Virginia JLARC Audit: State Workers Favor ChatGPT Over Enterprise Copilot, Exposing Shadow AI Risks

On Tuesday, October 13, 2026, Virginia’s Joint Legislative Audit and Review Commission (JLARC) presented its sweeping study, Governance and Oversight of Artificial Intelligence in Virginia, after lawmakers directed the review in 2025. According to reporting by The Center Square and JLARC’s official findings, state employees are adopting artificial intelligence at a breakneck pace, while the Virginia Information Technologies Agency (VITA)—the central body responsible for setting IT standards across the state's executive branch—is lagging behind on basic guardrails, risk tiers, and training.

What makes this legislative audit compulsory reading for Windows administrators, Microsoft 365 architects, and CISOs everywhere isn't just the public-sector politics. It is a textbook case study in "Shadow AI": even though Virginia rolled out an enterprise-protected version of Microsoft Copilot in August 2025, state workers have continued flocking to public, consumer-grade chatbots, sending more prompts and exchanging more data with public ChatGPT than with their state-provisioned Copilot environment.

Let’s unpack JLARC’s findings, examine why uniform bureaucracy is failing both low-risk innovation and high-risk security, and translate Virginia’s governance gaps into actionable lessons for your own organization.

The Enterprise Copilot Paradox: Provisioning Isn't Adoption​

In the enterprise IT world, there is a comforting myth that once you buy the licenses and light up an enterprise AI tenant, your data-leakage worries are over. Virginia’s telemetry tells a starkly different story.

According to JLARC’s review, 97 percent of surveyed executive-branch and independent state agencies permit staff to use general-purpose AI tools—such as ChatGPT, Anthropic’s Claude, Google Gemini, and Microsoft Copilot Chat—as well as domain-specific coding assistants like GitHub Copilot, Claude Code, and OpenAI Codex. During a single one-month window examined this summer, JLARC estimated that 76 percent of employees at VITA-supported agencies used a general-purpose AI tool to draft emails, summarize documents, or write and improve code.

To its credit, VITA began providing an enterprise version of Microsoft Copilot with enhanced commercial data protections to all supported agencies back in August 2025. As enterprise admins know, an organization-managed Copilot deployment offers crucial administrative controls: tenant data isn't fed back into public foundation models for training, and IT can govern feature access.

Yet despite having enterprise Copilot on the desktop, state employees frequently reached for publicly available tools instead. According to The Center Square’s reporting on the JLARC review, during the period analyzed, state workers sent more requests to ChatGPT—and exchanged more data volume with it—than they did with the state-provided Microsoft Copilot.

Clarifying the Risk vs. Reality​

Before anyone hits the panic button, it is vital to apply strict journalistic precision to what the telemetry actually shows:

• No Confirmed Data Breach: JLARC explicitly noted that VITA’s network usage data does not reveal the specific contents of the prompts employees entered or how they used the outputs. The 76 percent adoption figure and the heavy ChatGPT traffic do not prove that sensitive citizen records or classified state data were leaked.
• The Governance Gap: Rather than proving a breach, the data exposes an unnecessary vulnerability. Because Virginia lacks a statewide acceptable-use policy (AUP), mandatory staff training, and a clear, publicly posted list of approved AI tools, employees are navigating powerful consumer platforms without guardrails.

Why do users bypass a free, work-provided enterprise Copilot to use public ChatGPT or Claude? Based on general industry knowledge, this happens for three reasons: muscle memory formed on personal devices, confusion over which browser bookmark represents the "safe" enterprise login (especially given Microsoft’s shifting naming conventions around Copilot Chat), and the absence of network-level blocks or browser warnings steering users away from unvetted consumer endpoints.

Section Summary​

• Widespread Use: An estimated 76% of VITA-supported state employees used general-purpose AI in a one-month sample period, with 97% of surveyed agencies allowing some form of access.
• Shadow AI Dominates: Despite having access to an enterprise version of Microsoft Copilot since August 2025, employees sent more requests and data to ChatGPT than to the state-managed Copilot.
• Missing Guardrails: Virginia currently lacks a statewide acceptable-use policy and a clear directory of approved general-purpose AI tools.

When "One-Size-Fits-All" Oversight Breaks Down​

Beyond general-purpose chatbots, JLARC separately audited function-specific AI—systems built for distinct operational tasks like answering routine public inquiries, transcribing meetings, or detecting fraud. As of June 2026, roughly one-third of the 80 surveyed agencies were actively using or piloting at least one function-specific AI tool, and another 30 percent were actively considering or pursuing one.

Here, JLARC uncovered a classic bureaucratic Catch-22: VITA’s governance rules apply almost uniformly across the board, regardless of how dangerous or benign an AI tool actually is.

As study project leader Drew Dickinson pointed out during Tuesday’s presentation, VITA recently amended its policy to mandate annual audits for "high-risk" AI systems—yet failed to actually define what constitutes "high risk." Furthermore, while VITA requires agencies to identify risks during the AI approval process, it hasn't given them a standardized assessment instrument to do so. Nearly half (48 percent) of executive-branch agencies using function-specific AI cited unclear state requirements as a major or moderate constraint on their work.

Uniform requirements create the worst of both worlds:

1. Stifling Low-Risk Productivity: Treating a simple internal meeting transcription tool or a basic document formatter with the same heavy compliance machinery as a public-benefits decision engine delays innocuous projects and frustrates agency IT staff.
2. Under-Protecting High-Risk Systems: When oversight isn't tailored by severity, systems that directly affect residents' rights, public safety, or government benefits may not receive the rigorous auditing and human-in-the-loop safeguards they demand. At least eight other states already differentiate their AI requirements by risk tier.

The Agentic AI Blind Spot: Governing from 2023​

In the fast-moving world of enterprise technology, November 2023 might as well be the Bronze Age. Yet according to JLARC, November 2023 was the last time VITA updated its AI architecture standard.

In the nearly three years since that document was touched, the software industry has pivoted hard from passive conversational chatbots to agentic AI—autonomous systems capable of planning multi-step workflows and taking independent action across organizational files, databases, and APIs without a human approving every click.

JLARC warned that VITA’s outdated standards make no reference to AI agents and lack specific access controls for tools that can autonomously modify, delete, or share state files. For Windows and Azure administrators watching Microsoft push autonomous Copilot Studio agents and deeper OS-level agentic integrations, this is a glaring cautionary tale. Static software governance written for the early days of text-generation LLMs is completely unequipped to handle service principals and AI agents that possess read/write permissions on live production storage.

Commonwealth Chief Information Officer Mike Watson told the commission on Tuesday that VITA agrees with JLARC’s recommendations and already has training and technology pilots underway. Watson also offered a pragmatic observation that every enterprise architect will appreciate: before agencies simply bolt on more AI tools, they need to prepare their underlying data, business processes, and core technology infrastructure.

High Stakes: 850 Legacy Apps and a $270 Million SNAP Cliff​

To JLARC’s credit, the report is not an anti-AI hit piece. In fact, the commission highlighted two urgent areas where moving too slowly on AI could cost Virginia taxpayers dearly.

1. Rescuing 850+ Aging State IT Applications​

As of August 2025, VITA had cataloged more than 850 state IT applications in need of modernization, including over 300 requiring urgent attention. Anyone who has maintained decades-old government codebases knows the pain of technical debt. JLARC recommended that VITA establish an expedited review process for AI coding assistants (such as GitHub Copilot or Claude Code), publish clear technical and risk-mitigation guidance for developers, and evaluate a statewide procurement contract so agencies can safely use AI to refactor legacy software.

2. The $270 Million SNAP Error Rate Problem​

Under federal changes enacted in H.R. 1 in 2025, states whose Supplemental Nutrition Assistance Program (SNAP) payment error rates reach 6 percent or higher are required to help pay for benefit costs out of state coffers.

In fiscal year 2024, Virginia’s SNAP payment error rate stood at 11.50 percent. In January 2026, the Virginia Department of Social Services (VDSS) estimated that if the commonwealth’s error rate remained at that 11.5 percent mark, Virginia would face $270 million per year in state benefit cost contributions starting in fiscal year 2028. (It is important to note that this $270 million figure is a conditional projection based on the FY2024 error rate holding steady, not a guaranteed invoice or a direct measurement of AI savings.)

As of September 2026, VDSS had already deployed two AI-enabled tools to help prevent and detect SNAP payment errors and identified several more candidates. JLARC urged VITA to give priority review to VDSS proposals aimed at driving down those error rates before the FY2028 fiscal cliff arrives.

The Legislative Dilemma: Why Writing AI Law Is So Hard​

While much of the report focuses on internal executive-branch IT operations, JLARC also zoomed out to look at the General Assembly’s efforts to regulate AI across the broader economy—a topic of intense interest in Richmond, where Governor Abigail Spanberger recently urged federal lawmakers to establish baseline national AI guardrails while demonstrating that innovation and safety don't have to conflict.

JLARC cautioned state lawmakers that regulating AI through statute is fraught with traps:

• Definitional Quicksand: Without a universally accepted technical definition of "AI," laws easily become overly broad. JLARC pointed to a 2024 Colorado AI law whose sweeping definition of "AI systems" sparked widespread confusion over what software was covered, eventually forcing Colorado to repeal and replace the measure with a narrower definition.
• The "Black Box" Compliance Problem: Laws requiring deployers to explain the exact principal reasons behind an adverse AI-influenced decision often collide with technical reality, because the internal weights and reasoning of advanced neural networks are not always fully explainable.
• Constitutional and Legal Overlaps: Many existing Virginia laws (such as consumer deception statutes) are technology-neutral and already penalize bad outcomes regardless of whether a human or an algorithm pulled the trigger. Meanwhile, poorly drafted state rules risk running afoul of First Amendment speech protections, due process, or the U.S. Constitution’s Commerce and Supremacy Clauses.
• A Shortage of Tech-Law Brains: Perhaps most tellingly, JLARC found that as of September 2026, the state’s Division of Legislative Services (DLS)—the nonpartisan agency that actually drafts bills for lawmakers—did not have a single attorney on staff (or on retainer) specializing in technology law.

To fix these gaps, JLARC recommended that the General Assembly fund ongoing technology-law expertise for DLS, direct the Joint Commission on Technology and Science (JCOTS) to publish an annual legislator’s guide to AI by October 1 of each year, and pass legislation mandating annual AI literacy training for all state employees. Because JLARC is an oversight body, its recommendations do not establish binding rules on their own; they require executive execution by VITA or statutory passage by the legislature.

Practical Takeaways for Enterprise IT and Windows Admins​

You don't need to work for the Commonwealth of Virginia to recognize the warning signs in JLARC’s audit. If your organization has rolled out Microsoft 365 Copilot, GitHub Copilot, or enterprise ChatGPT, ask yourself how your environment would fare under the same scrutiny:

1. Audit Your "Copilot vs. Consumer AI" Traffic: Like VITA, you may have paid for enterprise Microsoft Copilot licenses assuming employees would naturally gravitate to them. Inspect your firewall, Secure Web Gateway (SWG), or Microsoft Purview Insider Risk and Cloud App Security logs. Are users still pasting corporate data into personal ChatGPT or consumer Claude accounts because nobody showed them how to access the enterprise-protected endpoint?
2. Make Training Mandatory, Not Optional: VITA staff initially told auditors they were considering voluntary AI training, even while 76 percent of agencies pursuing AI said poor staff AI literacy was holding them back. Voluntary cybersecurity training rarely reaches the people who need it most. Bake AI acceptable-use rules into mandatory onboarding and annual security awareness recertification.
3. Retire Pre-Agentic Governance: If your corporate AI policy was written in 2023 or 2024, it probably talks about prompt hygiene and text generation. Update your security baselines today to explicitly address agentic AI—establishing strict least-privilege RBAC, service-account boundaries, and human-approval gates before any AI agent is allowed to modify, delete, or share files on your network.
4. Tier Your Risk Reviews: Stop putting a developer who wants an AI regex helper through the same six-month security review as an HR director deploying an automated resume-screening algorithm. Build a clear Low / Medium / High risk matrix so your security team can fast-track safe productivity wins while focusing deep auditing on systems that touch sensitive data or human livelihoods.

Virginia’s audit is a refreshing dose of transparency in a tech cycle usually dominated by vendor hype. The tools are already on your users' screens—the only question is whether your governance policy is living in 2026, or still stuck in November 2023.