Services are eating software

Software used to sell tools. AI lets companies sell the completed job.

That is the whole shift.

For most of the SaaS era, the bargain was simple: a vendor gave your employees better buttons, cleaner workflows, and a database that made management feel less blind. The customer still supplied the labor. Humans still reconciled the books, worked the ticket queue, reviewed the claim, chased the lead, fixed the spreadsheet, and handled the weird edge case.

AI makes that bargain feel unfinished. If a system can read the inbox, classify the request, search the CRM, update the record, draft the response, and escalate the exception, why buy another seat? Why not buy the outcome?

That is why Sequoia's "Services: The New Software" thesis feels important. The next category-defining AI company may look like an accounting firm, brokerage, claims processor, medical billing shop, recruiting agency, or back-office outsourcer.

From the outside: service business.

Under the hood: software economics trying to escape the SaaS box.

The quick skim

• Old SaaS captured the tool budget. AI-native services can go after the labor budget.
• The product changes from interface to outcome. Not "software for accountants" — "we close your books."
• The moat moves into workflow data. Every exception handled becomes training material.
• The hard part is trust. Customers do not just need automation. They need someone accountable when the workflow breaks.

Why the old SaaS deal was incomplete

Traditional software was a phenomenal business because it scaled. Build once, sell many times, charge per seat.

But in many categories, the software line item is smaller than the payroll line item around it. A company might spend a modest amount on bookkeeping tools and far more on the people who operate them. The tool is useful. The work is expensive.

AI attacks the expensive part.

That does not mean every service becomes a pure robot overnight. The interesting version is messier: agents do the repeatable work, humans handle the judgment calls, and the company sells a finished workflow with a service-level promise.

That is not classic SaaS. It is closer to a labor-replacement wedge with software gross margins hiding inside.

Where the moat may form

The first-order story is automation. The better story is compounding.

A normal services firm hires more people as demand grows. An AI-native services firm should, in theory, improve with every client, every exception, and every workflow variant. The work produces the data that improves the system that does the work.

That creates a different kind of flywheel:

Layer	What improves
Workflow data	More examples of real edge cases
Agent tools	Better actions, retrieval, and escalation paths
Human review	Clearer places where judgment is still needed
Customer trust	More willingness to hand over bigger workflows

The moat is not just the model. Everyone can call a model API. The moat is the operating loop around the model: how the company captures exceptions, converts them into process knowledge, and proves reliability to customers who cannot afford hallucinated work.

The uncomfortable part

A lot of AI startups will accidentally become low-quality agencies with a chatbot attached.

That is the trap.

If the company has to add humans linearly with revenue, it is still mostly a services business. If the AI layer gets brittle across customers, it becomes consulting. If the promised outcome is too vague, margins disappear into custom work.

The winners need a narrow enough workflow to standardize, a painful enough labor budget to matter, and enough repeated volume for the system to learn.

My mental model

The old software question was: "Can this workflow become a product?"

The AI-native question is sharper:

Can this service become a learning system?

If yes, the company may look boring from the outside and compound like software underneath. If no, it is probably just another agency wearing an AI hoodie.