TLDR: Legal content is perfectly structured to assist Legal Agents. Legal Agents became possible by GPT fixing many of the downstream issues. Now there are three more obstacles on the road to Autonomy.
Intro: This is our first-ever double header. The previous one was a bit dark, and we want to end the year on a bright note.
Before we throw any numbers at you, a bit of context. We covered a similar eruption of new apps twice before. In 2016, we recorded 44 new companies in a couple of months. We called it the Hottest Summer. During the crypto craze of 2018, we found 55 ‘legal tech’ companies raising a Billion in coin. Back in 2015, ProductHunt report, we uncovered a growth of 43 apps in roughly the same amount of days. In the early days of the Apple App Store, we noticed a similar growth. But nothing quite like this.
We were warned four months ago when we found 81 apps in 28 days. 550+ new solutions in one month set a new record. Now, will these new interfaces all be displacing current legal solutions and services? Not yet, and we’ll explain when this displacement will happen. These new apps will face similar struggles as the App Store, and ProductHunt platform posed. However, now there are different market conditions. In “Data, Distribution & Deals” we discussed market conditions that affect supply. This analysis will explore the demand side. Why do we need agents, and are they better?
In order to answer, here is a brief history of how we got here. We first discussed agents in 2011: Call My Agent. Basically, we wanted systems to learn from our content, derive instructions and tasks, and complete these tasks. They could perform their tasks autonomously in the background. This would allow us to access the results instantly at runtime.
In 2013, we created Loupe as a concept that would track regulations and instantly update affected contracts or briefs remotely. You would need a dashboard (DESH) to keep track of all edits deployed on each document. We even made an Apple Watch interface.
The same principle applies to legal research. We could dynamically convert phrases in documents to background queries. That was called Monocle and even made a teaser video. We submitted eleven patent proposals. Little did I know that each submission costs a fortune. Only one was patented, making me an inventor.
Why did we bring this up? Cooking these concepts in the lab, we discovered a few fundamentals of legal work. First, autonomous legal research is about anticipating legal issues, and calculating them beforehand. Second, the legal issues are already present in legal content. You just need to add the right context, as shown in the image above. In short: legal data already has the instructions and tasks baked in.
The biggest problem back then is actually what everyone now knows: calculating has immense costs. The OpenAI breakthrough is that no one has to pay that much. By creating GPT, OpenAI has democratized Legal Agents. Give GPT the right legal data, and anyone can create a Legal Agent.
What is a Legal Agent? Based on a given context, a legal agent can serve up a legal end-product like a contract, or research brief. If you ask GPT, it will provide a more technical answer. Something like “entities that can perceive their environment through sensors and act upon that environment using actuators. They’re designed to achieve specific goals or perform certain tasks.”. Not helpful, except it does provide the key ingredient for agents to work: context. Legal Agents get Legal context from instructions and legal data.
Before GenAI, every legal technology solution had to derive context from the user manually. As a Legal Knowledge Engineer, I had built legal decision trees and the process was very tedious. The tools were cumbersome and we could never encompass all cases. We just focused on the most complicated ones for commercial purposes. Now, with the right instructions, and a fined-tuned model can figure it out. Better yet, you can ask to explain the answer. This was the real breakthrough of TaxGPT. So are we there yet? Not yet. Let’s highlight one key difference we uncovered while covering GPTs: Copilots answer your questions, Agents take actions. So what is the next stage?
The launch of GTPs on November 6, 2023, made instructions more explicit for ChatGPT. However, OpenAI demoed instructions months earlier on March 14th, 2023. Remember TaxGPT? Yes, the first agent by OpenAI was targeted at Tax. Now there are around 90 TaxGPTs and counting. However, none can submit a tax form to your local IRS. While the current generations of legal agents can give you perfect answers, most can not take any action. Part of this constraint is security by design. Large Language Models will prevent any kind of code execution outside a sandbox. However, developers, that have APIs for actions ready in their apps, can now use GPT output as input. As the Copilot Battle states, this is the footrace, so hurry. Wait! There is one more thing.
Besides structured legal issues, we made one more discovery in our Ancestor era. A quick step back: The best definition of Artificial Intelligence I heard here at 4:59 min on the Waveform Podcast. The short version: AI is planning tasks in the right order to execute them. Sticking with the TaxGPT, here is a simple plan:
- Upload your bookkeeping*;
- Classify entries;
- Find the right deductions;
- Fill out the tax form and;
- Submit tax form to Tax Office.
*Note: Another Agent did my bookkeeping.
What tax and legal data provide is an understanding of how to structure tasks and actions to solve legal problems. Legal data has a complete recipe. Given this recipe and the ability to take action, Agents will be able to act autonomously. So the next stage of legal solutions is autonomy. And now you know why Sam was fired.
On the roadmap to autonomy, we have to travel through the Three Universes from “Will lawyers be replaced by GPT” which are:
- All legal data needs to be processed in GPTs. As stated in the previous post, it eventually will;
- Learn how to construct Agents to allow them to safely take actions, so allow autonomy;
- Interceptors do not squeeze developers, so allow fair competition;
This third milestone is, economically, the most difficult one because of Interceptors. Example: Google Bard now supports extensions and they are all Google apps. OpenAI was supposed to be open, but now it’s just Microsoft Copilot. Ideally, GPT becomes an open utility, just like GPS. Politically, it seems we are beyond that possibility. Potentially, agents can save the world, if we agree we want them to.
Meanwhile, we counted over 24,000 GPTs created since November 6th. That is 750 GTPs a day.