Building AI Inside the Courthouse
Guest co-author - Chief Judge Susan Chehardy, Louisiana Fifth Circuit Court of Appeal
Nothing gets us more fired up than case management. We know that may sound strange coming from two judges, but we are systems people. We love figuring out how to make courts more efficient, more effective, and more accessible to the people we serve. Give us a complicated workflow, a procedure that takes too many steps, or a problem everyone has learned to work around, and we want to figure out how to improve the process.
That is why we are excited about what we are building at the Louisiana Fifth Circuit Court of Appeal. Our court recently purchased two artificial intelligence servers and the equipment necessary to operate them, including additional power and improved air conditioning for our server room. We have partnered with the ALEA Institute(judicial disclaimer: this is not an endorsement) to develop applications using an open-weight model running on that equipment. It is an ambitious undertaking for a court, and we are just getting started.
Our court already uses some commercial GenAI products. But this project gives us another path, one that allows us to build tools around the particular needs of our institution, with the added benefit of learning what operating inside of our own AI infrastructure actually entails. We intend to begin with filing compliance and intake, bookmarking of pleadings and records, then explore assistance with classification, routing, and eventually the preparation of neutral bench memoranda. Before we could achieve any of this though, we had to begin by informing the technologists on how the analog work of our court actually gets done.
Learning From the People Who Do the Work
When a document arrives at the court, it moves through a series of decisions that may seem routine from the outside. Written rules explain some of them, but they do not capture everything an experienced employee knows. A clerk who has handled hundreds of similar filings may recognize a missing exhibit or an unusual procedural problem almost immediately. Translating that experience into a useful automated application requires more than handing a developer a copy of the court rules.
Over the past several months, employees throughout our court have participated in interviews and completed questionnaires we developed together about their workflow responsibilities. We asked them to walk through their workday, explain the documents they process, the decisions they make, and describe the problems they encounter repeatedly. Their answers have helped us see where technology might remove unnecessary steps and where the existing process depends on human experience that should remain firmly in place.
The ALEA Institute is leading the coding and development. We did not have all of the technical expertise to undertake a project like this ourselves, but our dedicated IT professionals have been working with ALEA for months and are learning quickly. It has been a tremendous learning experience for everyone, including the employees who are discovering how integral their everyday knowledge is in the design of a novel system. Instead of buying a ready-made product and rearranging the workflow around it, we are starting with the people who know the work and building outward.
Following a Filing Through the Court
Consider a lawyer preparing to file a writ application. The argument may be sound, but the document might exceed a page limit, contain incorrect pagination, or refer to an exhibit that was never attached. Under the usual process, an employee may discover the problem only after submission, when correcting it requires additional communication and processing. We would like to catch these avoidable problems before they enter the system.
Our planned tools would review a proposed filing against applicable rules and alert the filer to potential deficiencies. A lawyer or self-represented litigant could learn that the document appears too long, that the pagination needs attention, or that an attachment may be missing, and have an opportunity to fix it before submitting. That is especially promising for people navigating court procedures without a lawyer. The program would provide practical feedback about filing requirements, not legal advice or a determination that a pleading should be accepted or rejected.
After submission, the same system could help employees understand what has arrived. It might suggest a document classification, identify appropriate workflow, or flag something that needs closer attention. Employees would verify those suggestions, conduct the official filing review, and follow established procedures. We are not asking AI to decide where a case goes. We are trying to give the people responsible for moving it through the intake process better information at the moment they need it.
We also hope to add checks that address problems courts are beginning to see in AI-assisted legal work. By connecting to a trusted legal citation verification service, the system could flag nonexistent cases, inaccurate quotations, or other questionable authorities in pleadings. We would also like to screen incoming documents specifically for hidden instructions intended to manipulate an AI application reading the filing (prompt injection attacks). Those capabilities remain future development goals and would never be put to use to replace a lawyer’s responsibility for a filing or an employee’s independent review. These would be additional opportunities to identify problems before they move further through the courthouse.
This work is not intended to replace court employees. Like many courts, we face constrained budgets and cannot simply hire more people whenever the workload grows. We have dedicated public servants performing critical work, and we want to increase their capacity by providing them with tools that can improve the speed and efficiency of the jobs they are currently performing. If technology can help with repetitive checking, sorting, and organizing, employees can devote more time to the exceptions, complicated problems, and human interactions that benefit from their professional experience. The goal is not fewer people doing the same work. It is giving the people we have more resources to do the work that matters.
Why Build It Here?
We did not purchase AI servers because we believe commercial, cloud-based GenAI tools are inherently insecure or necessarily beyond a court’s control. Many commercial, cloud-based services can be configured with strong security and governance protections, and we already use some of them. We simply do not know where this technology is ultimately headed. Model capabilities are changing quickly, and so are the economics. Token prices, subscriptions, and the costs of operating AI will likely look quite different a few years from now. Rather than bet entirely on one approach, we decided to pursue both and learn from the experience.
Running open-weight models on our own equipment gives us another set of options. We can experiment with models suited to particular jobs, understand what they cost to operate, and decide where commercial services remain the better fit. A task such as checking page limits may not require the largest or most powerful model available, just a tool that performs that particular job reliably, under conditions we can test and understand.
Operating our own infrastructure also brings unique responsibilities. Servers require power, cooling, maintenance, cybersecurity protections, and people who know how to keep everything running. Being on premises does not automatically make a system immune to security issues. That is one reason the experience our IT professionals are gaining alongside ALEA matters so much. We are building institutional knowledge, as well as a new GenAI tool.
For now, we are keeping the AI environment separate from our existing case management system. We depend on that system every day, and an experimental application should not be modifying official records or disrupting established procedures. Our initial tools will produce information for employees to evaluate. Any closer integration in the future would follow rigorous testing and a careful determination of what safeguards are needed.
Beyond the Clerk’s Office
If the first phase proves useful, we hope to apply what we learn to research staff work. Appellate staff attorneys spend substantial time reviewing filings and records, identifying issues, researching the law, and preparing bench memoranda. We envision an application that could help organize those materials and prepare a preliminary, balanced memorandum with hyperlinked cases, grounded in the record, briefing and legal authorities, that fairly presents the competing arguments and identifies questions requiring closer attention.
That is a different challenge from spotting a missing exhibit, and we would not treat success at one as proof of readiness for the other. No matter how sophisticated any future application might be, research attorneys would continue to verify the sources, review the analysis, and remain responsible for their work product. The system would not recommend how a judge should rule. Its purpose would be to help our attorneys get better oriented in a record and begin the difficult work of analysis, not hand a judge a conclusion dressed up as a finished product.
Learning by Building
We are still at the beginning. The servers have been purchased, supporting infrastructure work has begun, employees have described their workflows, and ALEA has begun development. Some ideas will prove harder to implement than we expect, and some may turn out to be better handled by traditional software. We want to discover those things through testing, rather than mistake an impressive demonstration for a dependable court application.
For years, courts have often adapted their work to software designed elsewhere. We see an opportunity to vastly improve on that relationship, beginning with the problems an employee encounters, developing a tool around them, and improving that tool with the people who will actually use it. Commercial products will continue to be utilized as part of our workflow, and we are certainly not suggesting that every court needs its own AI servers. But there is real value in courts learning how these systems work inside and out, and taking an active role in shaping tools designed specifically for this important work that only the judicial branch can perform.
What excites us is not simply having two new servers in the courthouse. It is watching our employees, IT professionals, and development partners learn together how to make the courthouse work better to serve the ends of justice in the most efficient and effective way we can. If we succeed, the result will not be an artificial judge. It will be a court whose judges and staff members have better tools, more capacity, higher productivity and, as a result, more time to serve the lawyers, litigants, and public who depend on them.
