People who sell tools race against the model. People who sell the work race with it.
Everyone building something with AI today does the same math in their head: what happens to my product when the next version of the model can do this on its own? The question is honest and the fear is fair. If what you sell is the tool, every new model shortens the distance between you and anyone else. But if what you sell is the finished work, the same update that would threaten the product arrives as a margin increase: faster, cheaper, harder to catch.
Intelligence and judgment
It's worth separating two things that usually come in the same package.
Writing code is, almost entirely, intelligence. The rules are complicated, but they are rules: translate a spec, test, find where it broke. Deciding what to build next is judgment. It depends on experience, on taste, and on having been wrong before. Which technical debt to accept, when to ship before it's ready, what to cut when the deadline shrinks.
AI crossed the intelligence threshold in software engineering first, for a simple reason: it's the profession where the machine can check on its own whether it got it right. It compiles or it doesn't, the test passes or it doesn't.
Where agents are being used (% of tool calls)
Software engineering |================================== 49.7%
Back-office automation |====== 9.1%
Other |===== 7.1%
Marketing and copywriting |=== 4.4%
Sales and CRM |=== 4.3%
Finance and accounting |=== 4.0%
Data analysis and BI |== 3.5%
Academic research |== 2.8%
Security |= 2.4%
Customer support |= 2.2%
Source: Sequoia, "Services: The New Software" (2026)
Half, against single digits for everything else. It's not that the other professions are harder. It's that nobody has yet built the verification loop that engineering got for free.
The client never wanted the tool
Here is the part the Brazilian market has always known and the product market keeps forgetting.
No client wakes up wanting to buy software. They want the problem solved. A company that pays for an accounting system and then pays the accountant to close the month isn't buying two things: it's buying one, and paying the license as a toll. What it wanted was the month closed.
Selling a tool forces the client to become an operator. They buy the license, learn the interface, find out a piece is missing, and hire someone to connect the ends. Selling the work skips that whole line.
Where that leaves Nimbuu
The service is the door and the software is the house. A website, a system or an app leaves Nimbuu already installed inside the ecosystem: it charges through Pay, gets edited in Build, and keeps running without depending on a call back here.
None of these products started as a startup idea. Each one started as a problem that came up in the middle of a delivery:
- There was finished work to get paid for, and the checkout that existed was painful to integrate. It became Pay.
- The client needed to change a line of text without filing a ticket. It became Build.
- Several projects at once needed coding agents without stepping on each other. It became Stratuu.
It's the reverse of the order people teach. First the work, which is what someone pays for today. Then the tool, which is what's left of work done well and repeated enough times to be worth automating.
The test
If the next version of the model makes your product obsolete, you were selling the tool. If it makes your delivery faster, you were selling the work.
Worth asking the question before the market asks it for you.