States Are Setting the AI Ground Rules - Now Build the Capacity to Match
AI can speed license approvals, improve K-12 outcomes, and standardize compliance across agencies. The promise is real, and residents will judge us by how well these systems serve them.
But there's no standard playbook. Without federal guardrails, states are writing policy and implementing it with limited staff, mixed vendor quality, and tight budgets. Agility is an advantage only if it's paired with the right talent, process, and transparency.
Build Trust With Transparent AI Use-Case Inventories
A public AI use-case inventory lists the systems your agencies use, where they're used, and why. It's simple: if an algorithm touches eligibility, benefits, taxes, or any public-facing decision, people deserve to know.
- Improves visibility into automated decisions and outcomes
- Enables third-party testing, bias checks, and academic research
- Helps agencies learn from each other's successful deployments
- Signals expectations to vendors and raises product quality
The federal government has produced recurring AI inventories since Executive Order 13960 in 2020. That baseline transparency helped agencies see where AI shows up in services and where oversight is needed. You can review the order here: Executive Order 13960.
Minimum Viable Inventory: What to Publish
- Agency, program, and point of contact
- System name, function, and vendor (if any)
- Decision impact level (e.g., informational, staff support, eligibility, enforcement)
- Datasets used and data sources
- Human oversight: who reviews, when, and how to appeal
- Evaluation: fairness/accuracy testing, monitoring cadence, and latest results
- Procurement method and contract reference
- Launch date, pilot status, and sunset or reauthorization date
Common Pitfalls and How to Avoid Them
- Underreporting: Without clear definitions, agencies miss tools that "assist" decisions. Fix with a plain-language scope and examples.
- Inconsistent entries: Free-text surveys produce noise. Fix with a standard template and required fields.
- Low response rates: Email-only polling fails. Fix with an executive memo, deadlines, training, and escalation paths.
- Stale data: One-off reports age quickly. Fix with quarterly updates and an inventory owner accountable for quality.
Case Study: California's Stalled Inventory
In 2023, California passed AB 302 directing the state to inventory "high-risk" automated decision systems. The first deadline passed with a public note that reported none were in use - a claim that contradicts multiple agency disclosures.
Here are examples already in the public record:
- Government Benefits - Covered California: Automated document processing for health insurance eligibility. Link
- Governance - CA Department of Finance: Generative AI to assess fiscal impacts of bills and effects on the state budget. Link
- Taxation - California Department of Tax and Fee Administration: GenAI tools to assist staff in responses to taxpayers and businesses. Link
- Government Benefits - Employment Development Department: A Thomson Reuters algorithm scores the likelihood of fraud in unemployment applications. Link
- Government Benefits - California Student Aid Commission: Two-way chatbot engagement for students applying for aid. Link
- Government Benefits - CalHHS: Data Exchange Framework uses algorithms to match healthcare records across systems. Link
- Transportation - Caltrans: Pilots in traffic safety, congestion, and staff research support including data analysis and report drafting. Link
The takeaway is simple: transparency is a talent and process problem. A quick email survey won't surface real use. Clear guidance, hands-on implementation support, and accountability are required to make inventories accurate and useful.
Scale Technical Talent Fast
Passing a bill is the easy part. Implementation requires people who can evaluate models, read vendor claims, write procurement language, and run monitoring.
- Use skills-based hiring: Drop degree screens. Hire for model evaluation, data engineering, and MLOps skills with practical exercises.
- Stand up a small "AI Implementation Unit": Central staff who own the inventory, publish guidance, and support agency pilots.
- Leverage fellowships: Bring researchers and practitioners into agencies and the legislature with term-limited placements and clear deliverables.
- Replace generic consulting with shared services: Build repeatable evaluation checklists, vendor scorecards, and model test harnesses once, then reuse.
- Upskill existing staff: Pair program experts with technologists and provide focused training pathways. A curated, job-based course list can speed this work: courses by job.
For evaluation standards and risk controls, align agency guidance to the NIST AI Risk Management Framework. It's practical, vendor-neutral, and widely referenced.
90-Day Action Plan for State Leaders
- Week 1-2: Issue an executive memo defining "AI/ADS," scope, and reporting obligations. Name an inventory owner and agency AI leads.
- Week 2-3: Publish a standard inventory template with required fields and examples. Offer office hours and a help channel.
- Week 3-6: Pilot the template with three agencies that have known use-cases. Fix friction before statewide rollout.
- Week 6-8: Launch a central submission portal with validation checks. Require executive sign-off for each entry.
- Week 8-10: Publish the inventory with plain-language summaries and an appeals contact for each system.
- Week 10-12: Release procurement clauses: data access for audits, evaluation requirements, human-in-the-loop controls, and sunsets for high-risk tools.
- Ongoing: Quarterly updates, monitoring reports, and a public change log. Tie funding to compliance.
Conclusion
States are now the front line for AI accountability. The path is straightforward: publish clear inventories, bring technical talent inside government, and give agencies hands-on support to implement well.
Do that, and residents get better services, fewer hidden risks, and more trust. Miss it, and even the best-written laws stall before they help anyone.
Your membership also unlocks:
