Regulatory Functions Most Amenable to AI-Driven Process Improvement

Summary of panel discussion at July 2025 conference co-hosted by RSC and Norm Ai
August 12, 2025

Download this summary (PDF)

Explore more event coverage


In brief ...

At a conference co-hosted by the Regulatory Studies Center and Norm Ai, speakers discussed challenges andd opportunities for artificial intelligence to streamline regulatory compliance burdens and improve stakeholder experiences with government services. This commentary summarizes the panel, "Regulatory Functions Most Amenable to AI-Driven Process Improvement" featuring Reeve Bull, Susan Dudley and Michael Mandel.

 

On July 8, 2025, the George Washington University Regulatory Studies Center and Norm Ai co-hosted an event titled Can AI Streamline Regulation and Reduce Compliance Burdens?The event convened academics, industry insiders, and current and former government officials to discuss how the government can leverage artificial intelligence (AI) technology to enhance its regulatory abilities, and how this novel technology should be regulated. The second panel of the day featured moderator Susan Dudley, Founder of the Regulatory Studies Center and former Administrator of the Office of Information and Regulatory Affairs (OIRA), and panelists Reeve Bull, Director of Virginia’s Office of Regulatory Management (ORM), and Michael Mandel, Vice President and Chief Economist of the Progressive Policy Institute (PPI). The panelists discussed how AI is currently being used to streamline regulations and to facilitate comparative analysis between states; opportunities for future AI-driven enhancements to the rulemaking process; and strategies for preserving human input and accountability.

Dudley kicked off the panel by asking Mandel and Bull to describe how their respective work has informed their perspectives on AI-assisted regulatory functions. Mandel summarized PPI’s 2013 proposal for an independent Congressionally-authorized Regulatory Improvement Commission (RIC). The RIC would conduct periodic evaluations on existing regulations submitted by the public for retrospective review, identifying both redundant rules for removal and outdated ones for improvement. The RIC would then submit a package of proposals to Congress for an up or down vote. The motivation behind this proposition, Mandel explained, was to develop Congress’s ability to deal with regulatory issues and to have it take responsibility for the outcomes of the laws that it passes. AI could theoretically aid this endeavor by identifying suitable rules that Congress could vote to rescind or modify. Mandel emphasized that he only envisions AI acting in a support role and that Congress must be responsible for making the final decision if the proposed system is to be legitimate.

Bull responded to Dudley’s initial prompt by giving an overview of his last three and a half years of work at Virginia’s ORM, a state-level oversight office modeled after OIRA that works with each of Virginia’s 66 agencies. It reviews the benefit–cost analyses that Virginian agencies conduct for all of their issued rules and guidance documents. During Bull’s tenure, ORM has achieved a 26.8% reduction in regulatory requirements and a 47.8% reduction in the length of guidance documents (cutting approximately 11.5 million words), saving Virginian taxpayers approximately $1.2 billion per year. It has also created a dashboard that enables Virginians to track the progress of their permit applications through the review pipelin

ORM has achieved these milestones without the use of AI, however, Bull announced, this is about to change. ORM has partnered with a small startup to create an AI tool that will facilitate additional streamlining beyond the current level. The AI tool uses natural language processing to identify discrepancies between regulations and their authorizing statutes. It also determines whether regulatory requirements are discretionary (rather than mandatory) or based on outdated statutes. All of these scenarios indicate opportunities for further trimming. The AI tool also enables Virginia’s regulatory requirements to be juxtaposed with those of adjacent states to determine whether certain industries or vocations are excessively regulated. This sort of comparative analysis is traditionally time-consuming; however, the adoption of AI has massively accelerated the data-gathering process and made it possible for ORM to quickly identify additional areas to target for cuts. Bull suggested that these AI-assisted deregulatory functions have potential applications both for other states and the federal government.

Dudley commended these technical accomplishments but pressed both panelists to elaborate on the role that human input should serve in a future where such AI tools are ubiquitous. Mandel expressed his skepticism about the idea that data-driven comparative analysis alone could determine whether more or less stringent regulation is desirable. In his opinion, normative standards cannot be established without human input. Bull responded that when conducting comparative analysis, the absence of either complaints or evidence of harm in less stringently regulated states provides a sufficient justification for the states imposing the more burdensome requirements to carefully review their regulations and determine if it is possible to reduce the burdens. Bull also agreed that it is essential that all final decisions are made by policymakers, with AI merely providing relevant research.

Mandel also expressed concern about proposals such as the recently introduced Leveraging Artificial Intelligence to Streamline the Code of Federal Regulations Act of 2025 that aim to amend or rescind ”out-of-date” and “redundant” regulations as identified by AI models while undercutting and bypassing the role of Congress in rule-making. More generally, he suggested that applications of AI tools need to be fundamentally aligned with the deliberative structure of government institutions. Figuring out how to align incentives and reform institutions without compromising their core purposes may ultimately prove more challenging than any of the impending technical hurdles.

Dudley next asked both panelists whether the AI revolution necessitated any reform to the style in which regulations are written, so as to maximize their readability to AI tools. Bull stressed the importance of ensuring that all regulations and guidance documents are available online in a machine-readable format as a necessary first step. Furthermore, research initiatives such as QuantGov have shown that AI tools, which are often trained on colloquial text, sometimes struggle to decode unusually formal text, such as legal documents. These tools also struggle to interpret imprecise modal verbs, such as “may.” Longer-term work will be needed to standardize regulatory language and reduce structural irregularities. Both panelists agreed that any significant technical or stylistic changes to regulations will have to be implemented incrementally.

Dudley concluded the panel by asking whether there are certain types of regulations that are more amenable to AI-driven analysis (e.g., performance standards over design standards). Mandel responded that AI-driven analysis is currently more applicable to digital industries than to physical industries. Furthermore, he suggested that this new type of analysis will likely expose a blind spot in the current system regarding regulatory outcomes. AI-driven analysis may reveal a stark discrepancy between the incredibly detailed prescriptions in the Code of Federal Regulations and the relatively sparse data on the real-world outcomes of those prescriptions. Closing this feedback loop by evaluating regulatory text against its effects in the physical world is essential to ensuring that AI-driven analysis actually improves outcomes.