Exporting custom software
in the AI era is like
exporting cattle on the hoof

In 2025 Uruguay exported approximately 390,000 tons of beef for USD 2.68 billion — a historic record that accounted for nearly 20% of the country's total goods exports (source).

But those numbers weren't built by selling live cows. They were built by exporting premium cuts, with value added at every link of the chain.

Exporting cattle on the hoof is the simplest possible model. The animal leaves the country, someone else processes it, someone else adds traceability, someone else does the cuts, someone else packages it, someone else places it on a supermarket shelf in Tokyo or Houston. All that value — and all that margin — stays with others. It leaves the country behind.

Uruguay's beef value chain does exactly the opposite. It doesn't just process: it certifies, classifies, traces, refrigerates, debones, vacuum-packs and places the right cut in the right market at the right price. Individual animal traceability, natural pastures, controlled health protocols, an uninterrupted cold chain from packing plant to destination — all of that transforms a commodity into a premium product with its own identity.

Now apply that logic to software. Exporting traditional custom development — teams that only receive tickets, write code and deliver features — is, in essence, exporting cattle on the hoof.

The client abroad receives the raw input and adds the value that's missing — both before and after the code: functional analysis, product vision, architecture decisions, testing, commercial strategy, go-to-market techniques, project management, business integration. All of that is done by others. We don't capture that value — and therefore, neither the margin.

For years, this model worked very well in Uruguay. There are dozens of companies that sell software development hours, and there was a real differentiator: regional developers had experience and so they asked better questions, knew how to say NO when it had to be said, operated in a friendly time zone (nearshore), and rates were competitive.

AI is changing all of that. A developer with the right tools — Cursor, Claude Code, Copilot — today produces code that previously required years of experience. The volume of code generated per person has multiplied several times over. And as a result, the cost per line has collapsed. The gap (or at least a large part of it) between a Montevideo developer and a Bangalore developer is narrowing faster than we like to admit.

Just as Uruguay didn't stay in the business of exporting live cows, there's a software value chain that adds layers of processing, quality and identity. That's what we have to export.

If we could also add intellectual property and export products rather than custom development — as Amílcar Perea has been advocating for years — even better. But building product is hard, expensive and high-risk. Meanwhile, let's at least export these layers alongside the development:

Sit virtually with the end user, talk, and understand how they actually work — not how they think they work. Map flows, identify frictions the client doesn't even know they have, detect automation opportunities that were previously unthinkable. It's the equivalent of knowing which cut goes to which market: not every market wants the same thing, and understanding that delivers a lot of value to the client — part of which we can capture.

Does this feature still make sense in six months? Does it move us toward our strategic north or away from it? Is the roadmap responding to real market needs or to assumptions? Are they building on an architecture that's going to slow us down later? Those conversations are the design of the slaughterhouse process: they define what comes out, how, and for whom.

Designing systems that scale and stay maintainable isn't writing code: it's the engineering of the plant. AI can generate code; deciding the right architecture for a specific context is a different category of work where many more variables come into play — performance, operating cost, business evolution, regulatory constraints, vendor lock-in.

Integrating AI into products so they actually work — with well-designed prompts, validation flows, hallucination handling, more determinism and less probability — is a new discipline few have mastered. It's the equivalent of refrigerating and packaging correctly: without it, the product arrives in bad shape.

AI-generated applications fail in different and unpredictable ways. Testing well — with real coverage, well-thought-out edge cases and human validation at critical points — is the software's traceability: the guarantee that what ships is what was promised.

Cyberattacks in Uruguay have been growing without pause since 2020 (source). Any new development must not only follow good practices: it must be designed to be cyber-secure, with code reviewed to that end. It's a huge differentiator for those who can guarantee a secure development process. And it's the sanitary control: without it, the destination market closes the door.