This post is the eighth contribution to Notice & Comment’s symposium on AI and the APA. For other posts in the series, click here.

Over the last four years, the Commonwealth of Virginia has established the gold standard for regulatory modernization. Virginia agencies have streamlined over 35% of the requirements in their regulations and cut over 49% of the words in guidance documents. These changes are saving Virginians $1.4 billion per year, putting money back in the pocketbooks of hardworking citizens of the Commonwealth.

In 2025, shortly after reaching the initial target of a 25% regulatory reduction, Governor Glenn Youngkin’s Administration decided to supercharge these efforts by launching an AI-empowered pilot program to further streamline regulations. Working with a small start-up firm, Virginia’s Office of Regulatory Management used AI to identify regulations with high costs and low benefits, to compare Virginia regulations to comparable regulations in other states, to find and consolidate overlapping regulations, and to streamline regulatory and guidance document text. The project identified a wide variety of additional streamlining opportunities.

A handful of states are now looking to follow Virginia’s trailblazing example. And federal agencies have also begun exploring ways to use AI to pare back overburdensome regulations.

As these federal efforts get underway, agencies in D.C. can draw on the successes of their counterparts in Richmond. Though federal regulations and state regulations differ in certain important respects, there are substantial similarities. Here are some of the possible components of a federal AI-empowered regulatory modernization initiative.

Breaking the C.F.R. Down By “Mandatory,” “Discretionary,” and “Unauthorized” Provisions

Congress delegates power to regulatory agencies. Sometimes it specifies exactly how to act (e.g., “charge a $100 fee”). But it usually leaves the agency with some discretion (e.g., “charge a fee,” without specifying the amount). And it occasionally provides the agency with almost carte blanche authority (e.g., “regulate chemical emissions to promote the public interest”).

Virginia’s pilot project used AI to compare each regulation to the statutory authority it cites and classify each regulatory requirement as either mandatory or discretionary. AI can also flag requirements that may lack statutory authority (e.g., a regulation cites a repealed statute) or that may contradict the statute (e.g., agency charges $200 fee when statute only authorizes $100). With this analysis in hand, agency officials can then decide what regulations they must (unauthorized) or can (discretionary) modify. And they can consider seeking a legislative fix for any mandatory provisions that merit revision.

And though Virginia did not use AI to assess whether a delegation is too broad, an algorithm could potentially be used to scan the C.F.R. to identify regulatory provisions that may be in tension with the U.S. Supreme Court’s holdings in West Virginia v. EPA, Loper Bright, or other seminal cases. For example, with respect to West Virginia v. EPA, if the algorithm discerns that a regulation is using an obscure or vague provision of a statute, it could then conduct a preliminary cost-benefit analysis to determine if it imposes huge net economic costs or assess if it implements a politically controversial program based on a scan of the web or social media. If so, it could be flagged for closer analysis and possible overhaul. For Loper Bright, an algorithm could compare the authorizing statute to the regulatory text and assess the extent to which there is a disconnect between the two. The tool could be told to adopt the perspective of a regulatory lawyer or a circuit court judge, and ingest existing case law to determine the extent to which the agency has overreached. Given the recency of Loper Bright, there won’t be a lot of case law with which to work, but it can use what’s available, and the analysis will get better and better over time.

Admittedly, human beings are also quite proficient at performing these sorts of tasks. But AI can accomplish in hours what might take a human reviewer months or years. And though human beings must then review the AI outputs, our experience has shown that it’s far faster to process AI-generated material than to produce it in the initial instance. In this way, AI massively expands the scope of possible analysis, potentially allowing for a scan of the entire C.F.R. that would have been impossible if left to civil servants.

Creating a “Heat Map” Showing High-Cost, Low-Benefit Regulations 

Under Executive Order 12,866 and its progeny, agencies have been required to prepare detailed cost-benefit analyses (called “regulatory impact analyses” or “RIAs”) of all “economically significant regulations.” Less than 2% of regulations qualify as “economically significant,” and even those rules seldom receive a robust cost-benefit analysis involving full monetization of all benefits and costs. And regulations only receive a cost-benefit analysis at their inception: there’s no requirement to go back and reassess old rules.

Past presidential administrations have made limited efforts to introduce some measure of retroactive analysis by launching “retrospective review” initiatives. These efforts have usually left the agencies with complete discretion as to whether to modify existing rules. And, not surprisingly, they’ve yielded paltry results. The first Trump Administration established a forcing mechanism by requiring agencies to rescind two existing regulations for each new one adopted, but agencies still got to choose which regulations to swap.

Agencies’ limited success in chipping away at the mountain of federal regulations makes sense when you consider the amount of work required to prepare an RIA. Those documents tend to run into the hundreds of pages and are virtually inscrutable to anyone other than PhD economists.

But Virginia’s experience has shown that agencies can review a much larger array of regulations by doing a more streamlined cost-benefit analysis. Under Governor Youngkin’s Executive Order 19, agencies do a cost-benefit analysis of every single regulation and guidance document issued.

And though even Virginia agencies have not typically done cost-benefit analysis on existing regulations, as opposed to new regulations, AI tools can now conduct such analyses in a matter of minutes. In our testing, the overall timeframe is around 20 seconds for a very preliminary cost-benefit analysis and 15-20 minutes for a more thorough cost-benefit analysis that scrapes the web for studies to cite.

The automated cost-benefit analysis draws from public sources to the greatest extent possible. If it cannot find a source, it makes an educated guess (and discloses the fact it’s doing so). From our experience, the tool is very faithful if it’s explicitly told to acknowledge when it cannot cite a source. While public versions of these models have been programmed with an “eager to please” mentality, it was fairly simple to override that in the controlled environment of the tools that we used.

An agency would not, of course, want to rely exclusively on an AI-generated analysis. But the AI tool can produce a heat map that shows which regulations impose especially large costs and/or feature relatively low benefits. Officials can then use these heat maps to identify the regulations justifying closer scrutiny and possible reform. 

Comparing Federal Regulations to International and State Counterparts

Once an agency has a heat map of high-cost/low-benefit regulations in hand, it still must ask “is there a better alternative?” The best way to answer that question is usually to consider approaches in other jurisdictions. But figuring out how other regulators have approached a problem is a time-consuming task. 

Here too, AI offers a solution. Virginia used AI to compare its regulations to those in surrounding states (Tennessee, Kentucky, North Carolina, etc.) and to forward-thinking states further afield. For example, if Wyoming imposes half the mandatory training hours for a particular profession (without any evidence of poor results), Virginia should consider whether its training requirement is too high.

At the federal level, the U.S. can compare its regulations to those in similarly situated countries and determine if an alternative approach is warranted.

Federal regulators can also look at state regulations to determine if a nationwide approach is truly needed. In many cases, states have adopted regulations at least as effective as those prevailing federally. In those areas, the federal government may choose to step aside to allow for a more tailored, state-by-state approach if it possesses the statutory authority to do so. If it does not possess that authority, it can propose a legislative fix to Congress. Federal pronouncements such as Executive Order 13,132 have nodded to the importance of federalism but never actually pushed more power down to states. AI can allow the federal government to comprehensively assess whether a more bottom-up approach may be preferable.

Ensuring Free Public Access to Regulatory Text

One of the less widely known aspects of federal and state regulation is the practice of “incorporation by reference.” Rather than spell out compliance requirements in the text of regulations, agencies will often refer regulated parties to privately developed standards, which are made legally binding when incorporated in a regulation.

These standards are often copyrighted and sit behind paywalls, meaning the public has to pay to figure out what the law is. And the standards are usually written by large industry players, giving them a competitive advantage over their smaller peers.

Despite the problems with incorporating privately developed code into public law, regulators often are stuck with this approach as a result of resource constraints. Writing code takes time, especially when it involves highly detailed requirements such as what building materials to use or how to maintain a highway. 

Virginia called upon the power of AI to help tackle this problem. Algorithms can quickly scan highly detailed standards and pull out the key provisions. They can identify unnecessary overlap and flag provisions that may be overly burdensome. And they can recommend regulatory text to be integrated into the official code. Human officials will, of course, need to decide what those regulations ultimately say. But AI can make an otherwise unmanageable task much more routine.

Federal regulators should consider a similar approach. Incorporated documents have generally evaded any scrutiny because no one really understands them. AI can empower regulators to ensure that they are meaningfully reviewing each line of regulatory code rather than handing the exercise off to large industry players.

The Future of AI in Regulation 

The ideas explored above merely scratch the surface of how AI will upend the regulatory process. Agencies are already using AI to automate routine exercises like summarizing public comments, and they will increasingly use it for more complex tasks like modeling various regulatory interventions prior to adopting a rule.

Nevertheless, the regulatory streamlining potential of AI is perhaps its most powerful use case because governments have largely failed to reassess existing regulations up until now. In a world of limited resources, agencies have always focused on the “next big thing”: once problem X has been solved, the natural tendency is to move to problems Y and Z, rather than periodically reconsidering whether X is still an issue. AI is making it possible for agencies to undertake a periodic reassessment of existing rules without detracting from other obligations.

Over the last four years, Virginia pioneered an approach to regulatory modernization that is becoming the model for other states. Going forward, the Commonwealth has the opportunity to use AI to sustain and build upon its successes. As the federal government begins to tackle the same problems at a national level, it should closely consider the Virginia precedent and draw on AI’s potential to unleash economic growth by streamlining bloated regulatory codes. 

Reeve T. Bull recently served as the Director of the Office of Regulatory Management in the Office of the Governor of Virginia.