State AI Deployment Without Training Is a Risk Model, Not a Strategy

Use Is Outrunning the Training Baseline

The 2026 NASCIO State CIO Survey, answered by 51 state and territory CIOs in summer 2026, shows how far adoption has moved. Ninety-six percent said employees in the CIO's organization use generative AI tools in their daily work, up from 54% two years earlier. Ninety-four percent reported pilot projects, 82% reported projects in production and 35% reported enterprise-scale projects across the executive branch.

Governance has largely kept pace on paper. In the same survey, 98% of states reported enterprise AI policies and procedures, 82% advisory committees or task forces, 79% an adopted governance framework and 77% an inventory of agency uses. The item that trails is the workforce: only 37% reported addressing the impact of generative AI on operations and workforce. One CIO's survey response captured the gap directly, noting that the state still lacked broad understanding and socialization of its AI policy, which made enforcement and governance difficult.

Code for America's 2026 Government AI Landscape Assessment, based on research completed in March 2026, reaches a similar conclusion from public records: most states are still early in their AI journey, the overall picture is governance-heavy and measurement-light, and training public employees in AI tools is identified as a major driver of adoption.

Spartan X's reading of those two sources together: states have written the rules faster than they have built the means to make employees able to apply them. The training gap is where that difference shows up.

What New York's Pilot Measured

New York is useful because it measured the baseline before scaling. According to the April 6, 2026 announcement from the Governor's office, republished by the Office of Information Technology Services (ITS), ITS ran a training pilot beginning in fall 2025 that paired AI training with AI Pro, a generative AI assistant developed by ITS and powered by Google Gemini. More than 1,200 participants took part across eight agencies: the Department of Labor (DOL), the Division of Homeland Security and Emergency Services (DHSES), the Department of Health (DOH), Homes and Community Renewal (HCR), the Office for People With Developmental Disabilities (OPWDD), the Office of General Services (OGS), the Division of Human Rights (DHR) and the Division of the Budget (DOB).

The starting point:

87% of participants had received no prior training in generative AI.
41% had no prior experience using generative AI tools.
Participants came from administrative, communications, policy, legal, service delivery, operations and technical roles.

The state-reported results after the pilot:

75% of participants saved time using AI Pro.
90% said their understanding of AI improved.
86% wanted to continue using AI Pro.
Overall confidence using generative AI rose 36%. The release does not say whether that is a percentage-point or relative change.
Participants generated more than 170,000 prompts.

Reported uses included executive summaries and policy briefs, translating complex topics into plain language, summarizing large reports and consolidating duplicative documents, and building data dashboards. The release says careful human oversight was emphasized throughout.

Two limits matter when using these figures. They are self-reported outcomes from a state press release, not an independent evaluation, and the pilot delivered training and the tool together. It shows that a population starting from almost no training can reach productive use when training is built into the rollout. It does not isolate how much of the result came from training rather than from tool access.

Three Enforcement Models States Are Using

States that have moved from policy documents to training programs have taken at least three distinct approaches. The useful way to compare them is by enforcement point: what, if anything, connects completed training to the person actually using AI.

Tool-gated training

New York's April 2026 expansion offers AI training to all state workers and provides AI Pro to the 50-plus agencies and entities ITS supports. The release states that responsible AI training will be required for agencies that elect to use AI Pro. The training is delivered with InnovateUS, an online public-sector learning platform, as a two-part course designed for New York State employees. The announcement does not say how completion is verified for individual users. The same release describes New York as the largest state in the nation to provide this technology to its entire state workforce, a claim made by the Governor's office in its own voice.

Pennsylvania uses a stricter per-user version. According to a November 10, 2025 Office of Administration release, more than 1,500 Commonwealth employees had completed InnovateUS responsible AI training as a prerequisite for access to Microsoft Copilot Chat and ChatGPT Enterprise, with another 4,100 enrolled. That followed a year-long ChatGPT Enterprise pilot with OpenAI in which 175 participating employees saved an average of 95 minutes a day, according to the same release. Carnegie Mellon University, which has supported the Commonwealth's AI governance work, now advises it under a cooperative agreement for AI advising services. New Jersey's state AI training page likewise directs employees to begin the AI training course before using the NJ AI Assistant, with the course itself providing the route to the tool.

The strength of tool-gating is precision. Training reaches the people using the sanctioned tool, and access gives the state an enforcement point it already controls. The weakness is coverage. Employees using AI outside the sanctioned platform, whether through consumer services, AI features embedded in existing productivity software or agency-level deployments, are not reached by this model. As more AI features arrive inside software states already license, a tool gate will cover a shrinking share of actual AI use unless the gate moves with them.

Statutory annual requirement

Texas took a different route. The Department of Information Resources explains that Government Code Section 2054.5191 requires state and local government employees and officials to complete a certified AI awareness training program, and that Section 2054.5193 directs DIR, in consultation with the Public Sector AI Systems Advisory Board, to certify at least five such programs. For state agencies, the requirement covers employees who use a computer for at least 25% of their duties, plus elected or appointed officers. Entities certify compliance annually by August 31 on the same certification form used for the state's cybersecurity training requirement.

This model solves the coverage problem by attaching training to employment, not to a tool. Its tradeoff is depth. An annual awareness course, certified to be equal in length to the cybersecurity training programs, establishes a floor of literacy. It does not by itself prepare a caseworker for the specific failure modes of the tool in front of them, and entities track completion by their own methods without submitting records to DIR. Texas also runs a voluntary AI User Group for public-sector staff alongside the mandate.

Community of practice

Maryland's approach emphasizes participation and peer learning. Its AI Community of Practice is open to anyone at any level of Maryland government, including state, local, tribal and territorial entities, and meets on the first Friday morning of each month. The state publishes session decks on its GitHub profile, and sessions range from no-code chatbot demos to case studies on scaling AI pilots.

Maryland's Department of Information Technology began offering free asynchronous AI training to state employees through InnovateUS in September 2024. Its 2025 AI Enablement Strategy and AI Study Roadmap commits to strengthening and expanding the Community of Practice as a collaboration and upskilling mechanism across state, county and local government. On May 28, 2026, Governor Wes Moore announced an AI Innovation Lab within DoIT that gives agencies a controlled sandbox, hands-on training, pre-deployment vulnerability testing and a path to turn pilots into ongoing products.

This is the broadest participation model and a low-cost one to run. It also creates horizontal learning that a compliance course does not. Its structural weakness is that attendance is voluntary and depth varies. A developer presenting a no-code chatbot and a program director deciding whether to approve AI in a resident-facing workflow attend the same session for very different reasons, and nothing in the model sets a different requirement for the second person.

Where the State CIO Community Has Landed

NASCIO's 2026 State CIO Top Ten Policy and Technology Priorities, drawn from 51 state and territory CIOs, placed artificial intelligence first for the first time, overtaking cybersecurity after 12 straight years at the top. Government Technology's coverage of the 2026 survey release framed the results as a path to AI maturity, built on enterprise policies, responsible-use rules, advisory bodies, governance frameworks, use-case inventories and procurement terms. The survey adds one more marker of that maturity: 49% of states now require disclosure by software providers.

At the NASCIO Annual Conference in San Diego in September 2026, Florida CIO Warren Sponholtz said that activities educating government employees on AI are critical, according to the same Government Technology report. Spartan X's analysis is that this placement matters. When a state CIO describes AI education as critical, it becomes part of the operating model rather than an optional human-resources effort. The survey data suggests most states have not yet given it that status. Every governance item above is more widely in place than workforce-impact work.

What Separates Training That Changes Behavior

Three factors separate training programs that change how people work from those that produce completion records. These are Spartan X recommendations, not findings from the sources above.

Coverage of the people actually doing the work. A program designed for IT professionals will not reach how a benefits caseworker, a program analyst or an investigator uses AI features in daily workflows. New York's pilot data is instructive because most participants started with no generative AI background. They needed foundational orientation, not technical depth. New York's statewide Acceptable Use of Artificial Intelligence Technologies policy (NYS-P24-001, issued January 2024 and updated June 24, 2026) applies to state agencies, their employees and third parties using covered systems. A policy with that reach needs training in a form that reaches the same population.

Connection to the specific tools and failure modes. Generic AI literacy does not prepare an employee for the tool-specific way an output goes wrong. Hypothetical example: a caseworker asks an AI assistant to summarize a long eligibility file, and the summary omits a household change reported midway through the record. A literacy course teaches that AI can omit information. Tool-specific training teaches this caseworker to check the summary against the change-of-circumstance fields in their own system before acting. The second is the training that changes the risk.

An enforcement point. Voluntary training produces self-selected coverage. Tool-gating, statutory mandates, role-based requirements and requirements written into agency AI intake each create an enforcement point. Without one, training remains a resource employees may use, not a capability the organization can rely on.

A Decision Framework for State Executives

State CIOs and agency executives face a discrete choice about sequencing:

Training before deployment. Hold tool rollout until training is designed and resourced. This minimizes early risk but delays benefits, and it is increasingly hard to sustain when AI features arrive inside software the state already uses. It makes sense for high-consequence uses such as eligibility determinations or investigations, where an error affects a resident directly.
Training concurrent with deployment. Gate access to the sanctioned tool on completing required training, as Pennsylvania does per user and New York does at the agency level. The cost is administrative effort to verify completion and some delay for employees who have not yet completed the course. It makes sense when the state controls a central enterprise AI platform.
Training as remediation. Deploy broadly, observe use, and address problems as they appear. This approach has a real logic, because observed use reveals training needs more precisely than assumptions do. It also accepts that consequential AI-assisted work will be done by untrained employees during the discovery period. It makes sense only for low-consequence internal uses with clear human review.

Spartan X's recommendation is a hybrid that combines a broad floor with a targeted gate: a required baseline for everyone who uses a computer for substantial work, on the Texas model; tool-specific training as a condition of access to enterprise AI platforms, on the Pennsylvania model; and a community of practice for the informal use that neither can reach, on the Maryland model.

Who owns it and what would change the decision

Ownership is usually split, and that is where programs fail. The CIO's office owns the platform and the access gate. Human resources or the training office owns the learning system and completion records. Agency program leaders own the decision about which workflows may use AI. If no one owns the connection between those three, completion records will not match the people actually using AI.

The decision should change with evidence. If incident reviews show AI-related errors concentrated in a particular workflow, that workflow moves to the training-before-deployment track. If usage data shows most AI activity occurring in embedded features rather than the sanctioned platform, a tool gate alone no longer covers the risk, and the baseline requirement has to carry more weight.

Metrics Worth Tracking

The 2026 NASCIO survey describes states measuring AI success first through participation, including tool use, training and adoption, before moving toward project results and outcomes. A training program that cannot measure its own reach cannot show that it is managing the risk it was built for. Three measures matter most:

Coverage denominator. What share of employees who actively use AI tools have completed baseline training? Answering this requires knowing who is using AI. Where usage of embedded AI features is not logged, the state cannot calculate this number and should treat that as a gap to close.
Behavior, not just completion. Completion counts are the floor. Useful measures include supervisor review of AI-assisted work, incidents traced to AI misuse, and employees' ability to identify AI failure modes in their own tools.
Population not yet covered. Hypothetical example: a community of practice with 700 regular participants in a state workforce of 80,000 reaches less than 1% of employees. That is a useful start and a weak measure of organizational readiness.

Sources and further reading

Spartan X has served as an AI advisor to Colorado state leadership, including CDLE, OIT and the Governor's office, and extended that state advisory model to Florida in 2025. That work sits where this article's decision sits: between a state's written AI policy and the employees expected to apply it. Spartan X helps state teams decide where the training gate belongs, who owns it, and how to measure whether training reaches the people actually using AI.