by Dr. Gleb Tsipursky
Vermont’s chief data and AI officer described a promising approach this month: state employees are learning to build small AI tools for tedious work, with trained people operating and checking them. This practical approach offers a useful alternative to buying a large system before anyone knows what it does well.
Still, residents may never see the mistakes an internal tool produces. A draft notice can leave out a requirement. A service summary can misstate a deadline. The employee checking the output needs explicit authority to halt the process and send a case back for correction.
The National Institute of Standards and Technology’s AI risk framework encourages organizations to measure and manage the risks of AI systems. Vermont agencies can put that into practice with a one-page workflow card: the task a tool performs, what it may get wrong, who checks it, and how staff stop or escalate a doubtful result.
If a resident challenges an AI-assisted response, the agency should offer an ordinary contact route and a meaningful correction process. Over time, count how often employees halt outputs and how quickly problems are resolved.
Frontline staff can become skilled builders without losing responsibility for the results. A visible human stop switch would help Vermonters trust the service they actually receive.
Gleb Tsipursky, PhD, a behavioral scientist, CEO of Disaster Avoidance Experts, and author of The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026) https://disasteravoidanceexperts.com/aibook