Essay

SaaS Sold Access. ANSOs Sell the Work.

Why agent-native software organizations change pricing, margins, revenue per employee, and the unit of value in software.

June 7, 2026

Software economics used to be easier to explain.

You built a product once. You sold access to many customers. The customer paid by seat, usage, or subscription tier. The software helped people work faster, but people still produced the deliverable.

That model rested on a stable boundary: software coordinated work. Humans performed it.

ANSOs move that boundary.

An Agent-Native Software Organization is a company where agents become the primary execution layer. The system produces the thing the customer is paying for: the legal draft, the clinical note, the resolved support case, the booked freight load, the first usable software feature.

That distinction sounds technical.

It is economic.

The Seat Is the Wrong Unit

Traditional SaaS sold tools attached to workers: CRM for salespeople, help desks for support teams, project management software for operators.

That is why per-seat pricing made sense. More workers meant more seats. More seats meant more software spend. The pricing model mapped cleanly to the org chart.

Imagine a 100-person support team buying a $100-per-seat monthly SaaS product.

That is $10,000 per month, or $120,000 per year.

Useful business. Clean software margin. Easy procurement category. But the software is still priced as an accessory to labor.

Now imagine an agentic support system that absorbs 30 percent of the work handled by that same team. If each support employee costs $80,000 fully loaded, the labor pool is $8 million. Thirty percent of that workload represents $2.4 million of annual labor capacity.

That changes the buyer's reference point.

The product is no longer being compared only to a $120,000 software line item. It is being compared to labor, outsourcing, response time, quality, and operating capacity.

The seat stops being the atomic unit of value.

The work is.

The Pricing Ceiling Moves

This is the simple ANSO pricing formula:

ANSO price ceiling = labor capacity absorbed x buyer's willingness to share savings

If software absorbs $2.4 million of annual labor capacity, a $600,000 contract is not expensive SaaS. It is a 75 percent savings trade for the buyer.

That is the difference between helping someone work and doing part of the work itself.

In AI-augmented software, the customer buys assistance.

In an ANSO, the customer buys performed work.

A legal AI system that helps an attorney research faster is selling productivity. A legal AI system that produces the first draft under attorney review is selling labor absorption. Both may be valuable. They do not have the same economic ceiling.

This is where many founders will underprice agent-native products. They will charge like SaaS because SaaS is the model they know, even when the buyer is evaluating the product against headcount.

Gross Margin Is No Longer the Whole Story

The first generation of SaaS trained everyone to worship gross margin.

That made sense. At 85 percent gross margin, $1 million of revenue creates $850,000 of gross profit.

ANSOs complicate that picture.

Agent-native companies may carry real delivery costs: inference, orchestration, human review, compliance, implementation, and edge-case handling.

But lower gross margin can still produce better economics if the company captures a larger pool of value.

A legacy SaaS product sells a $1 million contract at 85 percent gross margin. It creates $850,000 of gross profit.

An ANSO sells a $1.5 million contract that replaces or absorbs $3 million of labor capacity. Even at 65 percent gross margin, it creates $975,000 of gross profit. The buyer still saves $1.5 million versus the old operating model, and the vendor generates more gross profit than the cleaner-looking SaaS product.

That does not mean gross margin stops mattering.

It means gross margin is no longer the whole story.

The better questions are: how much human work does the system absorb, how much shifts toward software over time, does model progress reduce delivery cost or expand output quality, and can the company price against outcomes rather than access?

Revenue per Employee Becomes a Clue

Revenue per employee is imperfect. It can be distorted by contractors, outsourcing, timing, and accounting choices.

Still, in the agent-native era, it points to the right question: how much work can a company produce without adding proportional headcount?

The early numbers are hard to dismiss.

Cursor reportedly passed $2 billion in annualized revenue with roughly 150 employees. That implies about $13.3 million of revenue per employee.

Lovable crossed $400 million in ARR with 146 employees. That implies about $2.7 million per employee.

Harvey is different because legal work is high-trust and review-heavy. Even there, the company has been reported around $190 million in ARR, serves more than 100,000 legal professionals across 1,300 organizations, and says customers have created 25,000+ custom agents.

These companies are not identical. Cursor and Lovable sit closer to agentic software creation. Harvey is a cleaner example of agent-produced work product under human review.

But they rhyme.

They point toward a company shape where output scales through agents, workflow systems, and review loops rather than only through hiring.

That is not just a management observation.

It is an economic one.

The Moat Moves Into the Workflow

The obvious objection is that model capability will commoditize.

That is partly right.

The weak ANSO answer is, "our prompts are better."

That will not be enough.

The stronger answer is workflow ownership.

The more an ANSO sits inside the customer's real operating process, the more it accumulates proprietary workflow data, review histories, edge-case handling, integrations, compliance posture, trust boundaries, and customer-specific operating logic.

Harvey is interesting because legal work is precedent-sensitive, review-heavy, and embedded in professional accountability. Abridge is interesting because clinical documentation sits inside Epic, billing, liability, physician time, and hospital operations.

In these markets, the model is necessary.

It is not the whole company.

The moat is the system that turns model capability into trusted work product.

The Pitch Has to Change

If ANSOs sell performed work, GTM has to evolve with the pricing.

Per-seat pricing will not disappear. It is simple and familiar. But the more agent-native the company becomes, the more pricing should move toward outcomes, workflow capacity, or a hybrid of platform fee plus usage, output, or success-linked economics.

That makes the sale harder. The founder has to explain what work the agent performs, where the human sits, what happens when the system is wrong, how accountability works, and why the buyer should trust software with a job that used to belong to a person.

The best ANSO pitch is not, "we replace your team."

It is more precise: here is the work our system absorbs, here is where your humans stay in control, here is the economic math, and here is what improves as the model layer improves.

That is a very different pitch from "AI-powered software."

Why This Matters

AI-augmented SaaS can be excellent. But an ANSO is making a different bet.

It is betting that software can move from coordination to execution. It is betting that agents can produce the primary deliverable under the right trust architecture. It is betting that humans will govern more and produce less.

If that bet is right, the economics will look strange at first: lower gross margins in some places, higher revenue per employee in others, more outcome-based pricing, more workflow-specific moats, and smaller teams doing work that used to require entire departments.

That strangeness is the signal.

ANSOs matter because they are not a new label for AI startups.

They are a new economic shape for software companies.

-Larry J. Erwin

Disclosure: I work at OpenAI. Some companies mentioned in this essay use OpenAI's models. The views here are my own.

Sources