The next competitive edge in AI isn't the model. It's the context.

On September 9, I defended my doctoral thesis in Artificial Intelligence at UNIRIO. Its subject sits at the frontier of the AI companies are adopting right now: context engineering.

I'll set the personal news aside and go straight to what I learned, because I think it applies to any business going through digital transformation.

The AI companies buy today is good at predicting. Credit, pricing, résumé screening, performance reviews. Among the 500 largest American companies, 97% already use automated systems to filter candidates. The problem is that these machines get it right and don't explain. An automated decision without an explanation isn't a better decision. It's an unquestionable one. And what can't be questioned doesn't correct its errors: it accumulates them.

I built one of these machines: a model that estimates salaries from 66,000 real employment records in Rio's IT market. It reached 99% accuracy. That was too good to be true. It was the explanation, not the score, that exposed the flaw: one variable was leaking the answer.

An explanation isn't the model's decoration. It's the model's fraud detector.

Then came the second discovery. I opened the model with the most advanced explainability technique there is, and the result was technically perfect and useless to anyone deciding. It said a person's experience contributed "−0.230". Minus 0.230 of what? The explanation was written for the engineer auditing the model, not for the manager, the candidate or the HR team who would act on it. The whole field solved the problem of explaining the machine to its creators and declared victory.

This is where context engineering comes in. The frontier of AI isn't only in bigger models. It's in what you put into a language model's context window: the knowledge that makes the answer useful for a specific decision. In the thesis, that knowledge came from five professionals who set salaries in Brazil. I turned hours of conversation into 18 context variables that no database held.

The result is EnriCX, a layer that sits after the model and the statistical explanation, without changing either, and turns abstract numbers into priced decision alternatives: this experience is worth R$ 2,291 less a month, and here's why; moving to another sector is worth R$ 7,411 more, but depends on an opening; building seniority is worth R$ 5,306 more and depends only on you.

Three takeaways for business leaders

The model will become a commodity. The scarce asset is your experts' tacit knowledge, which nobody has written down yet. Context engineering is the method for turning it into a system.

Explanation is decision architecture. The quality of an explanation changes the quality of what people decide with it. Whoever designs the system is accountable for that.

Be wary of any AI that only shows its hits. My model explains about a third of salary variation in reais, and I publish that number. The rest is human. Whoever shows you the error treats you like an adult.

My thanks to Professor Ana Cristina Bicharra Garcia, a pioneer of AI in Brazil, who pointed the way when all I could see was the limitation.

Amounts are in Brazilian reais and are model references, not promises. Learn more about the research.