The AI Pilot Era Has a Date On It

American Rescue Plan State and Local Fiscal Recovery Funds had an obligation deadline of December 31, 2024. The spend deadline is December 31, 2026.

That is this fiscal year.

Those funds paid for a large share of local government AI and innovation pilots between 2022 and 2025. Not all of them. Enough of them that the pattern is the story.

When people say the AI pilot era is ending, they usually mean it as a mood. Leaders lost interest. The technology disappointed. Neither is true, and both are more comfortable than the actual reason. The pilot era is ending because the funding mechanism that made a pilot cheap to say yes to expires on a specific day, and the day is already on the calendar.

The landing is harder than the takeoff was

A 2027 pilot cannot be paid for the way the 2024 pilot was. It has to compete for recurring general fund dollars, against police, fire, pensions and deferred maintenance, in front of a council or a board that has never had to weigh those things against an AI line item before.

The fiscal weather is not helping. NASBO projects FY2027 state general fund spending at $1.36 trillion, down 1.4% in aggregate, with the median state essentially flat at plus 0.6%. Thirteen states project FY2027 gaps totaling $26.3 billion. Seven states made mid year cuts in FY2026, the most since FY2021.

Cities are in the same weather. In the National League of Cities fiscal conditions survey, 52% of city finance officers said their city was better able to meet its financial needs in FY2025, down twelve points from 64% the year before.

None of that is a story about AI, and that is the point. The AI line item is now being judged by people who are making cuts elsewhere in the same meeting.

Six percent scaled is not a slow start

Here is the number that should bother anyone running an AI program right now.

NASCIO and Accenture found 6% of state CIOs reporting mature, scaled generative AI capability, against more than 90% who believe in the technology. The rest of the distribution is 31% defined, 50% developing and 13% exploratory.

Read that as a curve and it looks like early days. Read it as a distribution and it looks like something else, which is an entire sector parked one step short of production. Half of it is developing. Almost none of it has arrived.

Belief was never the constraint. So the fix is not another executive briefing, another strategy document or another round of enthusiasm. The fix is knowing exactly where these things die, because they die in the same places.

Five checks, in order

I have watched a lot of AI capabilities fail between the demo and daily use. Almost none of them failed because the model was bad.

They failed at a specific point, and after enough of them the point becomes predictable. There are five checks a capability has to survive, and they run in order.

Useful. Does it do something a person actually needs done? A surprising number of pilots fail right here, and the room stays polite about it.

Reliable. Does it behave on Tuesday the way it behaved in the demo? Not on average. On the specific ugly case that arrives at 6am, in front of someone who has no time to verify anything.

Governed. Who owns it. Who approved it. What is it allowed to touch. Who is accountable when it is wrong.

Integrated. Does it live inside the system where the work already happens, or does it ask someone to open a second tab? Every extra tab is an adoption tax, and adoption is a multiplier, not a rounding error.

Valuable. Did a measurable operational outcome move? Permit cycle time. Backlog cleared. Asset availability. Response time. Cost avoided. Something you could defend in a hearing without a slide.

The attrition has a shape. Most pilots die between reliable and governed. Most pilots that get declared successful and never scale died at integrated. And most deployed AI that quietly disappears in the next budget cycle died at valuable, usually long after anyone was still watching.

The federal numbers show the first half of that pattern at national scale. GAO reported that of 282 federal generative AI use cases for 2024, only 41% were implemented or operational, while 56% were still in acquisition or development. Different level of government, same two gates.

One thing has changed about which gate matters most. Two years ago the hard one was governed, because writing defensible AI policy took real work. It does not anymore. The GovAI Coalition gives governance, policy and procurement templates away free to any of its 900 member agencies that wants them. Governance is no longer the hard part.

Valuable is. Governance templates are free. Budget defense is not.

I call these the five gates to production.

What to do before the hearing

Take every AI effort in your portfolio and put each one against those five checks. Mark the first gate it fails.

Most portfolios cluster, and the cluster is the diagnosis. If yours clusters at reliable, you have a data and engineering problem and you know what to fund. If it clusters at integrated, you have a systems and vendor problem, and no amount of model improvement will touch it. If everything clears four gates and dies at the fifth, you do not have a technology problem at all. You have a measurement problem, and you have had it the whole time.

Then do the part nobody wants to do. Pick the small number that can clear all five before budget season, and stop spending attention on the rest. A hundred experiments do not become a hundred capabilities. They become about five, and choosing which five is the actual FY2027 job.

One warning on the last gate. Usage is not value. Seats are not value. Satisfaction scores are not value. A staff member saving four hours has not saved the public four hours of salary, and the distance between those two statements is where most AI business cases come apart under questioning. Code for America found only 7 states established on evaluation mechanisms while nearly every state has launched a pilot, and described the ecosystem as governance heavy and measurement light. That is the fifth gate, measured across the country.

The money moved. It did not vanish.

If you are waiting for the AI budget conversation to get easier, it will not. But it is also not going away.

State and local IT spending in 2026 is $160.2 billion and still growing 4 to 6%, and the growth is shifting toward consulting, managed services and integration rather than more software. The money did not disappear when the recovery funds did. It moved from the innovation line to the operations line.

The operations line asks harder questions. What did it cost. What changed. Who is accountable when it is wrong. Those are answerable questions, and the agencies that can answer them in the next twelve months will be operating production capabilities in 2028 while everyone else is still explaining a pilot.

The pilot era had a funding source. The next one needs a business case.