State and local AI needs a capacity-dividend rule

State and local governments are under pressure to do more with constrained workforces. Artificial intelligence promises exactly what managers want: faster research, shorter case-processing time, easier document review, better resident communications and less administrative work.

The productivity case is strong. The workforce design around it needs equal attention.

The Stanford Digital Economy Lab’s August 12 employment update analyzed ADP payroll records covering millions of U.S. workers through June 2026. Employment among workers ages 22 to 25 in highly AI-exposed occupations was about 19% below where it would have been if it had kept pace with similarly aged workers in less-exposed occupations. 

The comparable measure was 15% in the July 2025 data vintage. The adjustment appears mainly through reduced hiring, especially where AI use tends to automate human tasks. Experienced workers show no comparable gap.

For government, the lesson is practical: do not let an AI efficiency program become an accidental succession program in reverse.

States are already trying to widen talent pipelines through skills-based hiring. That effort should now extend beyond how people enter government to how they become capable once they arrive.

A useful operating rule is to treat a portion of AI savings as a capacity dividend.

If an agency saves 1,000 staff hours by automating routine preparation, it should deliberately reinvest some of those hours in supervised work that builds judgment. Junior analysts should verify AI-generated summaries against source records. 

New caseworkers should handle edge cases with experienced review. Early-career procurement staff should evaluate exceptions and explain tradeoffs before a senior official approves the decision. Communications staff should test AI-drafted notices for accessibility, ambiguity and unintended consequences.

The objective is not to preserve low-value work. It is to replace lost repetition with higher-value practice.

Route Fifty has documented how training city staff on AI can improve service delivery. The next step is to make that training part of operating design rather than a separate course.

Consider a permitting office. AI can summarize an application, compare it against standard requirements and draft a response. The junior employee should not simply click approve. The learning task becomes identifying which facts the model missed, recognizing when a rule conflicts with local circumstances, deciding what requires escalation and explaining the final decision to an applicant.

Or take a health department. AI can draft surveillance summaries or public notices. A newer employee can verify the underlying evidence, test whether the message fits different communities and explain where uncertainty changes the recommended action.

These are better learning experiences than routine transcription or formatting. They also require deliberate access to experienced reviewers.

That is why agencies should measure time to independent competence alongside hours saved and cases processed. How long does it take a new analyst to handle a standard case without correction? How accurately does a junior employee identify an exception? How often are escalations appropriate? How much rework occurs as responsibility expands?

These measures make workforce capacity visible.

A government can report that an AI system saved 10,000 hours and still be worse off if fewer people are becoming capable of handling difficult cases. Another agency can save the same number of hours, reinvest a fraction in structured coaching and emerge with a stronger bench.

The capacity-dividend rule also gives public managers a concrete answer to employee anxiety. The message becomes: we will automate work that should be automated, and we will use part of the gain to move people into more responsible work faster.

This approach matters most in agencies with retirements, hard-to-fill roles or thin specialist teams. When experienced staff leave, an organization cannot instantly purchase the judgment they accumulated over decades. It has to produce successors before that knowledge walks out the door.

AI can help. It can expose patterns, simulate scenarios, retrieve precedents and reduce the clerical burden surrounding expert work. But it creates resilience only when people get repeated chances to exercise responsibility under supervision.

State and local leaders should therefore add one line to every serious AI implementation plan: what learning opportunity replaces the task being automated?

Answer that question well, and AI becomes more than a cost-saving tool. It becomes infrastructure for building the next generation of public servants.

Gleb Tsipursky, PhD, a behavioral scientist, CEO of Disaster Avoidance Experts, and author of The Psychology of AI Adoption at Work: From Resistance to Results.