Why Budget Control Jumped Three Spots
The NASCIO 2026 State CIO Top Ten, based on responses from 51 state and territory CIOs, draws one headline—AI at number one for the first time, displacing cybersecurity after 12 straight years at the top—but a quieter shift underneath it matters as much for day-to-day operations. Budget, cost control, and fiscal management moved from sixth to third. NASCIO's own description of that priority is broad—managing budget reductions and federal funding, finding savings, and dealing with inadequate funding—so the ranking does not by itself show that AI caused the jump. Our interpretation is that AI cost growth is one plausible contributor worth planning for, because it is the budget pressure most directly tied to the priority that took first place.
The mechanism, as we read it, is straightforward. AI pilots typically run with limited user populations on exploratory budgets. When pilots produce results, agencies ask to expand. The expansion request lands in a budget process designed for fixed-cost software, and the math gets harder. NASCIO's 2026 State CIO Survey, which also drew on 51 state and territory CIOs, asked whether CIOs anticipate increased turbulence and disruption in the state CIO role and organization over the next five years: 68 percent said yes and 22 percent said maybe. The CIOs who elaborated pointed to higher public expectations, artificial intelligence, increasingly sophisticated cyber threats, and aging legacy systems. Separately, among CIOs who said funding constrained digital services, long-term state budget uncertainty was one of the factors listed.
How State Budgets Were Built for Fixed Software
Most state budget processes treat IT spending in two categories: capital expenditure (building or buying a system) or annual operating expenditure (software licenses, maintenance contracts, hosted services billed at a fixed annual rate). Both categories share a structural assumption—the cost can be specified in advance with reasonable accuracy, appropriated by the legislature, and allotted by the budget office across the fiscal year. As one consultant told GovTech, capital expenditures are more common in state and local government, while cloud and AI tools generally align more closely with operating expenditures.
Traditional SaaS licensing fits this structure cleanly. A state pays a fixed amount per seat per year, multiplied by the projected user count. The unit of cost is predictable.
AI inference does not work this way. A short prompt consumes relatively few tokens; a long document review, a regulatory analysis, or a multi-step workflow can consume far more for the same user in the same session. California's Department of Technology makes the same point in its Poppy rate documentation: generative AI workloads "vary significantly across users and departments depending on mission requirements, document-processing volume, and analytic activity levels." Agentic processes that chain model calls are the steepest case—GovTech reports that token use rises with agentic tools, especially without constraints, because they can run continuously. None of this is visible from a license count. The variable is not how many people have access; it is how they use the service and for what.
The next budget cycle does not wait for consumption data. In South Carolina, for example, agency budget plans for FY 2027-28 were due September 25, 2026, with agency budget meetings running through October 30. A CIO writing an AI line item without production consumption data is projecting from a fixed-cost model applied to a usage-based service. The likely result is a budget that is wrong in one direction or the other—too low and requiring a supplemental, or too high and inviting cuts.
What States Are Finding
California built a centralized abstraction to address this problem directly. The California Department of Technology's Poppy platform—a state-built generative AI assistant that moved from a limited pilot to statewide availability on July 1, 2026—offers departments three options under rates effective the same day: a free one-month trial (up to 50 users per department, for departments not already subscribed), a pay-per-user model at $8 per user per month, and seven committed consumption tiers ranging from $7,500 to $50,000 per department per month. Under both priced models, usage above the subscribed capacity is billed on a pass-through basis at actual vendor cost plus a 2.79 percent administrative fee.
CDT's FY2026-27 rate memo explains that the tiers are fixed dollar levels rather than user-count tiers "to better reflect the underlying consumption-based nature of generative AI services," letting departments pool usage across staff while keeping monthly billing predictable. The memo is also direct about the underlying uncertainty: "during the initial budget year (FY 2026-27), cost recovery may not fully reflect steady-state operations while usage patterns and customer demand continue to develop." CDT projects FY2026-27 revenue of $1,855,180 against expenses of $2,998,391, a $1,143,211 shortfall, and expects adoption growth to bring full cost recovery in FY2027-28, when it projects a $186,702 surplus. CDT says it intentionally structured the service to encourage early participation and is launching it with existing staff, partly by redirecting effort from under-recovering services. The tier structure gives departments a predictable base charge, but it does not eliminate variability: CDT carries the first-year cost-recovery gap, and departments that exceed their tier still pay pass-through overage.
Indiana's Office of the Secretary of State encountered the problem at renewal. Robert Fulk, CIO for the office, told GovTech that a vendor initially sold his office an AI solution with unlimited fixed annual pricing, and at renewal he learned the pricing was switching to tokens. "I've heard horror stories," Fulk said. "I can't have our costs increase times two because then all of a sudden, we're not hitting our budget." Fulk said past usage data let him look back at consumption to better understand what he might pay after the shift—and asked how organizations without that history would manage it. GovTech described his experience as representative of a broader shift in the AI economy.
Alaska has not funded an enterprise rollout. A $732,700 general-fund request for Microsoft 365 Copilot licenses for state employees, made in the FY2026 governor's budget, was not brought forward again in FY2027 "due to fiscal constraints, though broader expansion remains a future consideration," according to the Department of Administration's January 2026 budget overview. The same document says the limited rollout has resulted in uneven access across departments, with most employees lacking AI tools, and that departments continue to self-fund licenses as needed, leading to inconsistent adoption.
Georgia is building usage visibility into its AI program governance. According to GovTech, the state partnered with Darwin AI, which helps officials gain improved visibility into usage, said Nikhil Deshpande, chief digital and artificial intelligence officer for the Georgia Technology Authority. Each agency will have AI champions overseeing use at the agency level, while GTA monitors overall use, and the state is using role-based personas to understand how different positions may use AI to inform future budgeting. John Matelski, CIO of the National Association of Counties, summarized the vendor dynamic: "The vendor—or vendors—that can come forward and solve the predictability aspect of things is going to win trust and is probably going to be the vendor of choice."
South Carolina addresses the problem at the budget-submission stage rather than through platform architecture. The state's FY 2026-27 and FY 2027-28 agency budget guidelines direct that funding requests related to artificial intelligence should be coordinated with the Department of Administration's AI Center of Excellence, pursuant to a budget proviso, and designated "AI Related" in the state's budget system. Central coordination and tagging give the state a mechanism to see AI spending requests together and apply consistent cost assumptions across agencies—rather than discovering disparate AI budget lines when the consolidated request lands at the budget office.
Three Structural Approaches
Centralized platform with consumption tiers
A state IT agency aggregates LLM vendor contracts, manages variable billing, and offers agencies fixed monthly commitment tiers priced to recover costs at the platform level. Agencies get predictable base line items; the central IT organization takes on most of the forecasting burden.
This works when the platform has enough participating agencies to average out uneven usage patterns, and when the central agency can carry first-year shortfalls while consumption baselines develop. California's Poppy rate structure and its projected FY2026-27 shortfall illustrate both the advantage—departments can budget a predictable base—and the residual risk, which does not disappear but is split between the platform operator and departments that exceed their tier.
Per-agency consumption tracking before commitment
Track per-agency AI consumption during pilots and early production before committing to volume contracts or enterprise licenses. Build agency-level accountability for usage visibility—not just aggregate reporting at the platform level—so the CIO's office has actionable data by workload type when budget submission time arrives.
This works when agencies are in early production and haven't yet locked into volume contracts, and when there's organizational will to maintain agency AI champions who understand both the use cases and the cost trajectory. Georgia's model of agency champions plus central monitoring is one version of it.
Workload-tiered model selection
Not all AI tasks require frontier-model inference. Classifying tasks by complexity and routing them to appropriately priced models can reduce average per-query cost. The key budget implication: agentic workflows—those that chain multiple model calls within a single task—can consume substantially more than single-query assistant interactions. Mixing them in the same budget line produces projections that can be wrong in both directions.
Build cost profiles by workload type: single-turn assistant tasks (lower cost per query, high volume), research and analysis tasks (higher cost per query, moderate volume), and agentic process tasks (highest cost per query, currently lower volume but growing). Budget each workload category separately with different cost-per-query assumptions. When agentic workloads scale, the budget impact becomes visible before it appears as a billing event.
Five Questions Before Submitting an AI Budget Line
A budget office will need answers to these before an AI appropriation can be justified:
- What was actual AI consumption in the current and prior fiscal year, by quarter and by use case? If this data does not exist, the budget estimate has no empirical foundation.
- Which AI use cases transition from pilot to production in the budget year, and at what user scale? Each category carries a different cost profile.
- Is the request for a fixed commitment tier, per-user pricing, or consumption pass-through? The fiscal risk profile differs in ways the budget office needs to understand.
- What is the overage mechanism? If consumption exceeds the appropriation, does the agency submit a supplemental request, reprogram from another line, or restrict service access?
- Does the budget office have a process for mid-year AI cost adjustments? Traditional budget amendment processes assume cost changes are infrequent and large; AI consumption can drift month to month.
A budget request that cannot answer these questions is built on a fixed-license mental model applied to a usage-based service. The budget office will either see through it or approve it and discover the gap mid-year.
What Changes the Decision
If a state is in its first year of significant AI production deployment, our recommendation before the current submission window closes is to establish automated consumption reporting by agency and workload type, and to build an explicit contingency above projected consumption, sized from the month-to-month variance observed so far rather than from a fixed rule of thumb. Production data from a full fiscal quarter is far more defensible than pilot projections extrapolated to production scale.
If the state has 12 or more months of production consumption at close to the projected user scale, use that data directly, project by workload category, and document the methodology explicitly so the budget office can explain the estimate through the appropriations process.
The alternative—relying on vendors to keep offering unlimited fixed pricing—carries renewal risk. GovTech treats the Indiana experience as representative of a broader market shift, and Fulk expects an industry-wide "reset and shift" that will require careful review of contract renewals, including for tools with AI embedded in them. A state that built its AI budget model on a fixed-price contract should verify now whether that pricing applies to the next renewal period, rather than treating the renewal as routine.
A budget that a CFO can audit and a budget office can track is also a budget that survives a change in administration. States facing a gubernatorial transition in early 2027 are better positioned when AI program fiscal accountability is documented, not assumed.
Sources and further reading
- NASCIO 2026 State CIO Top Ten Policy and Technology Priorities
- NASCIO 2026 State CIO Survey: The State CIO in Motion
- California Department of Technology, FY2026-27 Rate Change Detail for Poppy
- California Department of Technology: Poppy Now Available Statewide (July 1, 2026)
- Alaska Department of Administration, FY2027 Governor Budget Overview (January 29, 2026)
- South Carolina Executive Budget Office, FY 2027-28 Guidelines for Agency Budget Plans
- GovTech: What Is Tokenomics? Government Budgets Grapple With the AI Era
- GovTech: AI Overtakes Cybersecurity in State CIO Priorities for 2026
Spartan X's state AI advisory work sits at the program design layer where these cost decisions are made: building AI roadmaps around budget realities, defining use-case scope in terms a budget office can track, designing consumption visibility into the architecture from the first deployment, and evaluating model options so workloads run on appropriately priced models before renewal terms arrive.
