Here is a scene playing out in public offices right now. A grant coordinator pastes a draft narrative into a free AI chatbot to tighten it up. A guidance counselor asks an AI tool to help word a sensitive email to a parent, student details included. A permitting clerk uploads a spreadsheet of applicant information to get a quick summary.
None of these people mean any harm. Most are trying to get through a pile of work faster. But each one may have just sent sensitive public data to a service the agency never approved, never reviewed and has no contract with.
That's shadow AI, and it is already common in government.
How common?
A Thomson Reuters Institute analysis published in July found that about 27 percent of government professionals admit to using AI tools their organization hasn't approved. Across all industries, the figure was 34 percent. Those are the people willing to say so in a survey.
The same piece cites a National Cybersecurity Alliance survey finding that 43 percent of AI users have shared sensitive information with AI tools without their employer's knowledge.
State security leaders see it coming from a second direction too. In the 2026 NASCIO-Deloitte Cybersecurity Study, state CISOs said they worry that employees will enter sensitive data into AI features built into platforms they already use. They also flagged vendors adding AI capabilities without enough transparency or state-level control. In other words, some shadow AI arrives through a software update nobody asked for.
Why it's a bigger deal for SLED
Every organization has data it shouldn't leak. Schools and local governments carry some extra weight.
Student records fall under FERPA. Law enforcement data has CJIS requirements. Health information at a county clinic or a school nurse's office has its own rules. Pasting any of that into a consumer AI tool can create a compliance problem, depending on how the tool handles and retains what it's given.
Then there's public records law, which is the part many agencies haven't thought through.
In Washington State, journalists with Cascade PBS used public records requests to obtain ChatGPT conversation logs from city officials. The logs showed staff using AI to draft constituent emails, mayoral letters, policy documents and press releases. In some cases, AI-written language went out in official communications.
A California law firm that advises public agencies, Liebert Cassidy Whitmore, warned in April that requesters "are beginning to treat AI-generated content and usage logs as potential public records." The firm noted that where the data sits doesn't necessarily protect it from disclosure if the agency can access it.
For Florida agencies, this deserves a close look. Chapter 119 of the Florida Statutes defines public records broadly, covering material "regardless of the physical form, characteristics, or means of transmission" when made or received in connection with official business. We're not lawyers, and how that applies to AI prompts and outputs is a question for your agency's counsel. But it would be unwise to assume those chats are private.
Why bans don't work
The instinct is to block it. Lock down the AI websites, send a stern memo, done.
It rarely holds. People use personal phones. New AI features appear inside tools you've already approved. And a ban with no alternative pushes the behavior further out of sight, where you can't see it at all. The Thomson Reuters analysis lands in the same place: "Governance, not prohibition, is the effective response."
Staff are reaching for AI because it saves them time. That need doesn't go away when the website is blocked.
What good AI governance looks like
You don't need a 40-page framework to get started. You need a few clear decisions, written down and shared.
Approve at least one tool. Give staff a sanctioned option that runs under your agency's control, with contract terms that cover data handling and retention. If you're already in the Microsoft 365 ecosystem, start by reviewing what your licensing includes and how its data protections are configured.
Define what data can go where. A simple three-tier rule helps: public information can go into approved tools; internal information can go into approved tools with care; protected data, like student records, CJIS data and health information, stays out unless a specific tool has been cleared for it.
Settle the records question before someone else does. Work with your records custodian and legal counsel on whether prompts and outputs must be retained, and where. Check whether your AI vendors keep inputs and whether you can retrieve them.
Ask vendors about AI. Add questions to your procurement and renewal process. Has AI been added to this product? Is it on by default? Where does the data go? Can we turn it off?
Train people, then keep talking. Explain the why, not only the rules. People follow policies they understand. Make it easy to ask "is this OK?" without getting in trouble.
Decide when to disclose. If AI helped draft a public-facing letter or a board item, does your agency say so? There's no single right answer, but there should be an answer.
Start with a conversation
The point is not to stop people from using AI. It's to make sure that when they do, the agency knows about it, the data stays where it belongs and nobody is surprised by a records request.
vTECH io helps Florida school districts, cities and counties put practical AI policies in place, configure approved tools inside Microsoft 365 and review security settings so sensitive data stays protected. If your organization hasn't had the AI conversation yet, or had it and never finished, give us a call at 888-277-5918.
Sources
- Shadow AI in government: Why unsanctioned tools demand a governance response, Thomson Reuters Institute, July 28, 2026
- Public Records Act meets AI: What recent records requests reveal about ChatGPT use in government, Liebert Cassidy Whitmore, April 28, 2026
- State CISOs less confident in protecting public data, Smart Cities Dive, May 5, 2026
- Florida Statutes, Chapter 119: Public Records, The Florida Senate