Private AI Applied: A Case Study in the Legal Industry
Secure AI for the legal industry.
Over the past two years, Frontier Foundry has been developing privacy preserving AI products that are now deployed across multiple industries and verticals. Since privacy preserving AI is so broadly applicable, we’ve found it helpful to showcase just how we configured our existing products to fit specific use cases. This communication has helped many of our prospective and current customers see how our solutions fit inside their organizations. And that’s really the rub with AI, isn’t it? There are products out there and it’s hard to see how they create value…if they create value. This is the story of one of our fine products, Limni, and how it came to live among attorneys in a large legal defense organization as an example of how AI projects should be done. Below you will find:
Leaders not allowing technology to dictate their vision
Rapid time from deployment to value
Privacy and security as core features
An AI system that is right sized for the requirement
Limni is our fully secure, privacy preserving large language model (LLM) that we are deploying with customers across industries. Limni is an agentic system with the capability to create custom, compartmentalized knowledge bases. It can run inside your secure environment, completely offline, or as an on-premise system to your preference. In short, it is secure, and your data stays with you.
For obvious reasons, this solution was very interesting to the legal community. For months, conversations have bubbled around the security of cloud-based eDiscovery platforms since data must be transmitted from the firm’s secure environment to the eDiscovery platform cloud. This creates a vulnerability, and the legal community is still understanding how to deal with client data in a multi-party cloud.
To wit (really leaning into the legal thing here), we found ourselves in a discussion with one of the largest legal defense organizations in the country and they had exactly this problem. They have huge volumes of client data coming in as multiple formats:
Video
Audio
Transcripts
Handwritten reports
PDFs
Photos
And much more
Limni was already out in the wild at this point, but the ability to use an LLM to parse and make quick sense of this mountain of data was critically important for the over 40 attorneys in their employ. The leadership of this organization had a clear vision, and they were not willing to be swayed by artificial barriers to entry. They found a tool that would work for them, evaluated the partners, and moved forward.
Wows in a Week
The next step in the journey was a small-scale pilot that we ran together. During this pilot, our client provided a small volume of PDF data that included handwriting. Using our proprietary process, we were able to upload the data, quickly train a legal specific model, and create some first-round outputs for their evaluation in under a week. We shared a screenshot of our results after just a couple of days and created significant excitement with the organizational leadership.
Once we understood that we were on the right track, we set to work over the next two months configuring our automated data ingestion, tuning our audio, video, and transcription capabilities, and preparing Limni for deployment on their system. As with all of our customers, running Limni inside of architecture they control is critical for us so that they do not have to make a choice between having AI without privacy or having privacy without AI.
An On-Premise Choice
Our partners ultimately chose to deploy the system on a server inside their office. We quickly configured the licenses to give all 40 attorneys access and watched as the first rounds of queries were run. We built in features to allow administrators and leadership to audit the system to ensure compliance with ethics, rules, and policies. These features allow users to view training data logs, export chats as PDFs, and monitor queries by user name. This level of security creates trust.
The Result
The system that resulted was complete by January 2025 after approximately 8 weeks of development time that included the holidays. The system is capable of reading handwriting, transcribing audio, and analyzing hours of video. Its simple user interface gives users access to a query field and the various knowledge bases the system has built in (each compartmentalized from the other).
The system now is live with over 50 daily users. Our partners deployed the system on an internal server, which never calls out to the internet keeping all client data inside their proprietary architecture. This feature inspires trust among their clients while providing them with a better end product, in this case their legal defense.
This system has been rewarding to work on for our entire team, but it also taught us something about how AI is built. When you build AI in pursuit of artificial general intelligence, you build big, and you are always in pursuit of more data. This creates a lack of specialization and the absence of privacy controls. However, most organizations do not want or need generalized AI that does not protect privacy. With many US states implementing new data privacy laws, the era of giving away your data for free is quickly coming to a close. For many organizations, this era never existed because they were never in position to give away data due to ethics, regulation, or other reasons. But if you build AI with the customer in mind, not an abstract goal to achieve AGI, you create something that is impactful and creates value in weeks, not years. You create something that can run on a laptop or internal server and does not require Three Mile Island to power.
In perfect candor, we did not set out to enter the legal market when we created the company. However, we found a home here because our tech is such an obvious fit. Working with our first legal partner showed us how smart technology and working within a vision wins over the enterprise SaaS sales model. We didn’t enter this project to sell as many software licenses as we could. We entered it to build something that mattered to its users and those that benefit from it. The result is a specialist system that knocks down the artificial barriers to entry that a lot of organizations see. Today, the system provides the value the organization needs to its staff attorneys. It reduces time spent on research and increases time spent with clients. It improves the quality of research and legal work product without requiring attorneys to spend hours watching video footage or listening to audio. It gives them the ability to generate reports and briefs and export them as Word documents for editing.
Ultimately, this use case is about caring about the mission. That’s the mission of the partner, not the mission of the vendor. If the goal was to sell as many licenses as possible, this would not be the result.
Connect with us: Substack, LinkedIn, Bluesky, X, Website
To learn more about the services we offer, please visit our product page.
This post was edited by Thomas Morin, Marketing Analyst at Frontier Foundry. View his LinkedIn here.