The people writing the government’s AI strategies may be working down the corridor from colleagues who already use the technology without knowing whether they are allowed to do so. In a survey published by the Inter-American Development Bank in September, 85% of respondents reported using generative AI at work, while 82% lacked clarity about their organization’s position on it (IDB, 2026).
The research covered 3,193 public officials across 19 countries in Latin America and the Caribbean. Because participation was voluntary and recruitment drew on professional networks, the results may favor people already interested in digital tools. They cannot establish how common AI use is across the region, but they describe an institutional problem: employees have found something useful before their employers have worked out how to accommodate it. Most of that activity remains close to the individual’s desk. 79% reported using AI to draft documents, compared with 10.5% using it for public service activities. Quite understandable; an official can change how they prepare a briefing far more easily than they can change the approval process it enters.
The productivity gains depend on what happens after the draft is finished. Cutting a briefing from three hours to one frees up the author’s time, but does little for someone awaiting a decision if approval still takes a fortnight. Faster drafting could even add to the backlog by producing more work than reviewers can handle. The survey does not measure these effects, so its adoption figures tell us how widely AI is being used, without establishing whether it helps government reach decisions sooner.
Some of the most useful knowledge may already exist inside these organizations without travelling very far. While 43% of respondents said they used AI frequently, only 19% believed most colleagues actively used it. The figures do not establish secrecy, but they give managers reason to investigate how much experimentation happens without discussion. Someone who discovers that a model repeatedly misreads a particular document has learned something valuable. Unless that experience is shared, colleagues may encounter the same error independently, each paying the cost of discovering it. Useful methods can disappear just as quietly when an employee moves on.
Guidance has to make those experiences discussable. Staff should be able to describe an unsuccessful experiment without uncertainty about whether admitting it will itself cause trouble. They also need explicit limits on sensitive information, with examples close enough to their responsibilities to be useful. Among occasional users, 44% said they did not know how to use AI effectively. Training should therefore involve reviewing flawed work, including plausible answers whose mistakes require subject knowledge to detect.
A public body could learn more by following a few tasks through to completion than by celebrating another rise in adoption. Time saved during drafting should be considered alongside checking and corrections, then connected to what changes for the recipient. That would reveal where AI earns its place and where it merely produces more material for an overstretched institution to process. The distinction matters to anyone waiting for that institution to act.