. Users increasingly enter products via AI assistants (ChatGPT, Claude, Gemini) rather than directly. Make sure your product is discoverable and usable through those entry points.

What this means for builders

Three opportunities.

Build agent-first replacements for vulnerable SaaS categories. The form-and-workflow category is a green field. Existing players (Zapier, Notion, Salesforce) are competing for incremental AI features; agent-first replacements compete on a different axis.

Build the picks-and-shovels for agent-as-interface. The infrastructure layer for agents (memory, tool integration, observability) is still maturing. LangChain, MCP servers, agent-evaluation frameworks, this layer has commercial space.

Build vertical agents. Generic agents are limited because they don't understand specific domains. Vertical agents (healthcare, legal, finance, real estate) trained on domain data and integrated with domain systems will eat parts of vertical SaaS. The opportunity window is roughly 2-3 years before vertical SaaS catches up with their own AI.

Timeline

Three predictions:

By end of 2026: AI assistants will be the primary entry point for ~10-15% of digital work. Form-and-workflow SaaS sees noticeable churn at the low end. Integration glue (Zapier, etc.) under serious pressure.

By end of 2027: ~25-30% of digital work mediated through AI assistants. SaaS revenue growth meaningfully slows in vulnerable categories. New agent-first products dominate certain niches.

By end of 2028: AI assistants are the default entry point for most knowledge work. SaaS as a category is significantly different, tools that survive are AI-native. Several major SaaS companies have either pivoted or been acquired.

These are speculative. The pace of change has consistently exceeded estimates over the past three years.

The deeper takeaway

"Agents replace SaaS" is overstated as a general claim and accurate in specific categories. The work for SaaS founders is to identify whether their product is in a vulnerable or safe category, and to plan accordingly. The work for builders is to identify the categories where agent-first replacements have a window, and to ship into them before incumbents catch up.

The sober answer is that 2026-2028 will see a lot of churn at the low and middle of the SaaS market, and remarkably little disruption at the top. The biggest SaaS companies in 2028 will mostly be the biggest SaaS companies of 2025, with substantially better AI features bolted on.

Further reading

• Best LLMs for AI agents
• The state of coding agents in 2026
• The AI tool procurement checklist
• MCP servers actually changed things