Key Findings

The hard truth is that AI is already shaping government, mostly to optimize old processes. If we keep chasing efficiency, we’ll entrench the status quo. If we instead aim for Adaptation, Listening, and Trust (ALT), we can reset how public institutions deliver value and rebuild legitimacy.

• Legislatures are sprinting—defensively. Since 2019, states have proposed or passed more than 1,600 AI bills, 735 in the first half of 2025 alone (~45 percent of the total).¹ Seventy-seven percent are controlling (guardrails, bans, audits), split roughly evenly between constraining government use and regulating private providers.
• Policy focus is fragmented across education, safety, health, elections, etc.—a sign of no shared focus or paradigm.
• Workforce training matters. States are pairing adoption with mandatory training. About 100,000 public employees are enrolled in AI workforce development; some states require completion before staff touch AI tools.²
• State and city approaches differ.
  • States are advancing a whole-of-government approach by establishing safe environments to experiment with AI, evaluate results, and then scale.
  • Cities are more oriented towards stand-alone pilots addressing urgent problems.
• Chief information officers (CIOs) are driving adoption. Unlike in prior tech eras, information technology (IT) leads are the main actors writing the rules, conducting outreach, and publishing what works.
• Ecosystem readiness is uneven but promising.
  • Philanthropy is beginning to shift from internal productivity to field-building.
  • Higher education is investing heavily in AI, and the best programs now deliver production-grade civic projects.
  • Peer coalitions are moving the field from expert-led to operator-driven practice.

Diagnosis: By optimizing machinery that residents already distrust, we are building momentum without vision or a framework.

We recommend a governance—not government—approach organized around a model to adapt, listen, and trust.

A — Adapt: Plan for the Demand AI Unleashes

• Forecast demand before product or system launch; stress-test capacity; reallocate budgets and roles in advance.
• Use AI agents for routine steps; reserve discretion for humans.
• Redeploy people to work requiring high-level judgment; train staff for compassionate decision-making and bias awareness.
• Pass enabling policies (not just controls) so services can iterate as demand patterns change.

L — Listen: Clarify Needs, Don’t Just Translate Text

• Create a public AI sandbox where staff and communities co-translate budgets, ordinances, and services into plain language with low or no-code tools.
• Train models for context engineering—institutional memory plus data architecture—so that they understand problems, not clever prompts.
• Combine structured signals (such as reports from 311, benefits information) with unstructured inputs (e.g., meeting transcripts, proposals) for richer insight.

T — Trust: Offer Two-Way Accountability, Not One-Way Transparency

• Stand up civic data trusts and community-controlled data; explore sovereign local models.
• Evaluate for trustworthiness (based on fairness, responsiveness, and usefulness) alongside efficiency.
• Tie resident sentiment to impact metrics, including environmental stress, service usage, and safety.
• Create public compacts with shared goals, timelines, and measures.

If we aim AI at efficiency, we’ll get faster bureaucracy. If we aim it at ALT, we’ll get a civic sector that residents actually believe in.