Many people are sharing the story of Reco that saved $500k a year using AI to rewrite a JavaScript JSONata implementation to Go. But most of those people are sharing the wrong lesson.

The headline makes it sound like magic. Take some AI, point it at some old code, get new code, save half a mil. But if you actually read what went down, the magic part wasn't the AI.

The Part Nobody Wants to Talk About

Reco didn't just throw their codebase at an LLM and hope for the best. They adopted the jsonata-js official test suite (1,778 tests), and ported it over to Go as a specification that would guide every step of the rewrite.

That's the boring part. That's also the part that makes it work.

→ The AI didn't decide what the Go code should do. The tests did.
→ The AI didn't validate correctness. The spec did.
→ The AI was the typist. The engineering team was the architect.

Remove those guardrails and you don't get a half-mil saving yarn. You get a half-mil clean-up horror story.

AI-as-Magic vs. AI-as-Tool

The online discourse split into two predictable camps. One side says this proves AI can replace developers. The other side says it proves nothing because AI just did grunt work.

Both miss the point. The real story is about what was already in place before AI entered the picture. A team that writes thorough tests and maintains a real spec can hand boring translation work to an LLM. A team without those things can't hand anything to an LLM safely.

AI is a force multiplier. But a multiplier on zero is still zero. 🤷

The Unsexy Investment That Pays Off

I think about this every time someone on my team pushes back on writing tests. "We're moving fast." "We'll add them later." "The code is simple enough."

Maybe. But you're also closing the door on every future shortcut. Good test coverage isn't just about catching bugs today. It's a machine-readable description of what your software is. That description becomes incredibly valuable the moment you want to migrate, rewrite, or — yes — hand work to an AI.

Reco's test suite wasn't written for AI. It was written because that's solid engineering. The AI benefit was a side effect of doing the fundamentals right.

What This Actually Means for You

If you're excited about AI-assisted rewrites, great. Start by asking yourself one question: could you hand your codebase to a new human developer with just your tests and docs, and would they know exactly what the code should do?

→ If yes, congrats — you're AI-ready.
→ If no, the LLM won't magically figure it out either.

AI technology does not eliminate the importance of engineering fundamentals. Instead, it highlights the benefits of such practices. Teams that put effort into so-called “boring” but necessary work – such as tests, specifications, and well-defined interfaces – were making a solid investment. Others found themselves producing fast, unprecedented amounts of garbage.

The Takeaway

Yes, the $500k quote is accurate. But the merit goes to the engineers who had already created a test suite that was valuable and reliable long before a single prompt was entered. The AI model merely represented the last step; the first 25 were solid engineering practices.

So, would an AI model be able to rewrite your current project using only the tests and the documentation you built along with it?