Government workers are not the AI holdouts they are often made out to be. In a large training program for state and local government staff, half of the people who signed up had never tried generative AI, and among those who were not using it at work, many could not say whether they were allowed to.
The doubt is about the job, not about change
Among the people not using AI, 28% did not think it applied to their work, 23% did not know how to use it, and 19% did not know whether it was permitted. Managers who were not using it sounded much the same: 24% did not know how, and 22% did not know if it was allowed. A worker asks a manager, the manager is unsure too, and the safe answer for both is to wait.
We covered a version of this in the private sector, where most employees do not know why their company uses AI. Buying the tool and publishing a policy answer the company's question. The worker's question is smaller: may I use it for this letter, this file, this phone call?
For example, a manager at a county unemployment office could write down three things a caseworker may do: turn a denial letter into plain language, summarize a long case file before a call, and draft a reply to an applicant. The same note would name what never goes into the tool, such as a Social Security number. One page like that answers "does it apply to me", "how" and "am I allowed" in a single read.
The newest workers gain the most
The strongest reason to aim at beginners comes from the customer-support case. The assistant watched each chat and suggested what to say next, so a new agent got, in the moment, the kind of answer that experienced colleagues had learned over months. Experienced agents already had those answers, which is why their gains were small while novices and low-skilled workers resolved 34% more issues per hour.
The government training data points the same way, with a caution about what it measures: people who started with no AI knowledge showed the largest gains in what they said they knew, and that is reported learning, not better work. Still, the two results agree, and they suggest that an office with a high turnover of new staff has more to gain than one made of veterans. If new hires in a benefits call center reach veteran speed about four months sooner, as the support agents did, that is four months of shorter waits for the families on the line.
Confidence decides how far use spreads
After training, people who felt highly confident planned to use AI for an average of 3.8 tasks, against 1.5 for people with low confidence. Those numbers go together, and they do not prove that confidence causes wider use. But it fits what a manager sees: a person who has watched AI work on one of their own tasks starts to see the next one. Confidence is not accuracy, though, so the same training has to teach people to check an answer and to know when not to use AI at all.
The last step belongs to management. Faster paperwork is not a shorter queue until someone decides where the saved hour goes. A clerk who gets an hour back can answer more calls, or the office can simply do the same work with the same waits. The offices that win will be the ones that ask workers what they are unsure about, answer it for each job, and then point the time at the wait that residents feel. The report behind these numbers comes from volunteers and does not yet show whether waits fall, and that is the number worth watching.