Garfield Law: Inside the World’s First AI-Native Law Firm
Introduction
Jen Leonard: Hi, everyone, and welcome to the newest edition of AI and the Future of Law. I'm your co-host, Jen Leonard, founder of Creative Lawyers. Here, as always, with the wonderful Bridget McCormack, president and CEO of the American Arbitration Association. We're here, as we are every episode, to talk about all the wonderful and exciting things happening in the land of artificial intelligence and the future of law.
And as part of our new partnership with the Practising Law Institute this season, we've been thrilled to be joined by some guest experts who are doing really cool work in the space of AI. In this episode, we're joined by Phillip Young and Daniel Long from Garfield Law. We had the chance to talk about Garfield Law in an earlier episode that caught fire online among our audience, so we wanted to have Daniel and Phillip join us in the studio to share more about their projects.
Phillip is an ex–City commercial litigator and is now the named solicitor for Garfield Law. In that role, he reviews all the platform’s outputs from Garfield Law—they'll tell us all about how Garfield works. Daniel Long is a quantum physicist turned AI entrepreneur—my old job—and he is the lead architect of the underlying system and a former data scientist at Garfield Law.
So welcome to both of you to the podcast. And Bridget, hi, it's great to see you.
Bridget McCormack: Yeah, you too. Happy Friday, and I'm so excited to get to talk to you, Phillip and Daniel. It was really fun to talk about you, and I bet it'll be even more fun to talk with you. So thanks so much for joining us. This is going to be a really fun conversation this morning.
AI Aha! Moments
Jen Leonard: So every episode we like to start—before we talk about the law—by hearing how you're using AI in your personal lives to do something unique, interesting, and maybe even magical, to inspire other people to try it as well. We call this our AI Aha! segment. So, Phillip, maybe we could start with you. What have you been using AI to do recently?
Phillip Young: I've found that I'm now using consumer-grade LLMs as a sort of personal lifestyle assistant. The best example I can give is one of those little niggling things that would otherwise be a pain in the neck, but an LLM can answer it really quickly.
I don't know how it is in America, but British smoke alarms always seem to go off by accident at 2 a.m. It's never during a normal time—only ever at 2 in the morning—and that's the last time you want to be hunting for the manual to figure out how to reset them. So last week when that happened, you wake up at 2 a.m., not even sure who you are, let alone what's going on, and you just type into ChatGPT: "Here's a photo of the smoke alarm. It's going off and it shouldn't be. How do I fix this?"
And off it goes—it comes back with the instruction manual and gives you detailed instructions on how to reset the darn thing. You don't have to hunt through the folders behind me for the instruction manual. You get it done and go back to bed. That's been my latest experience in making my life a little nicer.
Bridget McCormack: Jen and I use it for tech support all the time—I call it tech support because I am not a quantum physicist. When I need to fix a smoke alarm, I definitely need tech support. But it does make life easier in that way. How about you, Daniel?
Daniel Long: People think quantum physics is complicated, but I find I have some really basic questions. For years I’ve been trying to understand music theory to get better at guitar. I have musician friends who have explained it to me once or twice, but it never quite makes sense.
So I've started asking the most simple, dumb questions and then building up the complexity with Claude (which I use most of the time). Finally things are making sense to me—things that took years to understand otherwise. So that's me with the guitar and Claude, trying to make some music.
Garfield Law, the First AI-Native Law Firm
Bridget McCormack: Well, we're excited to have you all here because you've launched a pretty interesting new legal practice—Garfield AI. Why don't you tell us a little bit about what Garfield is and what it does? Our listeners know a bit about it, but we probably have new listeners since then who only heard about it secondhand, and I'll bet you could do better. Maybe you can give us an introduction to Garfield.
Phillip Young: As you ladies mentioned, for many years I was a big-ticket commercial litigation and arbitration lawyer in the City of London. From time to time, my brother-in-law—a lovely guy called Andy, he's a plumber near a city in England called Sheffield—would have the classic problem that small businesspeople have: he just wouldn't be paid.
And it wasn’t usually because the debtor was insolvent or had a genuine grievance with the service. It was just that debtors know some small businesses don't have the time to chase a debt and find even the small claims court in England (which is supposed to be a simplified system). Intimidating! So, being the good brother-in-law that I was, I would step in and help him collect these debts and go through the court system.
But sadly, in our country, not everyone has a brother-in-law who's a litigation partner. So the problem is widespread, and I always thought to myself, "This is ridiculous—there has to be a better solution." Then when ChatGPT-4 came out around March or April 2023, I thought one could build something with this that could provide a solution.
I think part of the reason behind that is that I'm sort of 50% lawyer and 50% giant nerd. I've always played with computers since I was knee-high to a grasshopper, and I just got the computer side of it and the law side of it. I thought, yeah, you could build something to handle this sort of case, because these cases tend to be fact-heavy. They don't tend to be legally complicated—it's primarily process. It tends to be a sequence of what Dan (I think beautifully) refers to as "closed problem spaces." And that's an ideal challenge for an AI to tackle. So that's where the idea of Garfield came from.
Then I found Dan, and between the two of us over the last two years—along with the team we built—we built this product called Garfield that now seeks to do what a good English law firm would do to pursue a small debt claim, all the way up to the edge of trial. It is regulated as though it is an English law firm. So essentially it's an English law firm that provides services through AI, which makes it a first in our jurisdiction and, I believe, the world.
So that's a quick introduction to where Garfield came from and what it does. I'll hand over to Dan for a few seconds to talk a bit about the tech side of things, and maybe how he's found it having to work with a boring lawyer.
Daniel Long: Philip described a "closed problem space"—but the schematic he gave me when we first started looking at a small debt claim in the UK was anything but closed and simple. I don't know how many boxes and steps there were. Since then, we've managed to shoot things down a bit, but inherently there's a lot of different pathways. Luckily, they do tend to close off eventually, which is what we mean by a closed problem space.
On the tech side, when we first started, we initially tried to see if we could handle all of the individual stages we wanted to. The big fear was hallucination, and generally leading the user down a path they shouldn't go. Could we design the product from the ground up in a way that wouldn't lead a user into those issues was the initial question. Pretty quickly we could see this was going to be possible. And because the existing process is so time-heavy, we thought anything that streamlines a lot of those steps is going to be really valuable.
On the tech side, Phillip describes himself as a former City litigator—he does have a lot of coding experience. A lot of the early code is still his; when he's got time for it, he'll join in and write bits and pieces.
Phillip Young: Dan is being really kind. He and the other professional programmers, when they look at my code, they're always really kind about it—but I can tell that deep down, on some level, they're slightly dying. But they are able to turn it into professional-grade code.
Daniel Long: One thing we do—which will make sense to legal audiences—is numbering paragraphs in documents. As a non-lawyer, I'd never seen that before, and I certainly don't code like that. So we're always trying to remove lines in Phillip's code that are written in a case-law style rather than in a script.
Jen Leonard: And how did you connect with your small-business clients? How did you get the word out about what you were doing? Were you contacting them when they were already in court, or were you reaching out beforehand? How did you spread the word?
Phillip Young: We had some early adopters lined up even before we got regulatory approval. I have to say, the regulatory approval took quite some time—and rightly so, because we were asking our regulator to do something that had never been done before. They spent a lot of time working with us on all the ethical and regulatory issues it raised.
I think it helped that one of my smaller areas of practice when I was practicing law was regulatory law and professional discipline. I know the regulatory code really well, and I understand that field. So I had already thought about a lot of the issues while building and designing the architecture.
That process took eight months in total, which gave us a lot of time to line up some early adopters. What we were anticipating—because like all lawyers, I'm very cautious and careful (I always stop and think before I do anything, even if it's just going to make myself a sandwich)—was a very quiet, stealth rollout over six months without the world at large knowing we existed. That way we could get some claims to go quite far through the court process and see how they did at each stage. But I'm afraid the world had other ideas: somehow the "Financial Times" found out about this, and the next thing we knew, we were in the "Financial Times."
After that point, my life became a lot more hectic because suddenly there were journalists lining up to talk to us. And interest became a lot greater—many more users applied to join the platform.
Bridget McCormack: I probably should know this, but in the United States most small and medium businesses—even fairly large ones—don't have in-house legal help, and they largely can't afford outside legal help. Your brother-in-law the plumber had you to turn to. Is it the case that most small and medium businesses try to navigate the legal process on their own? Do they just give up? Do you have any sense of what people typically do with problems like these?
Phillip Young: It's exactly the same in England and Wales. If you're an "S" or an "M" in SME, you're unlikely to have a general counsel or much of a legal budget, and you would either DIY the process—if you could bear it and didn't feel daunted by it—or for larger-value debts you would instruct external counsel. But certainly for small debts (sums up to around £10,000), it's usually uneconomic to hire an external law firm. That's not a criticism of those firms.
It's just that historically it's been a manual process—very time-intensive and therefore naturally costly. And although in England we have a different rule than you do about cost recovery (in England the general rule is the loser pays the winning side's legal costs), that rule doesn’t apply in small claims. So that makes it even more uneconomic to hire an external law firm.
The result is that, depending on which survey you read, each year anywhere between £6 billion and north of £23 billion of debt just isn't collected—that is, small debt in England. In a big economy, that's a huge sum of money, and it's really inefficient.
Bridget McCormack: I didn't realize you were an expert in regulation—that definitely helps. In the United States, I don’t know of a single jurisdiction that’s ready to approve a solution like this, even though I think we have a similar problem. The small claims process here is similar: it’s usually only factual disputes, right? Those are exactly the kinds of disputes you could imagine an LLM being trained to handle quite well, especially if you do some reinforcement learning around the different stages of the process.
But I can’t imagine any U.S. jurisdiction allowing this anytime soon. So what are the cultural differences between U.S. and UK regulation, or what was your secret sauce? Why do you think you were able to navigate the regulatory path so successfully?
Phillip Young: I think there are a few factors. First, our regulator in England—the SRA (Solicitors Regulation Authority)—has a public duty to promote access to justice, and they take that really seriously. They've been well aware for a long time that there's a problem around small claims. It's not just people not being able to access the court; it's also people who do access the court as self-represented litigants, presenting poorly prepared cases to overworked judges. They take up a lot of judicial time, they don't necessarily get the quality of justice they deserve, and it makes an already overburdened system even harder to operate.
So the problem is on two levels, and the SRA is always looking for solutions to things like that. Secondly, there's been a lot of policy concern in England—from the regulator, the senior judiciary, and the government—about people turning to unregulated products, particularly LLMs (consumer-grade tools like ChatGPT), to help them navigate the court system.
If that's the only product available, people will turn to it. The problem is they get answers that are perhaps inaccurate, possibly misleading, or worse. In a litigation context, that's incredibly dangerous because of rules about contempt and not misleading the court.
Everyone's been very concerned about that and keen to get to a world where regulated products are available. Third, I think the SRA felt this was coming—it was only a matter of when. When we presented ourselves to them, initially they looked very closely at us as people and felt we were a good team.
We obviously understood what we were doing, we had the right experience and the right support, and they felt that if they were going to back anyone, they would back us to begin with. It's not like they've given us regulatory approval and then just let us go off into the blue yonder to do whatever we want.
They've been very sensible. We're doing quarterly check-ins with them—we're trying, to the extent we can (consistent with our duties of client confidentiality and such), to give them know-how and feedback. Because I think if you're the first, you have a sort of public-duty obligation to help the powers-that-be build out the regulatory framework to handle the next wave of products that people are clearly going to build.
I'm no genius, and if I thought of this, plenty of other lawyers will think of similar products. I've already been shown a bunch of products like Garfield being built in England, so I know plenty of others are coming. The regulator has been thinking about all of that too.
So those are really the factors that led to this. I wouldn't say—no disrespect to my profession—I wouldn't say we're an unusually creative legal profession in England any more than any other jurisdiction. If anything, most English people look to Americans and think you guys are five years ahead of us. I was quite surprised to hear that you might not be ahead of us on this one for a change.
I don’t know—Dan, did you have any thoughts to add from a technical perspective?
Daniel Long: One advantage of our approach is that we can track how users move through the system and keep an eye on their progress. That’s hard to do in a typical law firm, where so much information is unstructured. But in our system, everything is tracked; we've discretized everything in terms of where users are and which routes they go down. I'm sure our regulator finds the reports we produce on user journeys pretty unusual—and we're quite excited about that capability.
Bridget McCormack: That's really interesting. That hadn't even occurred to me.
Daniel Long: We're constantly analyzing whenever a user does something in the system. We even use LLMs themselves to figure out if something went wrong, what the user wanted to happen, and how the system could work better. (It means I don’t get much sleep, because feedback is constantly firing off to me, Phillip, and the rest of the team at all hours!) But we'd like to produce some interesting stats about how people are using the platform—what’s working and what isn’t—in a way that would be hard without an AI-driven system.
Phillip Young: There's an architectural point here as well, which goes back to understanding the regulatory framework we operate in. From day one, I was conscious that if we were going to provide a regulated legal service—and you do have to be regulated to do what we do, you can't conduct litigation in England without being regulated—then we had to account for all of our obligations.
You have to consider what's called the Hamid obligation (the duty you owe to the court as a lawyer). You also have to maintain client confidentiality, deal with conflicts of interest—all of that ethical stuff that will be very familiar to listeners in America, and it's basically the same in England.
So we built the architecture of our software with all of that firmly in mind from day one. I think the regulator took a lot of comfort from that, because they could see this wasn't the sort of product where—again, no disrespect to tech bros—a bunch of tech folks just built something and threw it at the law and hoped it worked.
This is a product designed and built by a lawyer and a technologist working together to make something accountable and responsible. And ultimately, they also took comfort from the fact that we have what we call a COLP (Compliance Officer for Legal Practice)—that's me—in this law firm.
So I'm accountable to my regulator, and I'm accountable to the judiciary, for the functioning of Garfield and its compliance with the ethical rules. I think accountability and responsibility are the two themes: if you're building a product like this, you have to keep those at the forefront of your mind the whole way through.
Jen Leonard: Philip, you mentioned the word "platform" earlier. As a user, are you engaging with the court system exclusively through a platform? How does that work with the small claims court where you are?
Phillip Young: The small claims court has the almost unique distinction in England of having an API (an application programming interface), which means users can exchange data with the court programmatically. That's what Garfield does—Garfield is essentially integrated into the County Court service. Garfield can send claim forms to the court and other documents to the court, and it can receive data back from the court. I think that's very much the future of law firms, frankly, because it makes things a lot quicker and easier. In fact, in England we're on the verge of phasing out paper and pen. In the next month or two, you won't be able to send a paper claim form to the court and get a claim issued. When I think about when I started in the 1990s as a lawyer, that was almost unthinkable. It shows how far we've come.
Jen Leonard: Could you resolve a dispute without going to the courthouse?
Phillip Young: You could, if you didn't need to go in front of a judge. Imagine that you're using Garfield: you've issued a claim form after sending the pre-action letters and not getting a satisfactory response. In response to the claim form, there's a range of possibilities. You might get paid—in which case, great, the claim is over. You might receive no response—in which case Garfield applies for default judgment (again a programmatic process that Garfield handles for you). You might receive an admission (which usually comes with a request for more time or lenience), in which case Garfield helps you go through the process of resolving that. Or you might receive a defense—perhaps even a defense and counterclaim—in which case you're going to a small claim trial (which is not a very grand thing; it's a one-hour hearing in front of a district judge in a county court). But if you go down that track, Garfield holds your hand through it. There's a compulsory mediation on that track as well, which Garfield will help you attend (it's by telephone). And if you settle, then you settle and the case is over.
Garfield handles all of those possible pathways a claim could take.
Jen Leonard: You mentioned, Philip, that some of the things happening with the API—and certainly things that would have been unthinkable in the '90s—have now become reality. How are you thinking about the next phase, or the future of legal services in an AI infused world?
Phillip Young: I have some thoughts on this, based on the last two years of everything Dan, I, and the team have done with Garfield. To explain where I think we're going, I need to say where we came from. When I joined the profession in the 1990s in England, all law firms looked pretty much the same. The only real difference between them was one of scale. Some were small high-street firms, some were larger City firms, and basically there was an office (or offices), and the parts the clients saw were a lot nicer than the parts the poor lawyers and everyone else had to work in.
There were always some elderly people who were the partners—who, in the 1990s, were, I'm sorry to say, overwhelmingly male and of a certain class. Then there were younger people who were reasonably gender-balanced (the young lawyers coming into the profession, like myself). And then there was a great strata of middle-aged ladies who kept the whole thing afloat—the secretarial and support staff.
If it hadn't been for them, every law firm in England would have collapsed. You know the old joke: you need ten people with their feet on the ground to keep one person with their head in the air. That was a law firm. There were loads of law books, and you'd have a court clerk, and that was your law firm.
Amazingly, where we're going to end up in the next ten years is a highly integrated kind of business that fundamentally exchanges data a lot more swiftly with all of the parties it deals with. For something like Garfield—primarily applicable to smaller matters, but you can generalize to bigger cases and transactional work—it's integrated with the court service and exchanges data programmatically with the court. That's only going to increase, particularly as our government builds more APIs into more areas of the court service. There's going to be one for higher-value civil claims, one for family cases, one for criminal cases.
All of that's coming. It might be that the arbitration community will have to start doing something similar with the big arbitration bodies—who knows. We're integrated with Companies House (our companies registry in England). For example, if a user wants to bring a debt claim against a corporate debtor, Garfield automatically goes to Companies House and checks the status of the debtor—are they solvent or insolvent, where are they, it pulls their latest filed accounts (if any) and gives a view on their solvency to help the user know instantly whether it's worth pursuing that party or if they're wasting their time and money. We're integrated with various other services as well. Inevitably, we'll integrate with the Land Registry, which has obvious ramifications for enforcement of judgments, and other services.
Fairly soon we're going to roll out a service—and I don't mind telling you this is coming in just a couple of months—where Garfield can send automated SMS messages and make automated phone calls. We'll control that so it can't become a nuisance or anything like that, but you can see how you'll end up with a law firm that has all of these automatic integrations.
That saves a lot of time and trouble, and it will also let firms exchange data much more effectively. The English court system is trying to promote a common data standard, and I think that will drive a lot of this. When I think back to practice, one thing that used to really annoy me was serving documents—trying to get documents to another law firm. Sometimes you had to kill trees to do it, sometimes you had to serve after hours when no one was available, sometimes you tried emailing ludicrously large PDFs. I think all of that is going to go away. It's going to become almost like a common data room where you just exchange documents, upload them, generate bundles, and the court can see everything in real time. If you have a hearing coming up and the judge is pre-reading the case, the documents are just appearing on their system.
We're going to have a lot more integration in how law firms interact with the world around them. Another integration we built: Garfield integrates with our users' accounting platforms, so you can pull unpaid invoices and data straight from their accounts. Again, that's a big time-saver. This is the future of law firms, I think—all of these integrations built in, and the movement of documents becoming much swifter and more effective.
Daniel Long: Something we joke about internally is that if you found the secret of the universe, it would probably be in JSON format. (That's a nerdy joke, but JSON is the data structure modern software uses.) Trying to populate PDFs or use XML—these older formats—can be frustrating sometimes. My dream is that someday everything would just be JSON documents floating around with shared schemas—super simple.
Bridget McCormack: I want that future right now. I was going to ask what your view is of what that will mean for the human lawyers who can have a practice where they're not spending time on busy work and maybe the frustrating work of finding the right documents and connecting them to where they need to be.
Do you think we'll need more lawyers—because lawyers will be able to create more value? Or do you think it might be time for people considering law school to think about becoming quantum physicists instead?
Phillip Young: I might express a view on that—and Dan may have his own as well. I'm an AI optimist as far as the law is concerned, and whenever I speak at conferences I always tell people I think there's going to be a need for a lot more lawyers. If the conference is full of lawyers, there are happy smiles; but if it's full of non-lawyers, everyone looks really miserable for some reason! (Goodness knows why—we are a wonderful profession.) But the problem we have in England, and I imagine it's a problem in every jurisdiction in the world, is that there's a huge strata of work that's uneconomic for lawyers to do at the moment.
I think what will happen is that, because the main reasons for that are the low value-adding but very time-intensive tasks needed to progress those kinds of matters—whether transaction or litigation— as that work is dealt with by technology, suddenly those transactions or cases are going to be economic to do, and therefore they will be pursued and lawyers are going to be engaged.
In a litigation context, of course, if a claimant decides to bring a claim they otherwise wouldn't have, that's going to be work for the defendant side as well. So I think there's going to be an awful lot more work for lawyers coming out of this. And I think the good news is that all that work that none of us went into law to do (but we were presented with as junior lawyers, where someone said, "Here's a room full of documents—can you do a due diligence exercise or a big discovery exercise?")—a part of you internally just died, but you did it because you had to. With this technology, even as a junior lawyer you'll be doing higher-level skills at a much earlier stage.
For example, in England the training of lawyers is very different at a small firm than at a big firm. If you were a trainee at a small firm, you'd be thrown in the deep end on day one and expected to do higher-level things like manage clients and shape case strategy and come up with the best legal arguments. At a really big firm, you're basically used as a document monkey—put in a big room and told to read a lot of documents.
I think even big firms are going to end up adopting a small-firm training model as a consequence of this technology. I think they've got no choice. I think they're going to be forced to by their client base. We're already—even just two years into this—seeing lots of big companies' general counsels insisting that their external law firms take up AI (and take it up aggressively) to drive down the legal spend those companies have. There's no choice here—it’s just going to happen.
Daniel Long: From my naive, non-legal perspective, I have that techno-optimist view: we could move to a future where you're not held back by all the mundane paperwork-filling stages, and trials happen a lot more quickly, and appeals can happen much more quickly. That could mean the whole economy benefits massively—where a legal dispute could be resolved in a matter of a month or two.
And it means the judgment from the judge would still be given as much thought and attention, but all the parts that can be compressed (that don't really require a judge) can be massively accelerated. I'm quite optimistic that we'll see economies that embrace this technology really benefit. (That's the techno-optimist side of me coming out.)
Bridget McCormack: I think I'm both of those—I think I'm an optimist for the legal profession and an optimist for the economy in general. So I tend to agree with both of you. Right now I think lawyers are only hunting in about 10% of the potential legal market because—like Phillip said—so much of it is uneconomic and they can't make ends meet.
But when all of that changes, there's the rest of the iceberg under the ocean—and an awful lot of work to be done that I think will give lawyers more satisfying practices and grow the economy. So I agree with all of that.
I'd love to talk about how you solved the hallucination problem. In many jurisdictions now, we hear a hallucinated case citation story every day. Even judges are getting into the action—I think they must have felt left out. And that's basically all we're talking about over here. I don't know if maybe in England you're more careful and nobody ever includes hallucinated case citations or quotes in their legal briefs.
But how did you solve for that with Garfield? Why was the regulator convinced they didn't have to worry about hallucinations with your product?
Phillip Young: I can assure you, people in England are no more careful than anywhere else. It's something we've been very conscious of—and so has Dan—since day one, because it was obvious even two years ago, when this journey was beginning, that hallucinations were a serious issue. So we designed our product very much with that in mind, and we built in a whole series of safeguards and guardrails. Some of those I'll let Dan talk about, but I'll mention one thing I think is really interesting—and it's an architectural thing.
Because we needed to ensure the risk of hallucinations was reduced to the bare minimum, we didn't just build an LLM-only product. In that sense, the press describing us as an "AI company" is doing us a slight disservice, because the product we built is a hybrid.
It has a deterministic expert system and a probabilistic LLM-based series of systems. The expert system is deterministic, so we let it be in charge—that then controls the universe of possible things the LLM can or might do. If you're not a techie, one way of describing it (rather inexactly) is that it's like having the result of my years of practice in charge of the LLMs, constraining what they can and can't do.
That helps minimize the risk of hallucinations. That's one thing. Another is that we tried to adopt the best practices a law firm should adopt. We tell the user at every stage what Garfield has understood from the documents they've uploaded and from the information it's been given. The user can always see what the platform understands and what the next options are.
And it's not autonomous—the user has to approve each action before the system takes it. That's exactly analogous to the way, for example, my former law firm operated: whenever we drafted something for a client, we would run it past the client for approval before it went anywhere. At the end of the day, the facts come from the client and the arguments and presentation come from the lawyer.
So you always need the client to check the facts. Garfield takes exactly the same approach. If the user spots that something has been misunderstood—just as might happen between a human lawyer and a human client—the user can correct it. That's a very important safeguard as well.
The third relevant factor is that, depending on what statistics you look at online, hallucinations are inherently a thing with LLMs. As a result, all the big players in the space have poured huge sums of money into research to mitigate them. And as models get better and better, the incidence of hallucinations does seem to drop dramatically.
Certainly from the SRA's perspective, the point they made to me was: if our product is operating at least at the level of a reasonably competent human, that's already more than good enough. Eventually, with technological progress, you'll reach a point where it's at the level of a superhuman—which is clearly way better. And I have to say, I've seen some products (I won't mention who) in the last few weeks that are specifically designed to fix hallucinations in legal documents and legal products. Those are coming.
They're going to be on the market in the next year or two. So this is a very well-known problem, and an awful lot of thought and work is being put into making sure it's manageable. I'm sure Dan has his own thoughts on this topic, since it's fundamentally a technological one.
Daniel Long: Yeah, I agree with all of these points. The thing I would add is that if you stood over someone and said, "Write me a contract right now" or "Fill in this claim form perfectly," they would probably get it wrong—they'd make mistakes. And that's true for humans, and it's true for LLMs. LLMs have an additional bias in that they want to please you far more than the average human does, and that only exacerbates the problem.
In other domains, the way this is resolved—for example, in software engineering, you wouldn't directly paste the first output of an LLM into your codebase and then deploy it and watch errors flood in. Instead, you use agentic systems that will write some code, then generate tests or iterate the code to match the tests, and then add new tests as it goes. All of these traditional software development ideas can be applied with LLMs. And because software engineering is a domain where you have very verifiable outputs, it lends itself nicely to this. The challenge in other domains is to pose problems in such a way that you can have verifiable outputs, or at least break problems down.
In Garfield's case, we don't just take in the user's message and directly output a response. In fact, there are many LLMs running in parallel, determining how different parts of the claim should be modified based on the user's instructions. Then, like Phillip said, all of these modifications are passed back to the user: for example, "You gave us this invoice. We saw this company number on it and it looked like your company number, so we set it as such." There may be mistakes, but it's on us to articulate as clearly as possible to the user what we've understood. Once we've extracted the information that way, we have the expert system which—like Phillip said—is a codified version of his brain that sets certain rules, like fundamental rules that might invalidate a claim (a £10,000 limit on a small claim, statutes of limitation, things like that).
We've got to look out for those. And because we're not using an LLM to judge "Does this claim look valid?"—we're actually using traditional flow control in our codebase to make sure conditions are met. By converting the problem into more of a software problem where we have verifiable data, we can make sure we're not hallucinating outputs.
That, coupled with the fact that we try to be as transparent with the user as possible—the system is making their life easier, but we don’t let them take their hands off the wheel and assume everything will be correct.
Phillip Young: I think there's an analogy there for the way law firms operate, because a review–correction–iteration process is how law firms operate. When I think about documents I've drafted in the past—whether witness statements or skeleton arguments (which you would call briefs) or other documents—you might do a first draft. If you're a junior lawyer, it might go up to a more senior lawyer; they might review and amend it. You might then re-amend it; it might go to a partner; they might change it again. The finished product that goes to the court or the client or the other side is often quite different from the very first draft.
It's a similar analogy with the way AI-based legal products should be built: it goes through a process of review and iteration and polishing before the final product is presented or goes out. Every lawyer in practice, anywhere in the world, understands the importance of that.
Daniel Long: One thing that just occurred to me—something I forgot to mention—is how Garfield tests itself. We spent maybe 30–40% of our time building the actual product, but the majority of our time is spent testing the product. That can be traditional software tests (plugging in inputs and expecting outputs), but we actually take it a step further.
James and the team were really keen on this and spent a lot of time improving it: we have Garfield talk to itself and pretend to be a user that has a small debt (or maybe doesn’t have one). Then we can see how well it responds to that simulated user and whether it correctly classifies them, and whether it hits any issues.
Because like I said earlier, we started with this very complicated schematic. We tried to simplify it, but inherently the process is very branching. For a claim-handling system to navigate that without mistakes, we realized we could actually use an LLM to simulate the system and run multiple scenarios in parallel. That allows us to flag issues really quickly.
Jen Leonard: It's amazing what you've built, and I wish we had another hour to learn even more about it. But we're coming to the end of our time with you, so maybe we could wrap up with one last question—and hopefully we end on a positive note. How have other solicitors and barristers responded to what you're doing? Have you become leaders? Are other people inspired to follow your example and think differently about how legal services are delivered?
Phillip Young: I was expecting a lukewarm welcome at best—possibly even a slightly negative welcome—because the profession has always been, shall we say, small-c conservative in England. But I've been really surprised at how welcoming everyone has been. When I look at the users on our platform so far, a sizable chunk of them are law firms—either pursuing their own debts or using Garfield to pursue their clients' debts in a way that's more economical for them, gives them more margin, and allows them to scale up faster as a firm. I think it's because everyone in the profession can see this technology is going to be useful, and they want it to be done right—because it is coming, love it or hate it.
And the fact that it's being built by a team of a lawyer and a technologist, and the SRA is regulating it, and everything's been done in accordance with the approach you should take if you're being responsible—I think that's made people feel a lot more comfortable. I think the profession also sees that there was a real problem here that Garfield is trying to fix for the benefit of the community.
One thing about lawyers worldwide I've always found is that people become lawyers because they genuinely care about justice. I've never met a lawyer who didn't become a lawyer because they care about justice and they care about people being able to access the courts if they need to. Lawyers do get concerned when there are justice gaps, and I think that's part of the reason why lawyers feel this is a good thing and they're in favor of it.
So yes, it's been received very positively—not just by the profession, but also by the judiciary. We're very fortunate in England that our senior judiciary at the moment (the Lord Chief Justice, the Deputy Head of Civil Justice, and the Master of the Rolls) are all very forward-thinking. And in fairness, so are the members of our Supreme Court and our Court of Appeal. They're very much in favor of AI being used to make the justice system quicker and more efficient and to provide a better quality of justice than before.
They have cautiously welcomed this development as well—without, of course, endorsing it in any way, because they're judges and they can't endorse a private product. But there's a sort of groundswell of warm feeling towards it that we've been very grateful for.
Jen Leonard: That's wonderful. I certainly don't have the expertise that either of you have—and this may be an oversimplification—but I think Bridget would share the takeaway that throughout our conversations with great people doing innovative things, the through-line is mindset around what's possible and the reason we're all part of this profession and what we're trying to do.
And that definitely comes through in your work. We're just huge fans of what you're doing and can't wait to see more. Thank you for spending your time with us today and sharing your experience. And thanks to everybody out there for tuning in—we'll look forward to seeing you on the next edition of AI and the Future of Law. Until then, take care and be well.