KF
Federal AI strategy is usually discussed as a technology question — which models, which use cases, which platforms. For the CIO who actually has to deliver it, the binding constraint is rarely the technology. It is the budget cycle. Federal funding moves on an appropriations rhythm measured in fiscal years, with money that comes in specific colors, expires on specific dates, and is planned eighteen to twenty-four months before it is spent.[1] AI moves on a different clock entirely — capabilities shift in months, and the system you scoped at budget formulation is not the system available at execution. The gap between these two clocks is where federal AI ambitions quietly die, and no amount of technical excellence closes it.
Two clocks running at different speeds
The appropriations clock is deliberate and slow by design. A CIO building next year's budget is making bets eighteen months out, defending them through formulation, and executing them in a fiscal year that ends on a hard date. That cadence served decades of IT investment, where the technology a CIO scoped at formulation was substantially the technology available at execution.
The AI clock does not cooperate. A capability scoped at budget formulation can be superseded before the money is even available to spend. The use case that justified the request may be obsolete; the approach may have been overtaken by something that did not exist when the budget was built. The CIO is executing a plan written against a version of the technology that no longer represents the state of the art — and the appropriations structure offers little room to adapt mid-cycle without re-justifying the money.
"A CIO scopes an AI capability eighteen months before the money arrives. By execution, the technology has moved and the plan is stale — but the appropriation is locked to the plan."
The color-of-money problem AI makes worse
Federal money comes in colors — different appropriations with different rules about what they can buy and how long they last. This has always shaped IT investment, but AI sits awkwardly across the categories in a way that compounds the timing problem.
- Is AI development or operations? The model and the application have the character of a capital investment; the continuous evaluation, model updates, and feedback loops have the character of ongoing operations. AI requires both, funded together, but the appropriation categories often want to treat them separately.
- The recurring cost is the real cost. Traditional IT front-loads cost into build and tapers into maintenance. AI inverts this: the standing cost of governing, evaluating, and improving the system is continuous and substantial, and it needs a funding line that recurs rather than a one-time build appropriation.
- Expiring money fights iterative work. Money that must be obligated by a fiscal-year deadline pushes toward big one-time buys, which is exactly the wrong shape for a capability that should be funded to evolve. The expiration clock rewards the spend pattern AI least wants.
Why so many pilots strand
This budget structure explains a pattern every federal observer has seen: the AI pilot that works, demonstrates value, earns praise — and then strands, never reaching production. The technology was not the problem. The pilot was funded as a discrete, time-boxed effort, and the money to operate and sustain it in production was a different appropriation that had to be planned, justified, and won in a later cycle. By the time that cycle comes, the pilot's momentum has faded, the champions have moved, and the production funding competes against fresh priorities. The pilot succeeds and dies in the gap between the money that built it and the money that would have run it.
The lesson is not that pilots are bad. It is that a pilot funded without a funded path to production is a demonstration of what the agency cannot afford to operate. The budget structure, not the technology, decides whether the pilot becomes a capability.
Funding for a capability that changes
CIOs who deliver AI despite the budget mismatch share a set of moves that align the funding to the technology rather than fighting it.
- Fund outcomes, not point solutions. Frame the budget request around a sustained mission outcome that justifies recurring investment, rather than a specific technical solution that will be stale by execution. Outcomes survive technology churn; solutions don't.
- Plan the production path before the pilot. Secure, or at least map, the operations funding before launching the pilot, so success has somewhere to go. A pilot with no funded production path is a budget request you haven't made yet.
- Build for substitution. Architect so the model and components can be swapped as the technology moves, so the eighteen-month-old plan can absorb a newer capability without re-justifying the whole investment. Flexibility in the architecture buys flexibility the appropriation doesn't provide.
- Right-size the recurring line. Budget the continuous cost — evaluation, governance, model updates — as an explicit recurring line from the start, not as an afterthought discovered after the build money is spent.
The portfolio move that fits the clock
The deepest adjustment is to stop treating each AI effort as a standalone investment scoped against a fixed plan, and start running AI as a portfolio funded for capability rather than for projects. A portfolio funded to build, operate, and continuously improve a set of AI capabilities can absorb the technology's churn, because the money follows the outcome rather than the obsolete spec. This is harder to justify in the appropriations process than a tidy project with a fixed deliverable — but it is the only funding shape that matches the clock the technology actually runs on. The CIO who wins this argument fields AI that improves over time. The CIO who funds projects against fixed specs fields a series of demonstrations that strand one cycle after another, each one a success that the budget structure would not let become a capability.[2]
