LitigusAI Blog · Under the Hood
Under the hood: how LitigusAI keeps its promises.
The origin story explained why LitigusAI exists. This post explains how it works. Fair warning: I am going to show you the engine bay without handing you the blueprints. The specific machinery is ours; the principles behind it are yours to inspect.
The Litigus Library was built for AI first.
Start with research, where most legal AI quietly cheats. The common approach rents access from a data vendor, meters your searches, and passes the cost along. We built our own library instead: the Litigus Library, 10.8 million opinions from more than 2,100 federal, state, territorial, and specialty courts, dating back to the 1600s and updated daily. It lives on our hardware, it is never metered, and no vendor sits between your question and the answer.
The difference runs deeper than ownership. A human researcher runs one search at a time and reads one case at a time. LitigusAI has no such limit, and the library was designed for how machines actually research: dozens of parallel searches in a single pass, results matched both by the words in an opinion and by its meaning, then ranked by the weight of the authority behind them. The library is ours, so nothing meters those searches and nothing rate-limits them. Context compression keeps each agent's working memory lean, so even a sprawling matter never slows the work down.
Here is the part that makes me proudest. Most search tools hand back a stack of cases and wish you luck. Our library evaluates each case against the specific proposition you need: does this opinion support your argument, cut against it, or stand apart on its facts? When the law simply is not there, the library says so instead of stretching. An honest "no authority found" has saved more briefs than any clever search ever will.
The research also never leaves the conversation. Run a search or hand off a deep research assignment, and the results stay in context: you can discuss the cases with LitigusAI on the spot, and it already has full access to everything it just found. No exporting, no re-uploading, no explaining your own research back to the tool.
The second set of eyes.
A second, independent AI, built by a different lab and trained differently, examines the work and reports what it finds: the bad citation, the missed authority, the argument that won't hold. The lead agent decides what changes. You decide what gets filed. Two minds with different blind spots catch what one mind alone cannot. When you want one yourself, just ask: an on-demand call brings in a review agent for a second opinion on the lead agent's work. I did not invent peer review; I automated it.
Gates don't take AI models at their word.
Early in the build I learned the lesson that shaped the whole platform: instructions do not reliably govern AI models. Architecture does. An instruction says "verify every citation"; a gate refuses to let the work continue until every citation has been verified. LitigusAI runs on gates: deterministic checkpoints the work cannot pass without meeting the requirement, enforced by code rather than by trust.
Citations get the full treatment. Before a case appears in your papers, LitigusAI confirms the opinion exists, pulls the actual text, checks that it says what we cite it for, and confirms it is still good law. No summary of a summary, no confidence score standing in for proof: the actual opinion, read and checked.
Hallucinations are an engineering problem.
Every AI system can produce a confident wrong answer. Anyone who says otherwise in August 2026 is misleading you. The state of the art is light-years ahead of where it was even a year ago, and on a platform built the way LitigusAI is, the residual risk is small. The bigger protection is independence. Two models from different labs, checking each other, do not share hallucinations; the odds are about the same as winning the Powerball. What remains is the risk you already manage with any associate's work, and the rule does not change: review the work product carefully before it leaves your firm.
LitigusAI never forgets.
The first time LitigusAI goes to work in a matter, it does what a careful associate does on day one: it reads the file. The agents walk the matter's folder, take in the documents, and build a map of the case: the parties, the counsel, the court, the claims, the events and their dates, and the documents themselves, with every fact tied back to the document it came from; nothing in the map is a guess the agents cannot source.
The map lives inside the matter, in a workspace the agents keep on your machine: subfolders for notes, drafts, research, and reviews, maintained the way a careful associate keeps a case file. Open the matter tomorrow, or next month, and the agents are up to speed in seconds, as if they never left. They know the case, they know what they wrote last time, and they know why. Each matter's knowledge stays sealed inside that matter, by design: facts from one client's file never bleed into another's. All of it stays on your machine, in your custody. I am not aware of any AI platform, legal or otherwise, that does this.
LitigusAI does the other thing a good associate does: before it drafts anything, it comes back to you with a plan and waits for your approval.
LitigusAI watches its own health.
When a background worker stalls, maintenance agents restart it before you notice, and the watchdogs even supervise each other. The routine failures never reach you; the rare problem that needs a human gets flagged instead of hidden.
It improves itself, with a governor on the engine.
LitigusAI reviews its own performance, refines its own skills and instructions, and carries the lessons forward: recursive self-improvement, running on a limited and deliberate basis. A deterministic governor decides which changes are allowed, and every proposed improvement passes through the same gates as a brief headed for court. LitigusAI gets better every week; it never chooses the rules that govern its own improvement.
It learns to write like you.
Every lawyer has a voice, and a good associate learns it. A dedicated adaptive agent studies your preferences, your edits, and your style, and the drafts start sounding like yours instead of ours: your structure, your phrasing, your formats. The longer you work with LitigusAI, the more it writes like you on the first pass instead of the third.
Setup is a conversation.
An AI concierge sets up LitigusAI by talking with you about your practice, your folders, your courts, and your formats. One conversation replaces an afternoon of setup.
Pick the team for the job.
All of our legal workflow agents are powered by the latest state of the art models; however, we provide users with the option to choose between two differently configured teams of agents. Our Standard lane is great for most tasks and is optimized to balance speed, quality, and cost effectiveness. Our Ultra lane is optimized to produce the absolute highest quality of work product, but it can take much longer and cost significantly more to use. When the stakes are highest and money is not an issue, this is the go-to lane for scholarly, high-end legal work product.
The tools we handed the agents.
An associate with good judgment and no tools is still stuck. LitigusAI ships its own document engine, running locally, with no model involved: it converts between Word, PDF, and a dozen other formats, pulls text out of documents, merges and splits PDFs, and applies Bates numbers to a production. The agents call these the way you would use the software on your own desk, except they never mis-click and never lose the numbering halfway through.
They also work inside the applications you already use. Our plugins put LitigusAI in the task pane of Microsoft Word and in the sidebar of Google Docs, working on the document that is open in front of you: it reads the passage you select or the whole draft, rewrites it, fixes the formatting, fills in the fields, and checks the citations. In Word it can do all of that with track changes on, so what you get back is a redline you review line by line, the same way you would review an associate's edits. For final papers, LitigusAI drives Microsoft Word itself on your machine to produce the PDF, so the document that reaches the court is rendered by the same software your court and your adversaries use.
The agents can also see. Hand them a scanned exhibit and they read what a text extractor misses: the handwriting in the margin, the signature block, the notary stamp, whether the redaction is a redaction or a black box sitting on top of live text, whether the filed copy matches the copy you served. Sight is how a lawyer checks a document, so we gave the agents eyes.
Sight comes with hands. LitigusAI can open a browser and work the way you would: pull the docket, read the judge's individual rules, check the part clerk's calendar, confirm what was filed and when. On your own machine it can operate the applications that never got an API, moving through the software you already run rather than demanding you migrate away from it. The restraint matters as much as the capability. The agents use a direct connection when one exists, the plugin when there is a plugin, and the cursor only when nothing else will reach. They do not type your passwords, they do not answer a security challenge pretending to be you, and they do not submit a filing. Preparing the filing is their work; pressing submit is yours.
Where the work actually happens.
LitigusAI is a desktop application, and that is an architectural decision rather than a packaging one. For the lawyers reading: the work happens on your computer, under the controls your firm already maintains. For the IT people who will be asked to approve it, here is the stack in plain terms. The interface is a React and TypeScript application built with Vite, running in an Electron shell signed and notarized for macOS and Windows 11. Alongside it runs a local Python engine on the same machine: it reads your files, maintains the matter's case map, and queries a search index written in Rust. Nothing about your documents requires a network round trip to be read, indexed, or searched.
Our cloud does the orchestration a laptop should not: it coordinates the agents, brokers the model calls, and handles licensing and billing. What it never does is hold your files. The client documents stay on your device; only the specific material an assignment requires is passed for processing, under zero-data-retention terms, and nothing survives the request. Compare that with the standard design, where the vendor's servers are the system of record and your entire matter has to live there for the product to work at all.
Your practice, from your phone.
The companion apps for iPhone, iPad, and Android pair with your licensed desktop by scanning a code. From the courthouse hallway you can read the full transcript of a matter, watch an assignment run as it happens, and approve, revise, or hold a plan the moment LitigusAI asks for a decision. Ask for a second opinion from the same screen.
Notice what this design avoids. Your phone is not carrying client files, and neither is a vendor's cloud on your behalf: the desktop stays the execution authority, and the phone is a secure window into work that never leaves your machine. A pairing session can be revoked from the desktop at any time, so a lost phone is an inconvenience rather than a data breach. The traditional approach puts the whole matter on someone else's server so the phone can reach it, which means the price of mobile access is custody of the file. No litigator should have to make that trade.
Security is the architecture.
You read the principle in the origin story: your files stay on your machine, in your folders, in your custody, and every provider that touches client content operates under strict zero-data-retention terms. The full security deep dive deserves its own post, and it is next in this series. The short version fits in one sentence: a file that never leaves your machine cannot be taken from someone else's.
None of this machinery exists for its own sake. Every gate, every reviewer, every verification pass serves one promise: work you can put your name on, delivered in a fraction of the time it used to take, with the judgment calls left where they belong: with you.
The private beta is open for applications. If you want to test LitigusAI on your own matters at no cost, submit an application; seats are limited, and we review applications as they open.