Who Governs Contracts Between AI Agents?

As AI agents begin negotiating, purchasing, and forming agreements on behalf of people and businesses, the legal infrastructure supporting commerce may need to evolve. In this episode, Jen Leonard and Bridget McCormack explore Legal Context Protocol, or LCP, a new open protocol designed to make legal terms discoverable, verifiable, and connected to agentic transactions.

The conversation begins with practical AI use cases, including Claude’s ability to send emails from inside the app and scheduled AI news roundups. Jen and Bridget then turn to Mythos, a powerful AI model reportedly capable of identifying software vulnerabilities at a level that raises serious cybersecurity and governance concerns. The main conversation focuses on LCP, why agent-to-agent commerce creates new legal risks, and why lawyers may have an essential role in designing the next generation of commercial infrastructure. 

Intro + AI Aha! 

Jen Leonard: Hi everyone, and welcome to AI and the Future of Law, the podcast where we follow all of the ever-changing dynamics involved in artificial intelligence and talk about what those developments mean for the legal profession. 

I’m your co-host, Jen Leonard, founder of Creative Lawyers, here with the fantabulous Bridget McCormack, president and CEO of the American Arbitration Association. 

For those who may be new to listening, when we have our one-on-one episodes where it’s just the two of us, we have three different segments. We start with our AI Aha!—something we’ve been using AI for that we find particularly interesting. Then we offer a little bit of a summary on things you might have heard about in the news and headlines in our What Just Happened? segment. And then we dive into our main topics that are relevant for the legal profession. 

I’m so excited today, Bridget, because this topic is new and fresh and cool—very much like the person creating it. 

We’re going to talk a little bit about Legal Context Protocol, or LCP for short. So we’ll learn all about that and what you and your team are doing around that. 

But first, do you want to kick us off with your AI Aha!? 

Bridget McCormack: I do. And like we sometimes do, I’m not sure mine is very sophisticated. 

But I noticed a little bit of a change in my behavior with AI that I thought was a good AI Aha! 

Steve and I were hiking our woods in northern Michigan, and I was using it like I always do to identify plants that I don’t know. 

The forest is full of ramps this time of year. Do you have ramps in Philadelphia? 

Jen Leonard: I don’t know what ramps are. 

Bridget McCormack: They’re like a little scallion-onion thing when you pull them up, but they just look like green leaves. And they’re on every menu right now. Basically, if you go to any restaurant, there will be some ramp dish. 

So while I was in the Claude app, I had it telling me what you would make with so many ramps—like probably miles and miles of ramps. I won’t make miles of recipes, but it gave me a bunch of ideas, like ramp pesto and ramp salt. 

Then I said, “Can you just type up the recipes and ingredients for the two or three that I liked and send them to me in an email?” 

Have you been using Claude to send emails while you’re in the app? 

Jen Leonard: No, I have not done this. How do you do this? 

Bridget McCormack: I just figured I’d try, because I knew I could look back at the chat. But I thought, “Well, I’ll just try and see if I can send an email to myself.” 

I asked Claude to do that, and sure enough, Claude just asked me for the right permissions on my phone and sent me an email with the information I asked for. 

Then, when I was still in the app, I remembered something I had to tell my sister, and I had Claude send my sister an email. 

Still just in the Claude app, in the woods, identifying things and sending emails. 

Bridget McCormack: So all in one app. I am identifying ramps. And the most exciting thing we identified—well, exciting to Steve, less exciting to me—was bear scat. We have bear scat in the woods. 

And lots of information about what bear scat looks like in early spring. So I learned so much about early spring bear scat and sent emails while learning it. 

So that’s my AI Aha! Doing all the things in one app instead of switching between apps. 

Jen Leonard: I can’t wait to play with this and try it out. 

I have two AI Ahas, and they’re not earth-shattering, but they are new capabilities I’ve been playing around with. 

One is that somebody at a workshop I recently did recommended Manus AI. Have you used Manus at all? 

Bridget McCormack: No. 

Jen Leonard: The interface is very similar to any of the general-purpose platforms, but it’s supposed to be much better at creating slide decks. 

I use Notebook LM for slide decks frequently, and I hear there are editable features now, so I need to go back and check that. But for a while, when you did decks in NotebookLM, the illustrations were amazing and everything looked great, but you couldn’t really change it around. 

I actually started creating this PowerPoint on Claude, but it never gets the colors right. I’ll upload the last version and say, “Replicate the colors and the format and the template,” and it’s just always off. 

So I decided to try Manus because this person had said it was really good with decks. I fed it into Manus and said, “Can you create this in the style of the last one?” And it generated it, but the color was still off. 

So I went into Canva and copied the exact color that we used in the last deck—all the numbers—and took a screenshot and uploaded it to Manus. I said, “Here’s the actual color that we used in the last one. Can you replicate this?” 

And it did. 

It’s such a small thing, but— 

Bridget McCormack: First of all, it’s amazing. And I’m glad to know it because I’ve been using Claude a lot lately, which has gotten a lot better. 

But I agree with you on the things it can’t quite do. That color thing—I think you discovered a jagged edge. It wasn’t able to identify the color just from you uploading a slide in that color. But when you give it the number that a machine can read, it can make sense of it. 

It can’t quite yet read the color it’s looking at. I think you just discovered a jagged edge. 

Jen Leonard: Claude could not do it. I copied and pasted the number into Claude, and it was giving me this teal color. I really needed a lime green color. 

Manus was able to do it. It’s like a Claude jagged edge. I don’t know. It’s very weird to me. 

Then the second one is sort of in the same vein, although not as advanced as what you were talking about with sending an email from Claude to you. 

It’s really just scheduled tasks. I’ve been doing a weekly roundup of news related to AI and the legal profession and AI generally. On Monday mornings, when I log in, it’s available to me and I can review it. 

Are you using scheduled tasks? 

Bridget McCormack: I have been using them for a very long time. 

I have a number of them set, and I’ve spent a lot of time lately thinking that I really need to revisit them because I set them a long time ago, and now I have things I care about more. 

You get an email update about what it found, and I’m like, “Yeah, I don’t really care about that. That’s not what I need to know. I need to know this other thing.” 

So I really need to go update my scheduled tasks. I went a little scheduled-task crazy when I realized it first had that functionality.

What Just Happened? — Mythos and Cybersecurity Risk 

Bridget McCormack: I feel like our What Just Happened? segments—I don’t know, should we do them every day? 

I don’t even know how to define what just happened since our last What Just Happened? because so many things just happened. 

Codex caught up to Claude Code, allegedly. I don’t know. That’s what I hear. I think we’re getting new models from Gemini soon. 

But I think the thing we’ve probably heard the most about that’s worth addressing is Mythos. What is Mythos, Jen? And what’s happening with it? 

Jen Leonard: For those who are listening and have busy practice lives, judicial lives, professor lives, you may have seen headlines floating around about Mythos and some of the interface between the government and Anthropic around Mythos. 

The underlying issue is that Anthropic, the creator of Claude, was preparing a new model for release called Mythos. And during testing, they realized that Mythos is really, really effective at finding software vulnerabilities—latent software vulnerabilities that some of the most advanced cybersecurity experts would never find, sometimes in 30-year-old systems where the vulnerabilities had never been disclosed before. 

Anthropic decided it was too dangerous for public release. Instead, it decided to release it to a limited group of technology and cybersecurity companies and Wall Street banks to start testing it and trying to develop a plan for protecting all sorts of systems we use: financial systems, personal systems, and the operating systems underlying many of the software programs we use globally. 

The challenge, of course, is that Anthropic here was acting conscientiously—although Anthropic itself also had two breaches of the Mythos code. One was found by Forbes, I believe, so it has its own vulnerabilities. 

But the other issue that has been keeping me up at night—although I saw Ethan Mollick say that anytime somebody says something is keeping them up at night, he’s sure they’re not actually up at night—is that all the other models will catch up to this quickly and have the same capabilities. 

I think Dario Amodei suggested it would be between six and 18 months from now that one of these other labs will have this capacity. And our systems all over the place are not prepared for what happens when a bad actor is able to exploit thousands of vulnerabilities that we didn’t even know about and weren’t prepared for. 

So we take an optimistic lens around AI, but this is the story this year that gave me the most pause and created the most angst for me. 

But I tend to be a worrywart. So what were you thinking around Mythos, Bridget? Did I describe it accurately? And what am I missing? 

Bridget McCormack: Yeah. No, I think you described it exactly accurately. 

In fact, you and I had dinner in New Orleans a few weeks ago, the night before a presentation we were doing. Up until that point, I was feeling kind of optimistic that there seemed to be this collaborative approach to figuring out what the vulnerabilities were going to be and pausing the release until it could be patched. 

That seemed like such a great sign of how we were approaching the most powerful model yet. 

And then you pulled me over to your camp a little bit when I said, “Well, what exactly could go wrong?” And you were like, “Well, what if all of the air traffic control computers are interfered with by some Mythos-like model from another lab that hasn’t taken care to workshop it in-house before it’s released?” 

So I think I became a little bit more alarmed after seeing you. Thanks for that. 

But it definitely seems clear that this is not just the regular kind of step increase in capabilities. 

You heard from the White House that they’re considering—although I think they backtracked on this a little bit—putting in some serious regulatory layers in the industry before allowing new models to be released. And that has to be a result of what everybody has seen with Mythos. 

I also saw this week that leaders of a number of the big tech companies are accompanying the president to China for his meetings. Maybe that’s a sign that everybody understands this is a moment for cooperation and not the alternative. 

But it’s interesting. I don’t think I’ve heard anything recently about when we will all actually get to use Mythos. Have you followed that at all? I don’t know where we are in the timeline of testing. 

Jen Leonard: No, I don’t know with Mythos in particular. 

It just seems like the speed at which these labs release new models without fully understanding what their capabilities are suggests to me that there will probably be a model in the next 12 months from somebody that people will be able to use for this purpose. 

And like you said, I always think it’s easy to try to sound smart by pointing out all the negatives about something. But I am really good at catastrophizing. 

So I was thinking about all the different systems. And it’s sort of like everything else we’ve talked about: because you have this highly democratized technology in the hands of millions of people, things are just very unpredictable. 

Trying to coordinate all of the underlying systems, industries, and security risks in a volatile global environment just feels troublesome to me. 

It’s interesting—I hadn’t connected the dots around the visit to Xi Jinping, but that makes a lot of sense. Some of the commentators we’ve been following for a long time have suggested that China and the U.S. really do need to collaborate around this for global security. 

So it will be interesting to see or hear what happens. I’m sure that in these presentations at the White House, one of the reasons that came out is because they have walked through some of the catastrophic risks associated with this enough to sober people to it. 

So that is Mythos at a very high level. 

But I’m super, super excited, Bridget, for our main topic today. And this is less of a conversation between two people who are very knowledgeable and more me learning from you today about what you’ve been working on. 

I love getting dispatches from you from around the globe—like, “I’m hunkered down working on this new build”—and then we get to hear about it. 

I’m always so curious. You’ve been working a lot this year on building what you’re calling a Legal Context Protocol. You and David Fisher have launched this together. 

First, who is David Fisher? And what is LCP? What have you both been cooking up over there? 

What Is Legal Context Protocol? 

Bridget McCormack: What are we cooking up in our little AI lab? 

First of all, David Fisher is the CEO of Integra Ledger. Integra Ledger is a blockchain company—but a blockchain company that has never issued a currency. It has focused on the legal industry. 

David is not a lawyer, but it occurred to him that legal agreements and commercial agreements need the kind of security and protection that financial agreements can get with blockchain. So he’s been working with the legal industry for a long time. I think his business is about ten years old. 

He started working with AAA before I got there in our election practice. He helped us with a blockchain solution to our online election tools. Then we worked together again recently. We offer Integra services after arbitration to users who want to hash an agreement—like the agreement that results from an arbitration—or any other documents. We offer that service. 

I always hear from people who are worried about deepfake videos and deepfake audio. And I think you’ve heard me say: sure, those are bad, especially given how good the technology is at making them believable. 

But I’m a little bit more worried about deepfake documents, because documents are relevant in just about every case. There aren’t that many disputes that turn on a video or an audio file, but almost all of them turn on some document. 

AI has made deepfake documents unbelievably effective and simple. So it may be time for lawyers to figure out how to hash documents so that if there is a debate later about whether a particular one is the correct one, they have some single point of truth. 

So I was starting to focus on agentic commerce. I’ve been following how quickly the payment firms, the frontier model companies, and the online retail companies were moving into a new market where agents are negotiating with agents—often in pretty simple transactions, but increasingly in more complex transactions. 

Agents are forming contracts with agents. 

And I started thinking, as I am one to do, about what new disputes are likely to occur as a result of a new set of financial relationships between agents. 

They may make fewer mistakes than humans. They may make more mistakes than humans. They may make less impactful mistakes than humans or more impactful mistakes than humans. It’s hard to know. 

But there will certainly be disputes. Even in a world where agents are negotiating agreements with other agents, things will go wrong. 

As somebody who provides dispute resolution services with my team, I was trying to think about what kind of expertise we will need. Do we need to recruit new people to our panel? Are there going to be new evidentiary issues we should be thinking about and maybe training our neutrals to be prepared for? What education do our neutrals need? 

So I started way down in the future of: what goes wrong in this new market? 

Even as the market was gaining steam and gaining steam, and the regulatory signals from D.C. were kind of all gas, no brakes, I wanted to make sure we were prepared for disputes. 

That got me peeling back the onion. 

It was actually at my governance panel at Legalweek, where I was talking with the other panelists about how complicated governance gets when agents are now doing the negotiating, and agents are doing more of the work than humans are. 

And I realized there’s a more fundamental problem. 

It’s not only that we will need new dispute resolution expertise. We will. And we’re working on that. We’re actually way ahead on that now. 

We may even need new dispute resolution options, which is exciting for us because we like to build new dispute resolution options, and AI gives us the tools to do that. 

But I started to get concerned that there was a more fundamental problem: all of the legal infrastructure that has supported e-commerce over the last 20 years kind of grew up around e-commerce—click-through terms, terms of service before you hit the next page—and none of that necessarily translates, and maybe not at all, to when agents are negotiating contracts with agents. 

So it seemed there was a more fundamental problem. There needed to be some understanding of how legal context attaches to agent-to-agent transactions. 

There have been all these other protocols in the e-commerce space, or in the agent space more generally: the MCP, the ACP, the UCP. They govern payments, how models find one another, connectivity, identity. Visa has a protocol around identity. 

These protocols follow the way internet protocols have worked over the last 40 years or so. They’re open source. Nobody owns them. There’s no money to be made here. This is not a commercial venture. 

But they are a way for the stakeholders involved in the market to agree to a universal way to do something—whether it’s payments, connecting with other services, or in this case, attaching the legal terms that govern a transaction. 

David Fisher was also thinking about this because, as someone who had been thinking about legal context for lawyers’ agreements for a long time, this seemed like a pretty big deal to him. 

We started working together and decided: what’s needed here is a protocol like the other protocols, one that will make legal terms—the most basic legal terms—findable. 

Stakeholders can build on more complicated ones if they need to, depending on the particular transaction. But the basic legal terms of the contract—what was consented to, what was agreed to—need to be findable at a known location, verifiable by the agents that are transacting, and then hashable with the payment so that if there is a dispute later, there is a single source of truth about what was agreed to and what legal terms governed the agreement. 

In many agentic transactions, you or I may have an agent booking travel for us, and maybe we’ve given it great context and a great harness, so it shouldn’t make mistakes. And probably it won’t in most cases. 

Maybe your agent ends up negotiating with United Airlines, and United Airlines also has great context and a great harness and doesn’t make mistakes. That’s great. 

Whenever there’s no problem with an agreement, the legal infrastructure doesn’t really matter. 

But as soon as there is a problem, the legal infrastructure matters. And if you have no way of knowing what it was, it will be a real stopgap in this market building quickly. 

So that was a very long first answer. But the bottom line is that Legal Context Protocol is like the other protocols in agentic commerce. It’s technology agnostic. It’s Apache 2.0 licensed. It integrates with every other protocol so that it’s totally interoperable. 

We’re co-stewarding it with Integra Ledger, but there will be lots of others contributing to it and others on the steering committee. 

The goal is, like with the others, to eventually donate it to the Linux Foundation—which is apparently how this goes. 

I’m not saying it will solve every legal problem in this new market, but it will be a big step forward. 

Jen Leonard: Okay. I’m going to try, as a layperson to these topics, to repeat back what I heard. You can let me know whether that’s accurate. 

Anytime we’ve been engaging online for the last 30 years, there are context protocols that guide our interactions with one another, but on different dimensions of the transaction. 

You mentioned ACP and UCP. They govern the flow of commerce. There are protocols that govern how payments are authorized and protocols that govern how agents connect with one another. 

And all of those have existed. Who makes the decisions on those preexisting protocols? 

Bridget McCormack: ACP was shipped by OpenAI and Stripe. UCP was shipped by Google and Shopify, but with lots of other contributors saying, “Yes, we all support this,” and then contributing to it as it builds to a place where everybody can agree this is the way it will work. 

It’s a way of avoiding a lot of mayhem in any given lane. 

Anthropic launched MCP, and they have already donated it to Linux. 

The way LCP will work, we expect it to mirror exactly how those work. It will be completely open to all contributors to figure out how far the community wants LCP to go. How much do they want to be at legal? What legal terms does everybody agree must be findable by agents? 

What are those bare minimums? 

There will probably be some differences of opinion about some of that. But eventually there will be some understanding, and the protocol will get donated to the Linux Foundation, which apparently is where all of these go. 

Jen Leonard: So LCP is filling a gap that these other context protocols don’t address, which is: what are the actual contractual terms that people are agreeing to through their agents, and how will those disputes be resolved? 

Is that accurate? 

Bridget McCormack: Almost. 

Dispute resolution is likely to be held as an add-on. That is to be determined by particular sellers in an agentic commerce market. 

If a seller in an agentic commerce market wants a dispute resolution provision in their findable legal terms at the LCP well-known URL, they can have that. But you don’t have to. 

You might have an agent if you’re selling something on Craigslist or Facebook, and you might not have dispute resolution terms. Or Craigslist might not require you to have dispute resolution terms. 

You and I enter into contracts all the time where there is a dispute resolution clause and others where there’s not. I will say, I think it’s always better to have thought about dispute resolution, and I think people should have dispute resolution clauses. 

But that’s not something the protocol will require. 

It is the ability for an agent to know where to go to find the legal terms that it is agreeing to. 

And the only other minimal thing that so far we have complete agreement on across the board is that the basic terms that are agreed to—whether they are only basic terms or whether there are more—are hashed and stored with the payment. 

That way, if there is a dispute later, there’s no question about what was agreed to. 

There might be a question about performance. There might be a question about authority. There might be lots of questions. But there should not be a question about the specific legal terms that were agreed to at the time. 

In the human, analog world, we have the contract and we can look at it. We can see what the terms were. 

Here, you’ll have a hashed set of legal terms with the payment. 

Jen Leonard: Our friend and frequent guest, Jae, had directed our attention to an explainer video Anthropic created about the MCP it created. 

So if people are looking for a description of a different type of context protocol, you can Google that and find it to give you an overview of what one of these looks like. 

I know your team has been doing a lot of work around starting to educate the rest of us in legal about LCPs’. 

Bridget McCormack: It’s interesting. 

I do think there are significant and interesting legal questions that this entire market raises. I’m excited for lawyers to have entirely new things to think about and to help their clients with, given how many of their clients are incorporating agents into their workflows—whether in commerce or otherwise. 

And I do think the commerce market is moving incredibly quickly, in large part because it has been moving without the legal industry for now. 

It’s the payment companies, the banks, and the frontier model companies that are already involved in this new market. 

But lawyers will come along. And I think there’s an opportunity for them to be extremely helpful in making sure the market works. 

Jen Leonard: Yeah. And I want to build this into presentations we do, because we’ve talked about some of the cool things on the horizon for lawyers. 

So much of the conversation is hand-wringing about lawyers not having enough to do, but these are some of the most fascinating legal issues I can imagine. 

It’s like a law school exam on steroids. 

Bridget McCormack: Yeah. And the LCP is really just a starting point. 

It’s a starting point so that the market can continue to build and grow. 

If we’re going to have agents negotiating contracts with agents or agreeing to terms with other agents, we have to have legal infrastructure. 

Legal infrastructure is how commerce works. 

So taking the legal infrastructure that has served humans in markets for thousands of years and building it into what is going to be a larger commercial market is a really interesting set of problems. 

There are so many. 

One of the things I was ideating with Claude last night was all of the next questions: what are the other places where you might want machines to be able to read some substantive legal terms, whether they’re procedural or even substantive legal principles that there is agreement about? 

What does that look like, and how should lawyers be involved in that conversation? 

I think I told you a long time ago that the ALI should make their statements into ChatBooks. That was a very 2024 idea, when ChatBooks was just what we did because suddenly we could actually chat with books. 

And that was amazing. It still is amazing. Don’t get me wrong. 

But probably the ALI needs to figure out how to make sure agents can read the statements. I think that’s the next step. 

I do think there is some really, really interesting work for lawyers coming as a result of this future market. It is exactly the opposite of, “Oh gosh, no one’s going to need lawyers anymore.” 

No, we need lawyers to help us figure out how to make sure the next generation of commerce works the way the last generation of commerce has worked. 

It’s super exciting. 

Jen Leonard: It is. It’s like upgrading the entire operating system of society by figuring out these puzzles and solving for them, which is really cool. 

I want to talk to you about one thing that I notice generally—not specifically to lawyers—which is that people use the word AI to mean one thing, when really AI is millions of different things and applications and platforms and solutions. 

And as a corollary, the concerns people have about AI frequently get glommed together without a nuanced understanding of the different concerns. 

You and David have been using a phrase in your work called “separation of concerns.” Can you talk through what people are conflating when it comes to concerns about AI, and how you’re trying to help people separate those in their minds? 

Bridget McCormack: I think what you’re asking about is something we’ve addressed in the two papers we’ve released. 

One paper, which David Fisher and I released a few weeks ago—or maybe a month and a half ago—was about the trust problem with agentic commerce, just to lay out and identify the problem. 

Then I did a second paper with our friend Dave Hoffman from Penn Law School, who is a super smart contracts professor. That was about how the legal infrastructure around e-commerce doesn’t naturally translate to agentic commerce, where the gaps are, and what the concerns are that we’re addressing with LCP. 

The concerns we’re addressing with LCP are not the concerns about AI and the mistakes generative AI makes—the hallucinations, the providing answers when it doesn’t know what the right answers are. 

LCP is totally agnostic on how good AI is, and how good your AI is versus your counterparty’s AI, and how good your AI is today versus your AI 12 months from now. 

We don’t know. We don’t care. 

I mean, I might care otherwise, but not for purposes of LCP. 

LCP is only a way to make the legal terms that govern the agreement discoverable. 

So the agent may hallucinate authority or hallucinate something important. That’s not a problem that LCP is solving at all. That’s an AI problem. 

This is a problem of machines being able to read, find, understand, and then make permanent the legal terms that govern whatever was agreed to. 

Those are the two things I feel like it’s important to separate. 

When you’re thinking about what LCP can do, it does not attempt to make up for any inadequacy in any particular AI model used in an agentic commercial transaction. 

Jen Leonard: So you’re leading this through the AAA, which is a 100-year-old organization as of last week. Congratulations on your 100th anniversary. 

Why do you think the AAA is the right institution to lead? 

Most of the players in this space are cutting-edge tech startups and people who are sort of born of innovation. So why the AAA? 

Bridget McCormack: The thing about the AAA being a 100-year-old nonprofit is actually the point. 

It is true that we don’t need to be the ones to build the technology, although I will say that Diana and her team are pretty amazing at building technology. 

But the legal frameworks to make sure that agreements are enforceable—that is kind of our central business. That’s what we have done for 100 years. 

The AAA has been administering disputes that have been resolved since 1926—over 9 million disputes. 

And I think the dispute resolution layer only works when there’s credibility around the legal infrastructure. 

So this feels like a perfect place for the AAA to convene. 

That’s all we are. We’re just a convener. 

And frankly, like I said, once there are a lot of other smart people in the mix contributing to the standard, figuring out the protocol, and figuring out where it should go, it will get donated. 

I think not having a commercial stake in transaction volume makes us, in some ways, the perfect convener. 

With lots of other protocols, nobody had commercial revenue to be gained from building the protocol. But the actors involved in shipping them and contributing to them obviously benefit from the market. 

There will be people who contribute to this protocol who will benefit from the market. 

But having a nonprofit whose mission is better dispute resolution—which includes, frankly, fewer disputes—makes us a really credible neutral in a way that’s hard for anyone else to be. 

I’ll tell you, David and I have done a lot of meetings with in-house teams at payments companies and other tech companies that are in the market, and they seem pretty delighted that the AAA is willing to play this organizing role. 

So I think that does make it a sweet spot for us. 

Eventually, I will turn our attention to how to make sure the disputes that this new market produces have great solutions. That will be something we spend a lot more time on at the AAA. 

But that only works if the legal infrastructure is solid. 

So it makes sense to me. 

Jen Leonard: Yeah, it makes total sense to me. 

You’re not focused on profit. You’re focused on getting it right. And you are at the helm of an organization that has a century-long track record in dispute resolution and trust. 

So I think it’s fantastic that you all are focused on this. 

I would love to dip back for a minute into some of the contractual issues at play. 

One of the things I know you and Dave Hoffman have written about is the concept of ratification. 

To make this really concrete for people who maybe are a little bit disoriented by the tech, can you walk us through what an example might look like in an agentic commercial setting without an LCP? 

What could go wrong that relates to ratification? 

Bridget McCormack: Yeah. This paper that Dave Hoffman and I published—which we may do a longer version of over the summer—was intended to make concrete how the kinds of doctrines that held during an e-commerce era, where there was a human clicking buttons, might not hold. 

So we opened the paper with a scenario that makes this point concrete. 

A mid-sized industrial manufacturer deploys a procurement agent, which is a kind of agent that businesses are using to handle routine supplier onboarding. 

As a result of a configuration error, there is a problem where the agent accepts counterparty terms that include a broad indemnification clause running in the supplier’s favor. 

Under the clause, the manufacturer agrees to indemnify the supplier against any third-party claim, including claims arising from the supplier’s own common conduct. 

Over the weekend, the agent completes 12,000 transactions across hundreds of suppliers. 

Monday morning, operations proceed normally. Goods are accepted, shipments move, invoices are paid. 

By the time everyone notices, ratification is already complete on most of those deals. The company has accepted the benefits with notice of the terms, even though no human ever read them. 

So the exposure is obviously substantial. 

There are lots of doctrines besides ratification that courts have relied on for a long time and will continue to rely on. And the question will be whether the company has a record that is clear, so that it can litigate those issues efficiently, or whether the terms and what went wrong are impossible to find and turn into decades-long litigation where everybody is trying to figure it out. 

That’s the kind of thing I think we will see. 

We will see mistakes like that get made. 

And if people are thinking at the front end about what the legal terms are—including authority, and how much authority the agent should have had—the disputes around that kind of large mistake will be far simpler to resolve and work through than if there is no way of finding the single source of truth about what terms were agreed to, what authority the agent had, and so on. 

We walked through a number of the doctrines that grew up around e-commerce that don’t really translate, to paint the picture about why it makes sense for lawyers to spend a little time thinking about this today instead of next year. 

Jen Leonard: It’s making me think back to our earlier conversation about Mythos—the scale and scope of what could go wrong is so much greater with agentic systems. That makes it so much more urgent to address. 

Bridget McCormack: Yeah, exactly. 

In commercial transactions, we’re talking about money—not air traffic control, not planes going down—but still, it will be a significant impediment to the success of the market if the legal infrastructure isn’t thought through at the front end. 

As we’ve talked about in other contexts, lawyers, for lots of reasons, are used to approaching problems after they’ve happened instead of thinking about how to head them off at the pass. 

And this is a moment for lawyers who want to be involved in the creativity at the front end of: what could we do to support all of this new commercial activity? 

We can be a partner and a helpful force in the success of a market that is coming one way or another. 

Jen Leonard: Okay. Last question, Bridget. 

On a practical level, if you’re a general counsel or a law firm partner listening to this podcast, what should you be doing right now about this? 

Bridget McCormack: I guess the one thing I would say is to make sure you have an inventory of where agents are already touching transactions. 

Is it just internal? Is it in procurement, customer service, marketing? It’s probably more than you think. 

I think agents are just creeping into our lives and doing more and more because they’re easier and easier to use. 

Then I would look at your form contracts and see whether they contemplate machine-mediated formation and performance. Most probably don’t. And some of those might fail in surprising ways. 

Then I would start watching the protocol layer. 

The companies your business depends on for payments and connectivity—Stripe, Google, Visa, the SaaS vendors—are forming these protocols to work together as a community to make all this work. 

But if you’re the general counsel at a business that’s going to be involved in agentic commerce, get involved in the LCP and contribute to it. 

And maybe start thinking about what else we might want to build—what other legal principles, procedural or substantive, we might want to think about building into future standards. 

The LCP is only an opener. There’s a lot more we can do. 

Jen Leonard: Well, that is a positive note to end on, both for society generally and for lawyers specifically. 

We have a lot of work to do, which is fantastic. 

Thank you so much, Bridget, for sharing this really cool project with everybody. 

And thank you to everybody out there for tuning in to this edition of AI and the Future of Law. We look forward to seeing you next time. Until then, be well.