A global study released this week shows Brazil leading the world in production AI agent adoption — ahead of the United States and the global average. The same study shows Brazil also leading on a far less celebrated indicator: the share of companies forced to roll back an AI project already in production because of a governance failure.
What changed
According to "The AI Production Paradox," a study by Sinch with an independent research institute that surveyed 2,527 senior executives across ten countries, 76% of Brazilian companies already run AI agents in production — above the 62% global average and the 67% recorded in the United States. At the same time, 80% of Brazilian companies had to interrupt or reverse an AI implementation because of a governance problem, versus 74% globally. In 39% of those cases, the rollback happened after a data or personal-information breach. The survey also found that 66% of Brazilian companies plan to increase AI investment.
Why it matters
Being ahead on the speed of putting agents into production isn't the same as being ahead on the ability to sustain those agents safely. The numbers suggest the opposite of a balanced race: Brazil moves faster into production and, proportionally, also gets it wrong more often — or at least finds out later, after the data has already leaked, rather than before. That fits a pattern we've discussed here repeatedly: governance infrastructure (permissions, auditing, monitoring) tends to get deferred when the pressure is to show results fast.
The impact for Brazil
The study's most concrete detail — nearly 4 in 10 rollbacks happened after a personal-data breach — should weigh especially heavily for Brazilian companies as ANPD enforcement on AI tightens, a topic we covered here when reporting on the delay of Brazil's AI bill. The Marco Legal's vote may remain stuck, but the LGPD is already in force and already applies to any AI agent processing personal data today. Companies celebrating the 76% production-adoption figure without looking at the 80% governance-rollback figure are reading only half of their own result.
Entercast's take
This is exactly the pilot-to-scale pattern we've flagged here repeatedly: moving out of pilot fast is good, but only if governance structure keeps pace with it, instead of chasing it after the fact. Before celebrating a lead in adoption, it's worth asking the inverse question the Sinch study itself exposes: how many of the AI agents already in production at your company have auditing, access control, and an incident-response plan defined with the same priority they had at launch? If the answer is "not sure," the number that matters isn't the adoption rate — it's the rollback rate.