Buried deep in San Francisco's municipal code is a provision requiring the Recreation and Park Department to report on a plan to restore an endangered frog habitat at Sharp Park, a picturesque wetland just off the Pacific Coast Highway. 

The department’s report deadline is August 31, 2009. Seventeen years later, the requirement remains on the books.

Elsewhere, the code instructs city departments to file biennial reports on “pedestal-mounted news rack zones.” Yet, San Francisco’s contract with newspaper vendors expired in 2024, and the racks that once held copies of the Examiner and Bay Area Reporter have long since been removed from the city’s sidewalks. 

The code runs nearly 16 million words, roughly 80 times the length of Moby Dick. 

These are just a few of the hundreds of provisions in San Francisco's Municipal Code mandating reports on programs and projects that no longer exist. Over the years, these outdated requirements have piled up. The code runs nearly 16 million words, roughly 80 times the length of Moby Dick. 

In 2024, the City Attorney's office partnered with Stanford University's Regulation, Evaluation, and Governance Lab (RegLab) to change that. 

Using an AI tool purpose-built for statutory research, they scanned the full municipal code and 16,000 pieces of legislation to build a comprehensive inventory of which reports are actually required by law. The proposed result was a 351-page ordinance to delete or consolidate 174 of those requirements — 36 percent of the city's reporting obligations that it has the power to change. 

It is the kind of administrative housekeeping that rarely makes headlines but could transform how government operates

It is the kind of administrative housekeeping that rarely makes headlines but could transform how government operates — freeing staff from compliance obligations that serve no one and making the rules that remain clearer, more coherent, and easier for the public and public servants to understand.

How It All Started

The project was kick-started when city departments came to the City Attorney's office with a request. The department wanted help identifying which of its reporting obligations were outdated, duplicative, or no longer serving a useful purpose.

The City Attorney's office (CAT), which was already working with Stanford's RegLab to identify projects where AI-powered statutory research could add value, saw an opportunity to expand that single department's request into something more ambitious: a systematic scan of every reporting requirement across the entire Municipal Code, for all departments.

This felt like a good initial project to work on, because it was clear, manageable, and something that we were hearing directly from our city clients. — Andrea Bruss, Director of Government Legal Reform at the City Attorney's Office.

"We had looked at some other projects that were just much more complicated, or time-consuming, or less clear," explains Andrea Bruss, Director of Government Legal Reform at the City Attorney's Office. "And so this felt like a good initial project to work on, because it was clear, scopable, manageable in size, and something that we were hearing directly from our city clients."

How It Worked: From AI Scan to Legislation

The project leveraged the Stanford RegLab’s Statutory Research Assistant (STARA) tool. 

STARA is a custom tool designed to assist with statutory research – the process of searching for and interpreting the laws that apply to a particular legal issue. The system uses a large language model to scan millions of words of legal text, identify provisions that match a defined set of criteria, and return structured results — analysis that would take a team of attorneys hours of painstaking work to do by hand.

The tool relied entirely on publicly available, published legal materials, including municipal codes and legislative texts. It was not connected to any internal City systems, and no confidential or nonpublic information was accessed or used.

The process unfolded in six steps:

1. Defining what to look for

The City Attorney's team first worked with the RegLab to develop a detailed description of what counts as a mandated report: any provision that directs a city department to produce a written report, study, or update for delivery to the Board of Supervisors (the city’s legislature), the Mayor's Office, or another entity on a defined schedule. 

The AI tool then translated that description into a rubric: a set of criteria that the model could apply consistently across thousands of provisions.

2. Ingesting and chunking the legal code

The team then loaded San Francisco's entire 16-million-word Municipal Code and nearly 16,000 Board of Supervisors resolutions into the STARA system. 

The AI model divided the legal text into smaller units, including individual provisions and subsections.

Each unit included a section of the text along with surrounding context, such as definitions and cross-references that appeared in entirely different parts of the code. Without this context, the legal text would make little sense. 

3. Identifying reports

Working unit by unit, the AI model read each provision and compared it to the rubric to determine whether the text required a department to produce a report. 

STARA identified 528 reporting requirements across city departments, with most concentrated in the Controller’s Office, the Office of the City Administrator, the Planning Department, and the Mayor’s Office of Housing and Community Development. 

The AI model also extracted key details for each requirement, including whether the report was recurring or non-recurring, how frequently it was due, which department was responsible, and who was supposed to receive it. 

STARA returned the results of this research in a searchable database, complete with citations. 

4. Validating the results

 The City Attorney's office reviewed every single output by hand, checking every citation and detail to confirm accuracy. 

The City Attorney's office reviewed every single output by hand, checking every citation and detail to confirm accuracy. 

Staff then analyzed the list to determine whether the city had the legal authority to change each reporting requirement — making sure, for instance, that changes were not restricted by the City Charter, a voter-approved ballot measure, or state law.

5. Consulting with departments

The City Attorney’s Office shared the validated inventory with client departments and worked with them to determine what legislative action to recommend. Each decision about whether to modify, consolidate, or retain a reporting requirement was ultimately made by the respective client department. 

As the subject matter experts and policy owners, departments were responsible for determining which requirements continued to serve a meaningful public purpose and which no longer did.

Departments flagged some requirements as obsolete. Others were duplicative, where the same information was already captured through a different process. Some were important but could be consolidated with related reports. And some they wanted to keep exactly as they were.

The initiative began in spring 2025. By June, the City Attorney had developed draft legislation presenting the departments’ recommendations for the Board of Supervisors' consideration. 

Results and Outcomes

Based on the statutory review and consultation with departments, the City Attorney's office drafted a 351-page omnibus ordinance that proposes deleting, modifying, or consolidating 174 reporting requirements — 36 percent of all alterable reporting obligations in the municipal code. 

“The ordinance proposes removing 140 reporting requirements, consolidating 34 requirements, and keeping 314 requirements unchanged.”

The ordinance proposes removing 140 reporting requirements (29 percent of all changeable requirements), consolidating 34 requirements (7 percent), and keeping 314 requirements (64 percent) unchanged.

Some proposed amendments would eliminate reports for defunct programs or agencies. For example, the ordinance would strike an 80-year-old requirement for quarterly reports from the Redevelopment Agency, an entity that has since been restructured and was dissolved in 2012. 

Others seek to consolidate duplicative requirements where the same information is being captured in multiple places. For instance, the Planning Department is required to produce quarterly housing production reports, an annual housing inventory, and a biannual housing balance report — all of which largely cover the same data on how many units are being built and at what income levels. The ordinance would fold these into a single consolidated report.

Still others take aim at outdated provisions with expired deadlines, such as the Park Code’s endangered-frog habitat report from 2008. 

City Attorney David Chiu introduced the ordinance to the Board of Supervisors in June of 2025. As of May 2026, the ordinance is awaiting action from a committee. 

Lessons Learned

San Francisco's experience offers several lessons for other governments considering AI-assisted regulatory reform. 

1. AI can dramatically accelerate statutory research – but deliberative work remains slow.

San Francisco’s work shows how AI can enable government teams to conduct intensive statutory research more quickly and at lower cost. 

For a related project, CAT used STARA to help the city's Commission Streamlining Task Force develop a comprehensive list of all boards, commissions, and committees established by law. By analyzing the full Municipal Code and Board of Supervisors resolutions, STARA identified all 111 previously documented bodies and discovered 33 more. 

Research that would have taken human experts hours and cost an estimated $3,000 in labor was completed in 20 minutes. The compute cost was just 86 cents. 

Research that would have taken human experts hours and cost an estimated $3,000 in labor was completed in 20 minutes. The compute cost was just 86 cents. 

The reporting requirements scan was similarly rapid, with AI helping to organize hundreds of scattered obligations into a structured inventory sorted by department, frequency, and recipient — giving staff a clear picture of the full landscape for the first time. 

But the months that followed — which involved manually verifying every output, consulting with departments, and drafting legislation — took just as long as they would have without AI. 

It did not replace the need for us to individually evaluate the legal issues associated with every single one of these outputs. — Andrea Bruss

"It did not replace the need for us to individually evaluate the legal issues associated with every single one of these outputs," Bruss said. “From our perspective, it wouldn't have been appropriate for [AI] to play that role.” 

2. AI could do more — if governments are willing to expand its role

The City Attorney's office used AI only for initial research, not to help classify which requirements were outdated, to analyze whether the city had the legal authority to change them, or to draft the legislation. That work was done entirely by hand. 

But there are ways AI could help at this stage too. Virginia’s Office of Regulatory Management, for instance, used AI agents to classify all provisions in its state code as either mandatory or discretionary and to flag duplicated requirements across multiple regulations. 

One could imagine San Francisco or other cities using a similar approach — deploying AI to flag requirements with long-passed deadlines, or to sort provisions by whether they originate in the city charter, a voter-approved ballot measure, or an ordinance the Board of Supervisors can amend on its own. AI should never replace legal research or analysis, but it can give staff a head start and make their analysis more efficient. 

3. Transparency and guardrails around AI's role help to build trust

Any effort to use AI for regulatory decluttering raises several risks. The technology could be misused, intentionally or not, to strip away requirements that serve a real public purpose. And, as with any use of AI, it is necessary to ensure that the technology is accurate, reliable, and does not perpetuate bias. 

In San Francisco, some were skeptical about using AI for statutory research. “Some policymakers, I think, have natural skepticism of how a government entity might use a tool like this,” Bruss says. “The use of [AI] certainly raised questions for some people about how, or why, or specifically what safeguards were in place.”

The City Attorney's office mitigated this risk in two ways.

First, AI was used only to identify which reporting requirements existed — not to determine which should be cut. We developed the recommendations on what to keep, modify, or eliminate with the departments that owned the requirements, through a case-by-case consultation process grounded in their institutional knowledge of which requirements deliver public value. 

Second, the team was transparent about the tool's boundaries: every output was individually verified by attorneys.

Defining clear boundaries for AI use, clearly communicating those guardrails, and keeping humans visibly in the loop can help to ensure that the technology serves as a tool for better-informed decision-making rather than a shortcut that erodes the careful judgment these changes require.

4. Custom AI systems outperform general-purpose tools for statutory research

A general-purpose AI model would struggle to complete the research on reporting requirements due to the length and complexity of legal codes. 

Legal codes are structured differently from ordinary reports or documents. Provisions are nested, cross-referenced, and defined by terms that may appear in entirely different sections. An off-the-shelf large language model prompted to find reporting requirements would miss obligations implied by context rather than stated in a keyword. 

Indeed, in benchmark tests conducted by Stanford’s RegLab, STARA outperformed both human researchers and comparable off-the-shelf AI research tools in accuracy, efficiency, and comprehensiveness in statutory research tasks.

Because STARA was built specifically to handle this complexity, it surfaced hundreds of reporting obligations scattered across San Francisco's 16-million-word Municipal Code. 

5. Opportunities for greater community engagement with AI remain

In San Francisco, community engagement around the reporting streamlining initiative happened through existing public participation processes. The ordinance was heard by the Planning Commission, where community members could provide feedback on proposed changes to planning-related reporting requirements. 

The legislation is now pending before the Board of Supervisors, where it is subject to public review and approval. Public hearings are important and necessary. But in cities across the country, there are numerous well-documented barriers to the public meaningfully participating in legislative processes. 

Hearings typically happen during business hours at City Hall, favoring organized interests over everyday people. Consultations often overrepresent the views of affluent, well-educated, connected residents who are part of the demographic majority. Decision-makers often solicit public input only after the key decisions have already been made. The materials under review – in this case, a 351-page omnibus bill – are often written in dense legal language that most people cannot reasonably be expected to parse. 

The result is that meaningful participation tends to concentrate among those who already know how to navigate the system, while the communities whose services depend on what the government reports and tracks may never learn that a change was proposed until after it takes effect.

Well-designed engagement processes can help to lower some of these barriers. Online platforms can enable residents to review and comment on proposed legislative changes – or even to participate in the collaborative drafting of legislation – on their own time, from their own devices, without needing to show up at City Hall. 

AI can go further, helping translate dense legislative language into plain-language summaries, surface the specific changes most relevant to a given neighborhood or community, and synthesize large volumes of public input into more usable insights for decision-makers. 

Used well, these tools could make the consultation process not just more accessible, but more substantive — engaging communities earlier in the process and giving them a real voice in shaping the outcome.

Conclusion

While the initiative's ultimate impact remains to be seen, San Francisco's effort to streamline its reporting requirements is a promising example of what is possible when the government pairs the right technology with the right institutional process.

Combining a well-defined problem, the right tool for the job, and an analysis process designed to produce legislative outcomes can serve as a replicable model for other governments. 

 AI did not decide which rules to keep or cut. It made the rules visible so that humans could work more efficiently. Combining a well-defined problem, the right tool for the job, and an analysis process designed to produce legislative outcomes can serve as a replicable model for other governments. 

For cities struggling under the weight of decades of accumulated code, AI can provide public servants with the tools needed to do what no one has had the time or resources to do before: read it all and start cleaning it up.