State and local government doesn't need another vendor pitch. It needs honest answers about what AI agents can do today, what they shouldn't be allowed to do yet, how to procure them without burning a year on an RFP, and how to run a small pilot that produces a result a city council will actually look at.
That's what this guide is. It's organized by role, then by procurement, then by pilot design. If you work in state IT, a city manager's office, a county clerk's office, or a state agency — read in order. If you're a vendor reading this hoping for talking points, save your time.
The state of play in 2026
Three things changed between 2024 and 2026 that matter for public-sector adoption.
- Models got cheap enough to deploy at agency scale. Inference costs dropped roughly 10x for the same quality between mid-2023 and mid-2025, which is what made per-employee agent licensing viable for governments operating on tight budgets.
- GovCloud and FedRAMP-authorized agent products shipped. Anthropic, OpenAI, and Microsoft all have variants that meet the data-residency and authorization bars most state agencies need.
- NASCIO put generative AI in the top three state CIO priorities for the third year running. It's not exotic anymore. It's a budgeted line item in most states.
If you're in a county or city, you're benefiting from the work bigger agencies did to negotiate the contracts and run the pilots. You don't have to invent it. You have to copy it.
Use cases by role
City manager / county administrator
- Council packet prep. Read agendas, attached reports, and prior meeting minutes. Generate plain-language summaries for residents and a one-pager for council members. Saves 3 to 5 hours per cycle.
- Constituent letter triage. Read incoming mail and email, classify by topic and urgency, draft replies for staff review. Public-records-friendly because you keep the human in the loop.
- Cross-department status snapshots. Pull the weekly numbers from every department's tracker, normalize them, and produce one dashboard summary every Monday morning.
IT director / CIO
- Help desk first-pass. Triage tickets, answer common questions (how do I reset my MFA), route the rest. Most agencies see 30 to 50 percent deflection on Tier-1 tickets within 90 days.
- Vendor security review intake. Read incoming SOC 2 reports, vendor security questionnaires, and DPAs. Extract the answers, flag gaps, draft the internal review memo.
- Incident response drafting. When something breaks, the agent assembles the timeline from logs, drafts the after-action report, and prepares the public statement. Humans approve before anything ships.
Permitting, planning, and code enforcement
- Permit application pre-screen. Agent reads the application against your code, flags missing items, generates the resident-facing checklist before staff touches it.
- Code violation correspondence. Draft notices, pull case history, generate the inspection summary. Cuts notice-writing time roughly in half in agencies that have piloted it.
- Public-meeting Q&A prep. Read the comments submitted on a planning case, cluster them by theme, generate the staff response document for the public hearing.
Communications and public information
- Plain-language translation. Take a 60-page agency report, produce a sixth-grade-reading-level summary plus a one-page resident handout. Critical for accessibility and Section 508 compliance.
- Spanish (and other) translation. Multilingual agents are shockingly good in 2026. Internal review by a native speaker still required, but you start from 90 percent there instead of zero.
- Social media drafting. Pull from approved press releases and policy memos. Never publish without review — but skip the blank page.
Procurement and finance
- RFP response evaluation. Read 20 vendor responses against your scoring rubric, produce a structured comparison matrix, and flag any non-compliant items. Final scoring still done by humans.
- Invoice and contract review. Match invoices against contract terms, flag discrepancies, route exceptions. Pairs with existing AP automation; doesn't replace it.
- Budget narrative drafting. From line-item data and prior year narratives, draft the new fiscal year budget narrative for staff to refine.
HR and workforce
- Job description generation. Standardized, classification-friendly drafts based on department input.
- Onboarding Q&A. A skill loaded with your employee handbook becomes the first stop for "how do I…" questions. Cuts HR ticket volume meaningfully.
- Performance review drafting. Pull from approved goals and prior reviews. Manager edits and owns the final.
Public health and human services
- Eligibility intake support. Triage applications, surface the documentation that's missing, route to caseworkers. Caseworkers approve every decision; the agent handles the paperwork.
- Plain-language explanation. Translate program rules and forms into resident-facing plain English, with translations.
- Caseload summarization. End of week, the agent produces each caseworker's "what changed and what needs your attention Monday" briefing.
Public safety (with caveats)
This is the area where you go slowest. The use cases that work today are administrative — report drafting, FOIA response prep, training material translation. The ones that don't work today are predictive policing, automated decision-making about people, and anything involving facial recognition. Those are policy decisions, not technical ones, and most state and local jurisdictions are choosing to either ban or heavily restrict them.
Procurement: how to buy without spending a year
Path 1: Cooperative purchasing agreement (fastest)
NASPO ValuePoint, Sourcewell, OMNIA Partners, and several state-specific cooperatives now have AI agent platform contracts in place. If your agency is a member, you can buy off the cooperative without running your own RFP. This is how most cities and counties are getting started in 2026 — six weeks instead of nine months.
Path 2: Existing master service agreement extension
If you already have a Microsoft 365, Google Workspace, or major SI contract, agent capabilities are likely already available as an add-on or scope expansion. This is the path for state agencies with existing enterprise agreements — sit down with your account team and ask what's covered.
Path 3: New procurement (slowest, but sometimes necessary)
If you genuinely need a new contract, the RFP best practices that emerged in 2025 are worth following.
- Mandate FedRAMP Moderate or StateRAMP authorization for any platform handling internal data.
- Require an enterprise data agreement that prohibits use of your data to train the vendor's models.
- Demand transparency on subprocessors — which models are running where, who can see your data, and where it lives geographically.
- Build in an exit clause and data portability — you should be able to leave with your prompts, your skills, and your logs.
- Require human-in-the-loop options for any tool category that touches the public.
Several states have published their AI procurement language as open templates — Texas, California, and New Jersey are good starting points. Don't write yours from scratch.
A 90-day pilot playbook
Most agencies that try to "do AI" agency-wide fail. The ones that succeed start small. Here's the model that has worked in dozens of cities and counties.
Days 1 to 14: Pick one team and one workflow
Pick a team that wants this, not one you're forcing. Pick a workflow that happens at least weekly, takes more than 30 minutes per occurrence, involves reading something, deciding something, and writing something, and has a clear good output you can recognize when you see it.
Permit pre-screen, council packet summarization, and constituent letter triage are the three workflows we see succeed most often.
Days 15 to 30: Build the agent
This is where Build An Agent Day fits — non-developers on the pilot team build a working agent against the workflow they brought. By day 30 you have a draft agent that runs on real (sanitized) data.
Days 31 to 60: Run it side by side
The agent runs in parallel with the existing process. Staff still does the work the old way. The agent's outputs are reviewed and scored. This is where you find out what works, what doesn't, and what guardrails you need.
Days 61 to 90: Make the call
- Roll it out. If the side-by-side passed, deploy with human-in-the-loop, define the success metrics, and start the next pilot.
- Iterate. If close but not there, give it 30 more days with the lessons applied.
- Kill it. If it doesn't work, write the post-mortem, share the learnings publicly (this is government — share the learnings), and pick a different workflow.
Risks to manage
- Public records. Agent inputs and outputs that touch government work are records. Log them. Make them retrievable. Many agencies are using a dedicated audit log on top of the agent platform.
- Bias and disparate impact. If an agent helps make decisions about people, you need impact assessments. NIST has a public AI Risk Management Framework that's the de facto standard.
- Vendor lock-in. Use open standards (MCP, open source skills) where you can. Avoid skills written in proprietary formats.
- Workforce trust. Bring labor partners and employees in early. Agents that take tasks off plates land well; agents that arrive as a surprise from the CIO's office do not.
- Public communication. Don't hide it. Be plain-spoken about where AI is used in your agency, why, and what oversight exists. Trust comes from transparency.
What to do this month
- Pick one team, one workflow.
- Find the cooperative contract you're already eligible to use.
- Send three people from that team to a hands-on workshop where they leave with a working agent.
- Run a 90-day pilot using the playbook above.
- Publish what you learned.
Frequently asked questions
Can state and local governments legally use AI agents?
Yes, with appropriate procurement, data agreements, and oversight. As of 2026, most states have published AI use guidelines or executive orders. Check your state's most recent AI policy guidance and your agency's IT acceptable-use policy.
Are AI agents safe for confidential government data?
The major enterprise platforms (Claude Cowork, ChatGPT Enterprise, Microsoft 365 Copilot, Google Workspace AI) offer data-handling agreements that prevent training on customer data and provide audit logs. FedRAMP and StateRAMP authorization should be required for systems touching protected data.
What's the fastest path to procure AI agent software?
Cooperative purchasing agreements (NASPO ValuePoint, Sourcewell, state-specific cooperatives) typically have AI platform contracts in place. Six weeks is realistic; full RFPs take six to nine months.
How do I run a low-risk AI pilot in my agency?
Pick a single team and a single internal workflow (not a constituent-facing process), set a 90-day timeline, run the agent side by side with the existing process, and measure the difference. Don't let the first pilot touch public-facing decisions.
What use cases should government agencies avoid right now?
Predictive policing, automated decisions about people without human review, facial recognition in public spaces, and any high-stakes determination (eligibility, sentencing, child welfare) without strong human-in-the-loop oversight.