Local government AI training needs workflow proofing
Andriy Onufriyenko via Getty Images
COMMENTARY | Training should be treated as the beginning, rather than the finish line, and judge what works after the excitement of being in the classroom fades.
A recent two-day Applied AI in Local Government workshop at Pepperdine University ended with a useful premise: public-sector leaders need hands-on practice with real government scenarios. That is a better starting point than another abstract seminar about what artificial intelligence might eventually do.
But cities and counties should demand one more step. Every training program should end with evidence that a specific workflow improved safely.
The need is becoming urgent. California has made discounted AI tools, workforce training and implementation support available to state agencies, cities and counties. At the same time, local leaders are trying to modernize vital systems that cannot simply be switched off while staff experiment. Access is expanding faster than most agencies can redesign work.
That gap creates a familiar pattern. Employees attend a workshop, learn several prompts and return to offices where their managers have not identified approved tasks, confidential-data rules or review responsibilities. Some staff stop using the tool. Others use it quietly. A few produce impressive demonstrations that never become a reliable service.
Local governments can avoid that outcome by treating training as the beginning of a short workflow test rather than the finish line.
First, choose one recurring task with a visible public-service outcome. Good candidates include summarizing public comments, drafting routine resident communications, organizing inspection notes or preparing the first version of an internal briefing. The task should occur often enough to compare results, while remaining low-risk enough for supervised practice.
Second, define the human operating rules before anyone opens the tool. Staff need to know which system is approved, what information may never be entered, who checks the output and when the work must be escalated. “Use your judgment” provides too little guidance. A named reviewer and a short checklist create clearer accountability.
Third, compare the AI-supported process with the existing process. Measure time, error rates, rework, completeness and the experience of employees and residents. Tool logins and prompt counts show activity, not value. A pilot that produces more drafts but also creates more correction work has not improved government.
Fourth, hold a brief after-action review. Ask workers where the system helped, where it failed and what they were reluctant to report. This last question matters because employees may hide mistakes or unofficial use when leaders celebrate adoption too aggressively. Psychological safety turns frontline staff into an early-warning system.
The resulting scorecard should stay simple. It might show that permit-response drafts took meaningfully less time, but required the same legal review. It might reveal that translation assistance sped up routine notices while performing poorly with specialized terminology. It might show no measurable benefit at all. Each result helps leaders decide whether to expand, modify or stop the workflow.
This approach also protects smaller jurisdictions from vendor-driven adoption. A county does not need to buy a broad platform because a demonstration looked impressive. It can begin with a narrow problem, test the process and invest only after the evidence supports expansion.
Managers play the decisive role. They must make time for practice, protect employees who flag errors and resist treating AI use as a performance target. Workers should receive credit for improving a service, not merely for using a fashionable tool. A responsible pilot may conclude that human work remains faster or safer. That finding still provides value.
Public transparency should follow when an AI-supported workflow affects residents. Agencies should explain the purpose of the tool, the role of human review and the measures used to judge performance. They should avoid claiming success based on projected savings before real results exist.
Local government leaders face legitimate pressure to modernize. Staffing shortages, outdated systems and rising public expectations make experimentation necessary. Yet the safest path also offers the clearest evidence of value: train people on real work, build review into the process and measure outcomes that residents can recognize.
A workshop can create confidence. A verified workflow creates capacity. Cities and counties should judge AI training by what still works, safely and accountably, after the excitement of the classroom has faded.
Dr. Gleb Tsipursky is a behavioral scientist and author of The Psychology of AI Adoption at Work: From Resistance to Results.