For twenty years, software investors underwrote a simple assumption: software scales better than labor. SaaS sold access to reusable code. Add another customer, seat, or department, and the marginal cost was close to zero. That was the economic magic behind 80–90% gross margins, high net retention, and the valuation framework that built modern enterprise software.
AI agents complicate that assumption. They are software, but they behave economically like labor. Every completed task carries a cost: inference, tool calls, orchestration, retrieval, verification, retries, exception handling, and sometimes human escalation. In traditional SaaS, usage was mostly evidence of retention. In agentic AI, usage is also a cost event.
The obvious story is that AI agents will replace seats. Why pay for fifty users when one person with agents can do the same work? That story is directionally right, but it misses the harder question. Replacing seats does not automatically create a better business model. A seat is predictable revenue. Work is variable cost.
That is the agent margin trap. Outcomes-based pricing sounds elegant: pay per resolved ticket, qualified lead, invoice processed, or claim reviewed. The customer only pays when value is delivered. But alignment is not the same as margin. An agentic AI company that sells more outcomes also takes on more work, more model calls, more edge cases, more failed attempts, and more dispute risk over whether the outcome was actually achieved.
The bull case is that inference costs will fall. They will. But inference is only the visible cost of autonomy. The hidden costs are verification, exception handling, trust, and liability. Those do not automatically collapse with GPU prices. In many workflows, they become the real cost center.
This is why the key diligence question for agentic AI companies is changing. It is not how much ARR they have, how many workflows they automate, or how accurate the model is. The real question is: what is the contribution margin per outcome after inference, tools, orchestration, verification, exceptions, and human escalation?
If that number does not improve with scale, the company is not AI-native SaaS. It is tech-enabled services with a better interface.
The best agentic AI companies will operate in domains where the work is frequent, narrow, measurable, repeatable, and low-liability. Customer support resolution, invoice processing, compliance documentation, sales research, data enrichment, claims triage, and recruiting screening are good examples. In these markets, success can be defined, failure can be measured, exceptions can be reduced, and margins can improve.
The weakest agentic AI companies will operate where the work is ambiguous, judgment-heavy, high-liability, and hard to verify. Strategy, complex legal advice, medical decisioning, financial planning, enterprise sales, and executive recruiting may demo well, but scale poorly. The edge cases become the product. The exception queue becomes the company.
Investors are trying to value AI labor with SaaS multiples. But labor and software are different economic objects. Software scales by replication. Labor scales by execution. SaaS monetized access. Agents monetize work.
The next great AI companies will not be the ones that automate the most work. They will be the ones that automate the most profitable work.
AI does not kill SaaS. It forces us to ask a better question: is this software, or is this labor with an API?