Vermont AI Council seeks more resources, outlines state AI pilots and public‑facing tools

The Vermont AI Council and Agency Digital Services (ADS) used a May 6 briefing to present their second annual report, describe ongoing AI pilots across state government and ask the Legislature to consider additional support to expand council capacity and oversight.

The council recommended an intermediate step before statewide regulation: serving as a clearinghouse of best practices for businesses and state agencies while the state builds staffing and technical capacity. For the record, Mark Holmes, chief technology officer of Agency Digital Services and cochair of the Vermont AI Council, told the committee, "For the record, I'm Mark Holmes. I'm the chief technology officer of the agency digital services as well as the cochair for the Vermont AI Council." Holmes said the council was established under Act 132 in 2022 and has already produced a code of ethics and guidance for state uses of AI.

Why it matters: Council leaders said Vermont is ahead of many states in publishing ethics and guidance, but that meaningful expansion of regulatory authority would require more staff and resources. The council and ADS described active pilots intended to improve internal productivity, public service search and disaster response, and said they are taking a cautious approach to public-facing generative tools to avoid publishing incorrect information.

Most important takeaways

- Scope and request to Legislature: The council asked for legislative support to broaden its role and membership and to fund more staff. Holmes and others told the committee that the current ADS AI division is one position and that a phased approach would let Vermont act as an information clearinghouse for businesses and agencies while monitoring models of regulation adopted elsewhere.

- ADS pilots and tools: Desiree Shunseis, chief data and AI officer, described a set of in‑production and pilot tools. "We put together an inventory. I believe there are around 30, maybe 34 use cases on it right now," she said. ADS has an internally hosted large-language model called ChatVT for staff drafting and productivity; Shunseis said it had "around 300 active weekly users before our expansion on on Monday." ADS also described a public-facing search interface used during flood recovery that layered AI on an older contracts database to let municipalities more quickly find relevant emergency contracts.

- Public-facing chatbots and approach: ADS said it intends to roll out agency pilots of curated chatbots that match user questions to vetted question‑and‑answer pairs rather than letting a generative model synthesize answers from the web. ADS staff said a public chatbot rollout is awaiting the new AI director completing background checks and stressed that answers will be human‑reviewed before publishing.

- Training, governance and staffing: The council and ADS said they have held multiple trainings ("probably 8 or 9" sessions), are planning learning‑management system modules and expect to hire a new AI director. Norwich University offered a four‑hour "AI for policy makers" course to the committee. ADS described combining the data and AI functions into a single office to address recurring data quality hurdles that block AI work.

- Pilot examples and limits: ADS described a pilot using AI to pre‑redact officer narratives for research, reducing human review time from weeks to about one week (with additional manual checks). ADS also said it uses AI for code reviews and report summaries to improve quality rather than as a shortcut to avoid human oversight.

- Procurement and contract terms: Committee members and ADS staff discussed embedding AI expectations into procurements and vendor contracts (for example, prohibiting vendors from using state data to further train their models or share it with third parties) as AI becomes embedded in camera systems, office suites and other products.

- Risks highlighted: Council members flagged emerging risks including agentic AI, cyber risk (AI vs. AI in attacks and defenses) and deepfakes. The council said it is monitoring other states' regulatory experiments (California and Colorado were cited) and will wait to see how litigation and rule‑making proceed before proposing Colorado‑style consumer protection rules for Vermont.

Discussion versus direction versus decisions

- Discussion: Council members and committee representatives exchanged questions on training rollout, the difference between agentic AI and current tools, how AI hallucinations occur and how to avoid publishing incorrect answers to the public.

- Direction/assignments: ADS said it will expand its ChatVT pilot to more internal users, build curated question‑and‑answer sets for public chatbots in agency pilots (HR, marketing, human services) and proceed with hiring an AI director. Norwich University offered, and the committee accepted consideration of, a four‑hour AI training for legislators.

- Formal action: The briefing contained recommendations and requests to the Legislature but no formal legislative motions or committee votes were made during the meeting.

Ending

Council and ADS leaders asked the committee for guidance on whether Vermont should evolve toward a regulatory body or instead strengthen an education/clearinghouse role first; they recommended a phased approach that builds capacity, governance and public engagement before pursuing broader enforcement authority.