You don’t need to be an AI expert to lead on this. You need to make three decisions: which tools are allowed, what information can go into them, and who checks the work. This page helps with all three, and with a low-risk way to try AI in one department first.
The real risks
These are the risks worth your attention. Each one has a practical fix.
| Risk | What it looks like | What helps |
|---|---|---|
| Data exposure | Someone attaches a voter’s email or another file with personal information to a chat, or drops it in an AI project folder, on a free plan. Depending on the plan and its settings, that chat may be kept and used to train future models. | Clear green, yellow, and red rules. Business-tier plans for yellow information. Training turned off. See the Safe Use Framework. |
| Wrong answers, published | AI states a deadline, statute, or procedure with confidence and gets it wrong, sometimes borrowing another state’s rules. If it ends up on your website, it’s your office’s mistake. | The people who know the work read every page before anything is published or sent. Anything outside their expertise goes to someone who has it. |
| Public records gaps | Office business done in personal AI accounts creates records your office can’t search, retain, or produce. | Work accounts only. Save final outputs. Follow your records management liaison officer’s (RMLO’s) guidance. See AI conversations can be public records. |
| Shadow AI | Shadow AI means staff using AI tools the office doesn’t know about, often on personal accounts or phones. A ban with no approved option doesn’t stop busy people. It just means you don’t know what’s happening. | An approved tool, simple rules, and a way for staff to ask questions without getting in trouble. |
| Vendor terms | Consumer and business plans have different rules on training, data retention, and admin control, and vendors change them often. | Have your office’s attorney review the terms before you buy, and recheck them at renewal. See Will They Train on Our Data? |
| AI taking actions | Newer tools, sometimes called agents, can send email, change files, and run on a schedule without anyone watching. | No connections to email, voter systems, or your Microsoft 365 tenant (your office’s private space in Microsoft’s cloud) without IT and legal approval. See Going Further. |
The risk of doing nothing
Saying nothing is also a decision. On October 2, 2026, we checked Online Sunshine and the list of 2026 Laws of Florida and found no Florida law that sets rules for how government offices use AI (how we checked, and what the law says). The laws you already follow still apply, like public records and the confidentiality of voter information, but the day-to-day rules are up to you.
If your office hasn’t set rules, assume some of your staff are already trying these tools on their own, with their own settings and their own judgment about what’s safe. You won’t know which tools, which accounts, or what went into them.
There’s also a cost in time. Two branded poll worker decks built from the state manual take about 4 to 6 hours by hand. In a test run, the AI’s part took about 20 minutes: it found the manual, downloaded it, and built both decks in the project folder. (Here’s how it works.)
Large counties have IT departments to help work this out. A three-person office with no IT staff has the most to gain from time saved and the least help getting started. Falling behind isn’t a crisis, but it’s easier to start small now than to catch up later.
The answer isn’t “yes to everything” or “no to everything.” It’s a short policy, one or two approved tools, and a small pilot.
A sample one-page office AI policy
Download the sample policy (a plain text file; open it in Notepad and paste it into Word to edit).
[COUNTY] County Supervisor of Elections: Use of Artificial Intelligence (AI) Tools
Adopted: [DATE]. Policy owner: [NAME AND TITLE]. Next review: [DATE].
1. Purpose. This policy explains how employees may use artificial intelligence (AI) tools for office business. AI tools are software that write, summarize, analyze, or create text, images, audio, or code in response to instructions. The goal is to save time on routine work while protecting voter information, public records, and public trust.
2. Scope. This policy applies to all employees, temporary staff, [volunteers, and contractors] who use AI tools for office business, on any device, in or out of the office. It includes AI features built into other software, such as Microsoft Copilot.
3. Approved tools. Only these tools may be used for office business:
- [TOOL AND PLAN], approved for [green only / green and yellow] information
- [TOOL AND PLAN], approved for [green only / green and yellow] information
To request another tool, contact [NAME AND TITLE]. New tools need approval from [IT CONTACT] and the office’s attorney.
4. Data rules. Before putting any information into an AI tool, decide what color it is (see the Safe Use Framework). If you’re not sure, treat it as the next color up and ask your supervisor.
- Green (public): published documents, statutes and rules, public notices, press releases, and published statistics. Any approved tool.
- Yellow (internal or unchecked): internal procedures, drafts, vendor contracts, staff email, and anything that might contain personal information. Approved business-tier tools only, with personal information removed first. For a report or export from any system, delete every column that holds personal information before the file goes in the AI project folder. The full export stays out.
- Red (confidential, exempt, or security-sensitive): Social Security numbers, driver license and Florida ID numbers, voter signatures, protected voter records, passwords and system credentials, network and cybersecurity details, ballot security details, and anything under a litigation hold. Never put red information into an AI tool unless IT and the office’s attorney approve a specific tool and use in writing.
5. Human review. A person is responsible for everything AI helps produce. Before AI-assisted work is published, sent outside the office, or used to make a decision, the employee reads all of it and checks every fact, date, number, name, and legal citation against the official source. Work outside the employee’s area of expertise is also read by someone with that expertise. Public-facing materials are also approved by [NAME OR TITLE].
6. Public records and retention. Prompts, AI responses, and outputs created for office business may be public records under Chapter 119, Florida Statutes. They are kept according to the State of Florida General Records Schedules (the state’s retention schedules; GS1-SL for state and local government agencies and GS3 for election records, adopted by Rule 1B-24.003, F.A.C.), as directed by the office’s records management liaison officer (RMLO). Save final work products in the office’s normal filing system. AI chats that are public records are covered by public records requests and litigation holds like any other record.
7. Accounts. Use only office-approved work accounts for office business. Don’t do office business in personal AI accounts, and don’t use work accounts for personal use. Turn off model training in every account where the setting exists. Yellow information may be used only in the business-tier accounts approved in section 3.
8. Prohibited uses. Employees may not use AI tools to:
- make final determinations about any voter’s eligibility or registration status;
- write, change, or paraphrase statutory ballot language or official ballot text;
- publish or send AI-generated content without human review;
- process red information; or
- connect an AI tool to the voter registration system, office email, or the county’s Microsoft 365 or Google Workspace environment (tenant) without written approval from IT and the office’s attorney.
9. Labeling AI-generated media. Images, audio, and video created or substantially changed with AI and used in office communications are labeled [LABEL TEXT, such as “Image created with AI”]. AI is never used to create realistic images of real people, ballots, voting equipment, or polling places that could be mistaken for photographs. [OFFICE DECISION: whether AI-assisted written material also needs a label.]
10. Training. Employees complete [TRAINING] before using AI tools for office business and review this policy [each year].
11. Reporting problems. Tell [NAME AND TITLE] right away if red or yellow information goes into an unapproved tool, if AI-generated content with an error is published, or if an AI tool does something unexpected. Don’t delete the chat or file. It may be needed to understand what happened, and it may be a public record.
12. Review. This policy will be reviewed by [DATE] and at least once a year, because AI tools and vendor terms change often.
Approved by [NAME], Supervisor of Elections. Date: [DATE].
Pilot with one department
Don’t roll AI out to the whole office at once. Try it with one department for 30 days, measure it honestly, and then decide.
- Pick one department and one lead. Choose a team that does a lot of repeat writing from public documents and has at least one person who’s curious about AI. Communications and poll worker training are natural fits.
- Pick a quiet month. Not during early voting, Election Day, or canvassing.
- Pick two or three green tasks. Choose tasks that use only public information and come up often: social posts from approved press releases, training slides from the public manual, or plain-language summaries of a new Division of Elections publication. Give each task its own project folder. The two poll worker decks prompt is a good first one, and the Use Case Library has step-by-step versions of the rest.
- Put the scope in writing. A short email from you approving the department, the tool (the Claude or ChatGPT desktop app, on which plan), and the tasks is enough. Have IT approve the install first.
- Measure the baseline. Before anyone uses AI, time each task the usual way a few times and write it down. Without a baseline, “it feels faster” is all you’ll have.
- Run it for 30 days. Use the approved desktop app, work accounts, one project folder per task, and your policy. AI Basics has a week-by-week plan your pilot lead can follow.
- Check in every week. Keep it short: what worked, what didn’t, any close calls with data, and any mistakes the review caught.
- Keep a simple log. One row per task, every time (template below).
- Decide at day 30. Use the go/no-go guide below, and write the decision down. It’s a record, and it helps the next department.
Pilot log template
| Date | Task | Time the usual way | Time with AI, including review | AI errors caught in review | Notes |
|---|---|---|---|---|---|
What to measure
- Time saved: the baseline compared with the time using AI, including review and fixes.
- Error catches: the AI mistakes your review caught, like wrong dates or made-up facts. Catches show the review step is working. A task that keeps producing them may not be a good fit.
- Data issues: any time yellow or red information came close to going into the tool.
Go / no-go at day 30
- Go: real time saved after review, no data incidents, and staff want to keep using it. Adopt your written policy and expand to one more department.
- Adjust: helpful for some tasks but not others. Keep the ones that work, change or drop the rest, and run another 30 days.
- No-go: review takes as long as doing it by hand, or the data rules weren’t followed. Stop, and write down why. That’s a useful answer too.
Questions to ask your IT provider or MSP
Many offices get IT from the county or from a managed service provider (MSP), an outside company that runs your computers and network. Before you approve any tool, ask them:
- Which AI tools already come with our Microsoft 365 (or Google Workspace) licenses? Is Microsoft Copilot Chat turned on, and does it show the green shield (enterprise data protection: your prompts stay inside our Microsoft 365 and aren’t used to train Microsoft’s AI)?
- Are we in a government cloud (such as Microsoft’s GCC)? Which AI features are available there, and which arrive later?
- Can staff connect outside apps, including AI tools, to our Microsoft 365 tenant on their own? Can that be limited to IT approval?
- Are our SharePoint, OneDrive, and Teams permissions tight enough that an AI assistant wouldn’t surface files people shouldn’t see?
- How would we keep, search, and produce AI chats for a public records request or a litigation hold?
- Can we install the Claude or ChatGPT desktop app on a county PC, or should AI work happen on a separate workstation off the county network? (See Safe Setup.)
- Should AI tools be blocked on computers that connect to the voter registration system or other election systems?
- How will we know which AI tools staff are actually using?
Questions to ask AI vendors
Ask these in writing, and keep the answers with your purchase records.
- Training: Is our data used to train your models by default? Can we turn training off for the whole organization? Can we turn off the thumbs-up and thumbs-down feedback buttons, which can send a whole conversation to the vendor even when training is off?
- Retention: How long do you keep our prompts, responses, and files? Can we set our own retention period? Can we export chats to answer a public records request?
- Admin controls: Do we get single sign-on, central user management, audit logs, and control over which connectors staff can use?
- Where data goes: Where is our data processed and stored? Is it kept in the United States? Which other companies handle it?
- Government terms: Do you offer government terms or pricing for a Florida county? Through which contract vehicle? Can our attorney review the terms and any data processing agreement before we sign?
- Public records clause: Will you accept Florida’s required public records terms? Florida requires public agency contracts for services to include clauses on keeping, providing, and protecting public records (s. 119.0701, F.S.). Your attorney will know how this applies to an AI subscription.
- Leaving: What happens to our data if we cancel?
Government offers to ask about
OpenAI (ChatGPT): OneGov 2.0
The offer takes effect October 1, 2026 (GSA: "expected to take effect") and runs through December 31, 2028. OpenAI says its $15-per-user monthly license fee is $0. GSA (the federal General Services Administration) says there is no platform fee, no minimum order, and no spend commitment, and that usage is billed at 50% off. So this is pay-for-what-you-use pricing, not a free ChatGPT seat. GSA says federal, state, local, and tribal governments can order directly from OpenAI, through resellers, or through cloud marketplaces. Neither OpenAI nor GSA names a ChatGPT plan or spells out sign-up steps for a county, so register interest with OpenAI's government team and have your attorney read the terms.
Anthropic (Claude): Claude for Government
It became generally available on September 30, 2026, and runs in a FedRAMP High authorized environment (FedRAMP is the federal security approval program for cloud services). Anthropic says there are no seat fees: agencies prepay for usage in fixed increments, with a hard cap. It is a desktop app only (no web or mobile) and includes Claude Code. Anthropic lists the Claude Code command-line version and Claude for Microsoft 365 as early access. Anthropic's product page now says it is available to U.S. federal, state, and local government agencies, with no seat minimum; ask Anthropic's public sector team about price and whether your county qualifies. Anthropic also says Claude Enterprise fits government uses that don't require FedRAMP authorization. On October 27, 2026, Anthropic announced a GSA OneGov agreement: 40% off Claude Enterprise, Claude for Government, and its API, which Anthropic says extends to state, local, tribal, and territorial governments, with no minimum spend. It names no Florida county and no sign-up steps, so have your attorney read the terms.
Anthropic: Claude for Government · General availability post · Get started with Claude for Government · Anthropic: OneGov agreement with GSA
Microsoft: Microsoft 365 G7 and Copilot in government clouds
Microsoft announced Microsoft 365 G7 on September 15, 2026. It became available to eligible GCC customers (GCC is Microsoft's Government Community Cloud) on October 1, with more purchase options from November 1. Microsoft hasn't published a price, and the announcement covers GCC only: it doesn't mention GCC High or DoD. Copilot features such as memory and Edit with Copilot arrive in phases after launch. The Copilot add-on is also listed for the government clouds, but Microsoft doesn't publish a government price. Microsoft's licensing page lists it for GCC, GCC High and DoD with plans including G3 and G5; its price page says the sovereign-cloud timeline is still being set, so confirm with your reseller or Microsoft account team.
Checked October 2, 2026 against the vendors' and GSA's own pages.
Federal references
Next steps
- Hand out the one-page decision brief to your deputy or senior staff.
- Print the AI Safe Use Checklist and hand it out at your next staff meeting.
- Share the FAQ with staff who have questions about AI.
- Read Going Further before anyone asks to connect AI to email or set up scheduled agents (AI that runs steps on its own).