Don't Build the Tool. Deliver the Outcome.

Sequoia's Services as the New Software makes an argument I have been circling for months without saying out loud. For every dollar spent on software, several more are spent on the services around it, the people who actually do the work. AI is now good enough to attack that labor budget directly, not just sell another tool to the people spending it. The products that win move from tool, to copilot, to autopilot, and as they do the money quietly shifts out of the software line item and into the services line item, which is the much larger one.

I agree with the thesis. But I want to push on the practical conclusion, because I think most builders are reading it backwards.

The trap of building another piece of software

The instinct, when a new capability arrives, is to wrap it in a product. See what the model can do, put a UI on it, charge a subscription, defend the moat. That instinct is exactly the trap. If your entire value is a thin layer over what a model does today, you have signed up for open competition with whatever model ships next quarter. You are not building a business; you are renting a head start. The next release closes the gap, and your "moat" turns out to have been the six-month lag in someone else's roadmap.

Software, in the old sense, was valuable because capability was scarce and expensive to reproduce. That scarcity is collapsing. When the underlying reasoning gets cheaper and better on a schedule you do not control, "we built a tool that does X" is a fragile place to stand. You are competing on the one axis guaranteed to improve without you.

What agents are actually good at

Here is what changes the math. Agents are genuinely, reliably good at a specific class of work: applying complex rules and connecting things that humans find tedious to hold in their heads at once.

Hand an agent a specification and it can turn it into code that actually runs, not a sketch but working software that respects the constraints you gave it. Point it at a tangle of data and it will connect several dots and reason across them: this invoice contradicts that contract, this booking pattern implies that demand shift, this clause interacts with that regulation. The value is not the generation. The value is the judgment applied at a scale and consistency no team could sustain manually.

That capability is not a product. It is labor. And labor is the thing people have always been willing to pay real money for, because they were paying for the result, never for the tool the worker happened to use.

Deliver the service, not the software

So the move is not to package the capability and sell access to it. The move is to do the work and deliver the outcome. Take responsibility for the result the service was always supposed to produce, and let the agent be the means rather than the product.

Concretely, that looks like:

• A revenue function that doesn't sell a dashboard but actually moves the number (catching the underpriced segment, the leaking channel, the demand signal nobody acted on) and is measured by the revenue it recovers, not the seats it fills.
• A tax or audit service that watches a continuous stream of transactions and surfaces findings quickly, the day they appear, instead of a quarterly review that arrives too late to matter.
• A savings engine that reads through contracts, spend, and obligations to find the loopholes, the duplicate commitments, the unclaimed entitlements, and hands back money found, priced against what it found.

In every case the customer is not buying software. They are buying an improved P&L, a clean audit, a fee recovered. The agent is essential and invisible. If a better model ships next quarter, that is good news (it makes the service cheaper to deliver), not an existential threat to a product that no longer has a reason to exist.

Why this is the more durable bet

This is the part of Sequoia's argument I want to underline. Selling outcomes is sturdier than selling tools precisely because it is harder. You have to own the messy reality of the domain: the exceptions, the regulations, the data that never matches the schema, the accountability when something is wrong. A model upgrade doesn't hand a competitor that accumulated context for free. The defensibility moves from the capability, which is commoditizing, to the delivery, which is not.

I have spent eleven years in revenue management watching the industry buy tools and still struggle to move the number, because the tool was never the hard part. The hard part was doing the work the tool implied. That gap, between what software promises and the outcome someone actually needs, is exactly where this next wave of businesses should live.

Don't build another tool that competes with next year's model. Deliver the service. Get paid for the result.

References

• Original post: Sequoia Capital, Services as the New Software
• Sequoia AI Ascent 2025 keynote, AI's Trillion-Dollar Opportunity