Imagine you want to open a small beauty salon in Virginia. Before you can serve a single customer, you must navigate a stack of rules — training hour requirements, facility standards, inspection procedures — spread across multiple agencies.

Many of those rules are buried in guidance documents that can stretch for hundreds of pages of dense legal text. Others are locked behind industry paywalls. Some rules were written to comply with statutes that have since been amended or repealed. A few flatly contradict each other.

The result is a regulatory code that, in many places, has become more of an obstacle course than a public service.

This system wasn't designed to be difficult. Like in many states, Virginia's administrative code accumulated slowly over decades, as agencies added requirements to address specific problems, with no mechanism to remove ones that had grown stale or duplicative. The result is a regulatory code that, in many places, has become more of an obstacle course than a public service.

Virginia spent the last four years trying to fix that.

With the help of artificial intelligence, the state cut 35 percent of its regulatory requirements, eliminated millions of words of guidance documents, and saved residents an estimated $1.4 billion per year.

What makes the story worth telling is not just the scale of those outcomes, but how the state got there: a methodical, grounded effort that deployed artificial intelligence not as a shortcut but as a force multiplier, offering a replicable model for governments willing to put in the work.

How It All Started

In 2022, Governor Glenn Youngkin issued an executive order creating the Office of Regulatory Management (ORM). The order instructed ORM to conduct a cost-benefit analysis on all new and existing regulations statewide with the goal of cutting regulatory requirements across the entire executive branch by 25%.

"Our focus was really on how to keep the good regulations. The restrictions that are truly necessary, in order to make people's lives better.”  - Reeve Bull, Former Director of ORM, VA

“Regulation can be good. Regulation can actually provide benefits,” explains Reeve Bull, who served as Director of ORM from 2024-25. “But over-regulation does more harm than it does good. It creates red tape, making it harder for businesses to operate and very difficult to start new ones. So our focus was really on how to keep the good regulations. The restrictions that are truly necessary, in order to make people's lives better.” 

The Early Years: Slow but Steady Progress

In its first three years, ORM made progress toward accomplishing many of the order’s key requirements.

ORM first developed a Regulatory Reduction Guide to give agencies a common framework for measuring regulatory reductions. Rather than relying on crude metrics such as page length, the guide instructed agencies to quantify the burden each requirement imposed on residents and businesses. 

For example, if an agency eliminated the fee for obtaining a professional license, that would count as reducing one requirement. If a department reduced required training for a certification from 1,000 hours to 500 hours, that would count as reducing half of one requirement. Each agency's reductions were then measured against its total discretionary requirements to track progress toward the 25 percent target.

The office then worked with agencies to analyze their regulations. Together with regulatory coordinators at each of Virginia's 66 executive branch agencies, staff methodically reviewed the state code, separating discretionary requirements from mandatory, and identifying redundant, conflicting, or ambiguous statutes. Agencies then decided which changes to pursue, initiating rulemaking where required or updating guidance documents to act more quickly. 

Using this manual review process, Bull says that ORM “exceeded or was approaching” their 25% reduction goal by early 2025. 

Laws Behind Paywalls

Virginia had made real progress. But they also ran into a new barrier. Roughly 70% of the 335,000 requirements contained in Virginia's Administrative Code were buried in "incorporated documents" — external technical standards and third-party publications referenced within regulations rather than published in the code itself.

Many of those documents contained copyrighted information and were locked behind paywalls. For ORM, this meant that the bulk of Virginia's regulatory burden was effectively beyond reach — difficult to review and rewrite, and inaccessible to the small business owners and ordinary citizens who must comply with it.

“We’d picked a lot of the low-hanging fruit, and then some,” Bull says. “So we thought, maybe it's time to revisit the AI project.”

ORM had long considered whether AI could both accelerate the review of agencies’ regulations and help address the problem of incorporated documents. “We’d picked a lot of the low-hanging fruit, and then some,” Bull says. “So we thought, maybe it's time to revisit the AI project.”

Enter AI: Reviewing Regulation at Scale

In spring 2025, ORM launched an AI pilot in collaboration with a startup called Vulcan Technologies. 

In addition to tackling the problem of incorporated documents, the pilot aimed to test whether AI could make the process of reviewing regulations and guidance documents, as agency staff had been doing by hand, more efficient.

Rather than a single model, the custom platform was built around multiple AI agents working in coordination. Agents are a type of AI that can carry out complex tasks, such as searching the web, reasoning, and communicating with other systems.

Using specialized AI agents allowed Virginia to tackle different parts of the statutory research process — such as parsing regulatory text, cross-referencing statutes, conducting cross-state comparisons, and synthesizing their findings into a single output — at a scope and scale that would have been impossible just a few years ago. 

How the AI Pilot Worked

The pilot unfolded over five phases:

1. Reading and Classifying 

First, the agentic AI platform read the entire Virginia Administrative Code (VAC), which encompasses the regulations of all 66 executive branch agencies, as well as those agencies’ guidance documents. The tool compared each regulatory requirement and guidance against the state or federal law authorizing it and assigned it to one of three buckets:

• Mandatory: required by state or federal law, 

• Discretionary: authorized but not required, or

• Unauthorized: unsupported by current law, either because the underlying statute had been repealed, the regulation exceeded its statutory authority, or the rule contradicted the statute 

Bull says that the vast majority of requirements were found to be mandatory or discretionary, while only a “tiny, tiny percentage” were unauthorized. 

This classification gave agency coordinators a clear roadmap. For every provision in their code, they could see immediately what they could change, what they had to keep, and what might require a legislative fix.

2. De-Duplicating and Simplifying 

Bull’s team then turned to consistency and duplication. 

They used the AI tool to flag any provisions addressed by more than one regulation or guidance document, both within individual agencies and across agencies. This analysis helped to identify cases where regulations were overlapping, duplicative, or contradictory.

The AI tool also assessed each requirement and guidance to identify ways to make them more readable and understandable. The tool could “take regulatory text or guidance document text and just simplify it,” Bull explains. “It could basically say ‘You are using 500 words to say something that could be said in 200 words. So, here is a way to rewrite it and actually make it shorter and simpler.’”

3. Cost-Benefit and Comparative Analysis

ORM then used the AI platform to perform a preliminary cost-benefit analysis on key regulations in the code. The goal was to identify outliers where compliance costs were high but public benefits were relatively low. 

If Virginia required 1,100 hours of training to become a barber but another state achieved the same safety outcomes with 600 hours, that indicates an area where requirements could be reformed. 

The tool also ran targeted cross-state comparisons, benchmarking Virginia's requirements against those of other states. For example, if Virginia required 1,100 hours of training to become a barber but another state achieved the same safety outcomes with 600 hours, that may indicate an area where requirements could be reformed. 

Both analyses helped agencies to gather evidence about which requirements were delivering public value and which were imposing costs on residents and businesses that could not be justified by the benefits they provided.

4. Extracting Requirements from Incorporated Documents

The team then turned to the problem of the incorporated documents. 

They used the AI tool to scan thousands of requirements housed in documents referenced in Virginia's regulations and classified each provision as either:

1) a legally binding requirement, 2) non-binding guidance, or 3) background information. 

For example, one health department regulation incorporated a publication from the federal Centers for Disease Control and Prevention that mixed mandatory immunization requirements with recommended vaccination practices. The document also included pages of clinical background with no legal significance. 

AI made it possible to extract what was legally binding, set aside what was not, and present the requirements in plain language

The AI helped to separate those layers in ways that would have required enormous staff hours to do manually, making it possible to extract what was legally binding, set aside what was not, and present the requirements in plain language that the public could more easily access and understand.

5. Reporting Back to Agencies

ORM then compiled the findings from its statutory and regulatory research into reports for each of Virginia's agencies. 

The reports provided to each agency: 

• A consolidated snapshot of the requirements and guidance captured in its regulatory documents, including those that were mandatory, discretionary, or unauthorized;

• Recommendations for how to consolidate, eliminate, or simplify requirements and guidance;

• Recommendations for language that could be added to the state code or guidance documents to make the key requirements and guidance captured in guidance documents more accessible. 

As with the manual review process, each agency reviewed the findings and decided which changes to pursue, which to set aside, and which required further review before action could be taken. 

Results and Outcomes

Over four years, Virginia's regulatory modernization initiative significantly reduced regulatory requirements. According to ORM’s End of Administration Report, agencies implemented changes that, in total:

• Reduced total statewide regulatory requirements by 35.7 percent (surpassing the original 25 percent target), 

• Cut over 12 million words from guidance documents, a 49.7 percent reduction equivalent to 20 times the length of War and Peace.

• Eliminated over 57,000 requirements found in incorporated documents.

Virginia estimates that these regulatory changes save residents an estimated $1.4 billion per year. For example, eliminating so-called “gold plating” requirements from the housing code – standards like excessive insulation or restrictions on stair riser heights, which the Department of Housing and Community Development found increased costs without providing meaningful increases in safety – reduced the cost of building a new home by 25.7%, or approximately $24,102.

As the AI pilot was completed in the last few months of the administration, it is difficult to separate its results from those of the broader regulatory decluttering initiative. Many of the recommendations emerging from the AI-powered analysis were handed off to the incoming administration. Shortly before leaving office, Governor Youngkin also issued an executive order directing agencies to continue using agentic AI as part of the required four-year periodic regulatory review process.

“It took months, and in some cases, years, to actually figure out where we should be really focusing our efforts,” Bull reflects. “Whereas the AI, in theory, can do that almost immediately.”

While Virginia introduced AI late in the process, Bull says that AI may actually be most helpful in the early stages of a regulatory decluttering effort, when states need to quickly scope the problem and identify the biggest pain points. “It took months, and in some cases, years, to actually figure out where we should be really focusing our efforts,” Bull reflects. “Whereas the AI, in theory, can do that almost immediately.”

Takeaways for Other Governments

Virginia's experience offers several important lessons about what distinguishes AI analysis that drives real change from analysis that goes nowhere.

1. AI can serve as a Force Multiplier

The Virginia pilot demonstrated that AI is not a single-use tool; it can support many different aspects of statutory research and policy research, from classifying requirements to identifying conflicts to simplifying text to benchmarking against other jurisdictions.

These functionalities can enable government teams to review regulations more efficiently. ORM used AI to complete its analysis in just a few months, compared to several years for the manual review process. The pilot was delivered by a lean team, with just one or two staff members working on the project part time together with a three-person team from Vulcan Technologies. 

AI allows teams to do more with less, conducting analyses at an unmatched breadth and speed.

2. Build Institutional Scaffolding

AI is only as effective as the policymaking environment in which it is used. 

ORM’s pilot was deployed in a context where there was a willingness and readiness to experiment with the new technology and to use what was learned. Clear metrics for success had been established in law and the initiative was championed by senior leaders. 

Before launching the pilot, ORM established working relationships with regulatory coordinators at each of the state’s 66 executive branch agencies, which created trust in the process and signaled a willingness to consider the resulting recommendations.

 Governments that skip this step are likely to find that even good analysis goes nowhere.

3. Invest in Data Readiness 

Virginia's experience also underscores the importance of getting your data in order before deploying AI tools. 

ORM's final report notes that the Virginia Administrative Code “is especially well-adapted to AI-empowered analysis: it is readily available online, very well-organized, and presented in a machine-readable format.”

But even in Virginia, gaps in data quality created problems. The AI-powered statutory research process worked best when regulations included explicit citations to their authorizing statutes, giving the tool a clear thread to follow. When citations were missing or incomplete, agents had to scan the entirety of state and federal law – a time-consuming process that also increased the risk of false positives.

States with fragmented or outdated regulatory records may need to invest in preparing that data for analysis before turning to AI — whether that means digitizing records into machine-readable formats, standardizing citation practices, or simply ensuring that every regulation is publicly available online. That groundwork isn't glamorous, but Virginia's experience suggests it is a prerequisite for getting the most out of the technology.

4. Build Safeguards Against Regulatory Evisceration 

Particularly in highly partisan policymaking environments, there is a danger that AI becomes a tool not for streamlining regulations, but for eviscerating them. A system optimized to hit reduction benchmarks could, if poorly designed or misused, sweep away rules that provide public value – environmental protections, consumer safeguards, workplace standards – alongside the genuinely redundant ones.

Virginia built two guardrails into its process to guard against this.

The first was a firm human-in-the-loop requirement. AI tools acted as what Bull calls "a small army of researchers," but officials at each of Virginia's agencies ultimately decided which changes to adopt. Executive branch agencies exist because policymakers have deep subject-matter expertise in the areas they regulate, and that expertise helps them avoid cuts that may inadvertently undermine public health, safety, or welfare.

Second, ORM’s benchmarks pushed agencies toward a more nuanced assessment of burden reduction. Rather than just rewarding agencies that deleted requirements, the Regulatory Reduction Guide gave agencies credit for modifications that meaningfully reduced compliance costs.

The cost-benefit analysis and cross-state comparisons built into the pilot reinforced this: agencies were asked not just whether a requirement could be cut, but whether it was delivering public value commensurate with its burden — and whether peer states had found less costly ways to achieve the same outcome.

The principle worth carrying into any similar effort: the goal is a regulatory code that works better for the public.

Neither guardrail is foolproof. But together they reflect a principle worth carrying into any similar effort: the goal is a regulatory code that works better for the public.

5. Look for Opportunities for Greater Engagement with Communities

Finally, Virginia’s experience shows that there remain more opportunities to involve communities in regulatory modernization efforts. 

ORM’s work included an overhaul of the Virginia Regulatory Town Hall website, where proposed regulatory changes are posted for public comment. Changes included posting regulations previously exempt from comment; requiring agencies to provide greater advance notice of proposed changes; and creating an online forum specifically for reviewing guidance documents. 

While intended to boost transparency and accountability, in practice, public comment periods are often procedural formalities rather than genuine invitations for democratic participation in the policymaking process. Comments are frequently solicited after key decisions have already been made, and regulatory documents are written in dense legal language that most residents cannot parse — limitations that concentrate meaningful participation among organized interest groups rather than the communities most affected by the rules. When residents do engage, agencies are rarely required to act on their concerns, leaving public input with limited influence on final outcomes.

While ORM's improvements to the Town Hall website show meaningful progress in making existing public commenting opportunities more visible and accessible, AI offers significant opportunities to redesign public engagement processes that are better, more scalable, and more impactful. Government can leverage AI tools to collect input in more varied and meaningful ways, aid in processing and making sense of that input, and ultimately use the public’s contributions to inform decision-making. 

Conclusion

Virginia's four-year effort to modernize its regulatory code shows what is possible when governments combine clear goals, strong institutional relationships, and the right technology. AI enabled the state to take its regulatory review process from years to months.

Virginia’s experience shows that AI is a powerful tool for regulatory modernization, but it works best when deployed into an environment that has already done the hard institutional work of building trust, establishing metrics, and preparing data. Get those foundations right, and the technology can deliver results at a scale that was simply not possible before.

Read more blogs in the Rethinking Regulation series

Read more from Reeve Bull: AI and the Future of State Regulation