Subscribe
Share
Share
Embed
Bots at Work is the new season from Fringe Legal, which explores how AI is changing the way work gets done, with a focus on real-world impact over hype. It looks at how operators, builders, and leaders are using AI to reshape workflows, decision-making, and business models, especially in professional services like law. The show focuses on practical insights, emerging patterns, and honest conversations about what works, what doesn’t, and what comes next as intelligence becomes cheaper and more embedded in everyday work.
Ab (00:32)
Helen, thanks so much for coming on the pod. Excited to have you on.
Helen Fan (00:36)
Hi Ab, it's nice to join your conversation.
Ab (00:40)
Yeah, and I've been following a lot of your writings, a lot of your experiments. And I wonder if we can start there. So you've been running this hundred days of AI experiment and you've just gone over the 50 day threshold. So you're halfway through. Can you take me back to what made you start? And yeah, where are you with the experiment at the moment? And just maybe tell people a little bit about what it is.
Helen Fan (01:06)
Yeah, so actually I call it OpenClaw Law LLP. So it's my personal technology experiment. I, so the story was like, I first published something called legal AI value stack. So that's a framework. after that, yeah, yeah, yeah, of course. So after that, I kept asking myself, can I actually walk that roadmap?
Ab (01:22)
We'll come back to that for sure. Yeah.
Helen Fan (01:31)
not just to write about it, but to do it. So I started building my AI native law firm from scratch. And it runs on the OpenClaw and inside this court. I'm the supervising attorney and I have two AI agents helping me. The first one is Morgan. So Morgan is my senior associate. And the other one I call it Cleo. So Cleo is my junior associate. And so Morgan, as my senior associate, she handles strategy.
on client management, email screening, and she can also break complex issues into smaller pieces and assign them to sub-agents or Cleo And for Cleo so she does legal research, dig into the case law and draft the first draft of the memos. So she's the one who do all the dirty work. And also they talk to each other with a very different from what you're seeing in other agents' experiments.
Ab (01:59)
Mmm.
Hmm.
Hmm
Yeah.
Helen Fan (02:27)
So I have a multi-agent and agent to agent communication. they can argue with each other. That's actually the most interesting part. I built a system where Cleo can challenge Morgan when she feels like she disagrees with Morgan. Every time when they disagree, I get what I call an argument report.
So that is a summary of exactly where they see the problem differently. And that can tell me where to focus my review, maybe to some degrees that can help to reduce the hallucination AI creates. But that's my understanding. didn't do it like an academic project. It's just my guess. So yeah, that's pretty much about my recent experiment.
Ab (02:55)
Hmm
Hmm.
Yeah.
I think that's super cool. And for anyone who may not be familiar with OpenClaw, it's a open source, I guess, local first AI agent framework. That's how I would describe it. And essentially, it is agentic. So unlike a chatbot, as Helen described, you are able to assign, delegate, and offload tasks. So they are able to continue working.
kind of asynchronously in the background. So I also have an OpenClaw instance. My OpenClaw is called Clawdia How hard was it for you to set all of this up? And has it been because what you described sounds a very smooth, issue-free operation as someone who runs a OpenClaw instance.
⁓ I think the reality may not be that. Is that the case or am I doing something terribly wrong because I spend probably as much time troubleshooting my instance as I do getting the benefits from it.
Helen Fan (04:21)
Yeah, absolutely. I totally got what you're saying now. So first OpeClaw's overall stability is not good enough. It breaks a lot every day. So agents can go in circles because I have two agents and they can talk to each other and never stop forever. and sometimes your rules that you set up in your soul.md
that will not be followed by your agent. And I don't know why, because sometimes it may be due to the limitation of the AI or it may be due to the setting of the Discord or OpenClaw. So it's very hard for a typical lawyer without any type background like me, it's very hard for me to keep it organized. And also I feel like the features of the OpenClaw aren't
Ab (04:57)
Mmm.
Mm-hmm.
Helen Fan (05:17)
comprehensive enough. For example, when you use Claude Code, you can try something like, how to say it, you can have, for example, if you want to do some high risk things, Claude Code every time will ask you before it does it. Yes, but in OpenClaw, you have to set up the rules beforehand. For example, ⁓ you have to,
Ab (05:30)
Hmm.
You can give it permissions, yeah.
Yeah.
Helen Fan (05:44)
write down a lot of rules in the soul.md or agent.md So only in this way, it can follow your security rules. So now I'm trying to, as I mentioned to you earlier, so I'm trying to migrate my OpenClaw things to Hermes Agent or to Managed Agent. So Managed Agent is a feature in Claude and Hermes Agent is another agent platform.
Ab (05:48)
Mm-hmm.
Mm-hmm.
Yeah.
Mmm.
Mmm.
Helen Fan (06:11)
got, I think, the first place in GitHub recently as an agent. I tried them both because I want to testing alternatives. And I think because
It's more like a very immature thing about agents. So you have to take additional efforts. You have to set up your guardrails carefully. And you have to supervise it every time. You cannot leave it when you sleep because it may, sometimes we call it like some prompt injection. And sometimes you may just...
Ab (06:28)
Yes.
Hmm.
Yes.
Yeah.
Helen Fan (06:49)
lost a lot of tokens in the process. So I think it's kind of very dangerous. ⁓ So I think if you want to try agents, you have to be very patient in the whole process.
Ab (06:52)
Yeah.
Hmm.
Yeah. And again, just for some listeners who may not know this, because well, for number one, all of this stuff is very bleeding edge. In fact, know, Peter who created OpenClaw, which who then got aqua hired by OpenAI, I think he still describes it as a rather technical product. And it's on purpose. It's designed to have friction in the process, because it is
very much changing by the day. Sometimes there's multiple updates each day. So there is a lot that happens. And the same thing with Hermes agent and they have different philosophies around how they work. I'm eager to hear your thoughts on that in a minute. but you get a huge benefit and you get to define and customize exactly how.
these agents should behave. you what I do like about OpenClaw is it does personify the agents. You have this heartbeat which checks in with the agent every X minutes, I think by default of 30 minutes or something, which can be changed. The agents have a soul. And by the way, these are all just marked down their text files, but you can define its characteristics as personalities, how it should behave.
⁓ And then it obviously has instructions like agent.md and skills as you might be familiar with. And I think some of those things are amazing. But to your point, the learning curve is steep, very steep because you can, and for what it's worth, the agent can help you fill that information in by talking to you and over time. But as you were saying, I have the same issue where
Helen Fan (08:37)
You know what
Ab (08:42)
The information is captured in the soul.md or elsewhere, but it just doesn't remember sometimes. It's kind of like working with an assistant who just somehow just gets amnesia one day. And they should know that. And of course there are the challenges of just monitoring the whole thing. I mean, it's a really cool project and I'm curious, what do you know today?
and day 55, 56, whatever day you might be on, that you wish you knew instead of the first week of getting started with this project.
Helen Fan (09:11)
Mm-hmm.
Yeah, so honestly, it's much harder than I expected. So that's the most honest thing I can say. As I mentioned, AI agents are fun deeply unsafe at the same time. So that was my earliest realization. You think you give them an instruction and they will follow it? They don't. And I posted about this on LinkedIn. So I call it the Security Guardrail framework.
just generated by myself based on my real experiment. Because sometimes I spend a lot of weekends just fixing bugs in my agent rules. You close one loophole and another one just opens. And also that is, think, especially hard with multi-agent because single agent is manageable. But when two agents are talking to each other, you need to constantly adjust to reduce friction between them.
and you have to make sure they stop at the right moment instead of arguing forever. And those problems come from three directions at once because the platform itself has limitations. The models have their own issues and the chat software, Discord, Telegram, isn't really designed to support this kind of use case. So I think that's the first point I want to mention and we just discussed a little bit on it. So for the second one,
Ab (10:39)
Yeah.
Helen Fan (10:40)
I would say the agent workflow tools that exist today are good for basic tasks. I use Claude Code and the co-work, checking emails, scheduling, and a simple task. So those work well. But multi-agent, agent arguing with each other or agents spawning sub-agents. So that's a completely different level.
⁓ There's no off-the-shelf solution for that yet. I had to figure it out myself. ⁓ as I just mentioned, when you have agent-to-agent communication, sometimes you can see the blind spots AI may have because the agent shows their confusion on one point. And also, sometimes you can assign different roles on the agent. For example, you can have
Ab (11:11)
Yeah.
Helen Fan (11:32)
someone as a business advisor and the other one as the legal advisor. So in this way, you can see the common ground between the legal and business perspective. I think it's very interesting.
Ab (11:35)
Mm-hmm.
Yeah.
Mmm.
Do you think as a law firm, your OpenClaw LLP as a law firm, it's functioning well? Because I mean, no business and no organization is perfect either. You still have friction, you still have people arguing with each other. Probably you don't get an argument report at the end of it. But do you think as an experiment is going well, is teaching you what the dynamics are between the agents pretty well?
Helen Fan (12:13)
Yeah, I think the point is that I deliberately make them to, you know, arguing. So, yeah, it's my purpose. It's not like, because only in this way you can see, I think the value of the multi-agent platforms. So when you have different agents, you can give them different roles. And I just want to supplement what I said.
Ab (12:21)
Mmm.
Mmm.
Helen Fan (12:39)
I'm not a litigator, but if I'm a litigator, maybe I can ask their agent to pretend they are my opposite party. So in this way, I can stress test my arguments. I didn't try it, but I think maybe it's valuable. And also, if you want to take it as a real law firm, would say definitely it doesn't work very well because you have too much chaos happening every day.
But if you take it as an experiment, I think it's very good because recently I realized very few firms, they are actually deploying agents in their workflows. ⁓ And I think in this way, I can kind of encourage more legal professionals to use AI, to use agents.
Ab (13:21)
Right.
Mmm. Yeah.
Yeah.
Helen Fan (13:30)
because agents
is very different from chatbots, as you mentioned. So chatbots is something you ask a question and you get the answer. It's a very straightforward, but for agents, it can really do the work for you. And I think that's a key point of the experiment. I don't want to make it as a real law firm or something. Yeah.
Ab (13:34)
Yes.
Sure.
I agree with you. And even if you are using, let's say, the Claude's managed agent services, which is an enterprise tool, which allows you to do some of this, but with better guardrails to be able to host it in a managed environment. I think a lot of what you shared are extremely valuable learnings because again, agentic design and all of these principles on how do you
Helen Fan (13:58)
Hmm.
Ab (14:15)
manage multiple agents, collaborating, communicating? How do you think about all of the potential conflicts, the loopholes and the edge cases? I think that's still key. So I do think from an experiment point of view, you know, whether the outcome and the outcome isn't to create a law firm, but you learn so much about how do you think about from a first principles bit, because I can tell you before I started going on the open claw rabbit hole, I was not thinking about creating so many
Helen Fan (14:17)
you
Ab (14:45)
concrete guardrails, but in my OpenClaw environment, A, it's in a separate machine, but they also have their own emails, for example, and they have their own accounts because I want to be very cognizant of the risk, right? To your point, it could be prompt injection, it could be something else. So you need to figure out how do we safeguard this, much like you would with a user in a real environment. Yeah.
Helen Fan (15:08)
Mm hmm. Yeah, and
I just want to add one point. So maybe we can discuss it later. So it's about my a native law from ⁓ transformation. And I also call it a legal AI value stack. So when I designed this experiment, I just want to kind of testify my theory. So the first step is about raw AI capability.
Ab (15:18)
Hmm.
Hmm.
Helen Fan (15:33)
The second step is AI plus workflow. And the third is AI plus data, property data. And the fourth is infra. And the fifth step is ⁓ hybrid model, the AI native law firm. So currently, I'm at step two and step three. So AI agent plus workflow and agent plus property data.
Ab (15:52)
Hmm.
Helen Fan (15:57)
So I ask the agents to run an argument report and also a debrief report every time I finish talking with them. So in this way, I can help the agent to generate more skills and playbooks based on my chatting. So I think that's the initial point of the self-learning data layer, which I think is very important in the transformation.
Ab (16:03)
Hmm.
Mm.
Yeah, and I think we'll have to talk a bit more about that because I do like the framing of the value stack. And I'll link to your post about that. But yeah, the five levels are the raw AI capability to start at level one, the AI plus workflow, the UI layer at level two, the proprietary data at level three, the infra layer, the system of record, as you called it.
Helen Fan (16:48)
Yeah.
Ab (16:49)
level four and then the hybrid model, right? So that's when you're not selling software to law firms. You're kind of becoming, you're creating a software layer itself. I like the framing because I like a good maturity framework because it helps people assess where they might be. What was the driver for coming up with this?
Helen Fan (16:51)
Yeah.
Ab (17:07)
Was it just so you can measure where you are in your journey or to help others sort of think about where they are?
Helen Fan (17:13)
I think there are two sides of it. The first side is about how to develop your legal AI company or how to, as a legal AI company, how to navigate when you face the ⁓ competition from the Claude or Gemini And the other side is about how to transform your ⁓ original legal work.
how to transform ⁓ a traditional law firm or traditional legal department. So these two questions, I'm thinking about them every day. And I think that's a motivation for me to, you know, just very naturally to have this idea. And I just want to share it with other people to get the feedback and to test it, whether it works. And yeah, so I think the turning point
It's about recent anthropics move in legal AI world. We can see it very clearly about a Claude legal plugin, right? And the Claude's words plugin. And recently we just got the first ever Claude legal webinar. So I think these things have just a...
make me to think deeply about this topic. And I think that's also why people kind of, you know, paid much attention on my framework because they found out, wow, Helen thing is right. So because it's kind of explains is a chaos around the Claude versus Harvey or other legal companies. People kind of very anxious, but they don't know how to start with them.
Ab (18:40)
Hmm.
lol
Mmm.
Yeah.
Helen Fan (18:56)
⁓ So for lawyers, they may think, okay, now I have Claude how can I use it to transform my work? But for legal AI founders, they are very anxious because they kind of feel like lose their jobs and lose their business. So my framework is just to help both of them, both of the groups to navigate. ⁓
Ab (18:56)
Yeah.
Hmm.
Hmm.
Where
where do think most AI companies are across the five stages of the value stack? Do you think they're still at that early stage or they're starting to move into the later stages using proprietary data and moving into what you call sort of the
Helen Fan (19:26)
Mm-hmm.
Yes. Yeah, yeah,
Ab (19:36)
the AGI kind of pathway.
Helen Fan (19:38)
Yeah, I think most illegal AI companies now are still at level one or level two. And I just wanted to clarify the level one and level two. So for level one, I call it raw AI capability. This is any way you are using AI as a chatbot. You ask it a question and have it, know, draft something. Even if you have a rag layer on top.
even if there is a plugin, if you are still in a chat window and AI is just responding to your prompts, that's level one. For level two, I call it AI plus workflow. And I think that's where AI kind of connects to how lawyers work, not a chat window. I see many companies, they have the word plugin features, which I think kind of very typical tools for contract review now.
And I think most of the companies is here level one or never two companies because but I think, know,
Ab (20:37)
And I think
is that where you think, you and you you're talking about the Claude, Claude webinar, and obviously Claude had their legal plugin that was released, which is essentially a collection of skills, more than anything else. Do you think that's where those companies who, I guess, for lack of a better word, and I don't, you I don't know how many of them, how many companies think that having a word add in is a moat itself, but just having a plugin doesn't give you
an automatic sort of dominion over the vertical. But I can imagine that's where it starts becoming a bit complicated, right? Where you're thinking, well, should I use this or I already pay for the Claude subscription for my organization or perplexity or OpenAI? Why not just use that to do some of those basic things? And I don't mean sort of things like drafting full documents, but if you're running a red line, for example.
Helen Fan (21:08)
Yeah, of course.
Ab (21:33)
That could now potentially happen comparing it to your playbook
Helen Fan (21:37)
Yeah, so that's why everyone wants to be at level three or level four in my framework. So for level three, it's a proprietary data. But here's the problem I have to say, like level three requires scale. So the kind of proprietary data I'm talking about is learning which class types cause the most friction in a given industry or which risk flags predict the deal failures.
Ab (21:41)
Hmm.
Helen Fan (22:04)
So that intelligence only emerges across hundreds or thousands of clients. So no single client's usage can produce it. So you need a lot of users and you need those users to trust you with their core data, their contracts, their negotiation patterns or whatever. Then your platforms learns from all that and turns into intelligence that benefits everyone. I have to clarify because sometimes there, you can correct me if I'm...
Ab (22:19)
Hmm.
Helen Fan (22:31)
if I'm wrong. So because sometimes you even you choose opt out, it means that your ⁓ AI platforms cannot train your data. But sometimes your behaviors data cannot still be counted into the training data. Is that right?
Ab (22:32)
Mm.
Hmm.
Yeah.
So I'm not an expert in this, but that's my understanding, right? Because I think there two different things. There is learning from your inputs, and then there's learning from your behaviors. Because behavior are usage signals, right? Which buttons are people pressing? What are they doing? And most companies, traditional SaaS and AI native, do look at that. Because you want to be mindful of, how do users utilize our platform? Are there features that are just not used?
Helen Fan (23:04)
Yeah.
Hmm.
Ab (23:17)
features that we think should be highlighted more, but because they're buried in a complicated flow and a hogmany of ⁓ massive menus, no one gets to them. And I do think that it's a signal, whereas the other information, is people's prompt and so on, which for most, they will opt out of. And a lot of the enterprise plans means you're automatically opting out of that. I think most CISOs want that.
Helen Fan (23:20)
Thank
Ab (23:43)
So I think you don't learn from what people are doing with your tool, but you do learn from how they're using the tool.
Helen Fan (23:49)
Yeah, yeah, totally agree. So yeah, so I also wrote ⁓ in my original article about level four, because when we talk about level four, we will think about Clio and a FileVine But we also wondering if every legal AI company wants to have the strongest defensibility, why not just become a Clio and a FileVine? But the truth is that it's too hard for you to become them.
Ab (23:51)
Yeah.
Hmm.
Hmm.
Yeah.
Helen Fan (24:17)
because I think most of them kind of have a very long term history and they have built up the credibility in their customers for many years. So for example, I know ⁓ Harvey has kind of a feature called Shared ⁓ Workspace. So you can upload to your firm some materials there and also share the materials with your clients. So it's...
Ab (24:22)
Mm.
Helen Fan (24:42)
can become ⁓ another version of iManage. clients are, think, some partners, just talk about this feature with me. And I think they cannot fully rely on this feature to store their data because they still trust iManage, they still trust Clio, but they cannot fully trust Harvey, which is very new, very new to them.
Ab (24:56)
right.
Hmm.
Helen Fan (25:07)
It's just like a VC backed startups and you know, just running for several years. So it's harder for law firms to make the determination like I just want to switch to Harvey or switch to other platforms. They still kind of get stuck with Clio or Filevine because they trust them. It's very simple.
Ab (25:12)
Mmm.
Hmm
Yeah, but I
think it's a more complicated problem than that because I think trust is one factor. And the other factor is if you use that, the workspaces example with a DMS, a DMS is more than just a storage, right? Although that's kind of at the surface at this most basic level, it's a document management system. There is a lot of complexity built in, including
Helen Fan (25:34)
Mm-hmm.
Mm-hmm.
Ab (25:54)
connections and integrations into other systems. It becomes your single source of truth. And even if the sharing capabilities handled by another company, let's say a Harvey in that instance, that's only for one scenario. How does that scale across the firm? And does it give you all of the controls that you need for a lot of other things? And this is why, whether we look at a Harvey Legora and other tools,
Helen Fan (26:06)
Mm-hmm.
Okay.
Hmm.
Ab (26:21)
they do connect with the document management systems because sure, you can store documents, but it's how you're storing documents, how they get exposed, who they get exposed to, retention periods, and all of the hundreds of additional questions that someone at one point had to think through. And you always end up with this complex scenario of the tool over here can do X, this tool can do Y, and then there is this sort of overlap.
Helen Fan (26:41)
and
Mm-hmm.
Ab (26:51)
that makes things complicated in the Venn diagram of things. And it just, if you can make that overlap bigger, then you may be able to get with one. Otherwise you have to have a multi-product strategy. also, mean, given how quickly things are changing. And I know you mentioned MCPs in your article as well. You know, that helps to change some of these things because at least my belief, I don't know what you think is
Helen Fan (27:13)
Yeah.
Ab (27:20)
As a technology buyer, you are absolutely thinking about all of these things. As a user, I never want the user to think about this. To the user, it should never matter where the documents are stored. You have a surface through which you're interacting to get your desired outcome. And that surface ideally connects to all the places that wherever the information is stored, so it's as frictionless for you as possible.
Helen Fan (27:24)
Mm-hmm.
Ab (27:50)
And I think we're slowly moving into that world through MCPs and previously APIs. But it does require a lot of cross-organizational collaboration and even internal collaboration, which is always the hard part.
Helen Fan (27:56)
Mm-hmm.
That makes a lot of sense. yeah, that's also a good topic we can discuss together. So how do you see the dynamic between Anthropic and Harvey?
Ab (28:11)
Yeah.
Yeah, well,
I won't single out Harvey, but I will say Anthropic and Legal Tech vendors. Let's just, we'll broaden the scope. I think it's tough, right? So, and I think Anthropic is probably the right player to pick here because them versus a OpenAI, they tend to focusing a bit more on enterprise and certainly enterprise knowledge work. I think both of them are equally focused on coding use cases, but...
I think it's challenging. You have these companies that have the raw intelligence. Most legal tech companies are not developing their own frontier models, which means they're simply plugging into and doing some fine tuning and other behind the scenes stuff on top of ⁓ OpenAI, on top of Anthropic, and other LLMs. So I guess the question becomes,
Helen Fan (29:06)
Hmm.
Ab (29:14)
What's the moat? on one side, and I'm just talking out loud, on one side, you'll have an audience saying, well, this is just a wrapper. I am of the view that there is nothing wrong with a wrapper. A wrapper requires you to have deep understanding of your user base. And there's nothing wrong with that. So many tools forever have been just wrappers built on top of things. And that's fine.
because that's how it should be, right? Like a generic company is not focusing on a niche use case. And it doesn't make sense for you to have to reinvent the wheel when you can. That's the whole point of, you know, libraries in the past or using APIs and now using, you know, LLMs as your intelligence layer. But I would absolutely think that as a company that's using
Anthropic or OpenAI as my intelligence layer, I must, I have to sleep with one eye open to say, is today the day they're going to release product X that competes with what we do? And it's not to say it's going to take all of your market share in one day. I mean, although we've seen examples of that, right? You wrote about when, when Anthropic released a number of tools before, how it's had a shift in the public markets.
They recently released Claude Design, which had an impact on Figma, for example. Not that it's a direct competitor, but slowly but surely it takes some market share away. So it's a tricky situation. think really the key question becomes as someone who wants to get that technology, you and I are not good examples because I think we're too much in the bubble.
Helen Fan (31:01)
Uh-huh.
Ab (31:01)
both as
users and also just people who are looking into this. I think the normal user in a law firm, particularly, or in-house, I'm not sure how much they're thinking about this. But I can imagine an in-house procurement person saying, look, we have perplexity, or we have Claude for everyone. Do we really need to look at this other thing? And it creates another point of procurement fiction before a decision can be made. So at the very...
Helen Fan (31:19)
Mm-hmm.
Ab (31:28)
least it caused delays in adoption and procurement of technology.
Helen Fan (31:34)
Yeah, so I just want to share my thoughts on it. So here's my prediction and I may be wrong. so if Harvey doesn't get acquired by Anthropic or OpenAI, its biggest revenue source may eventually be running as a plugin inside their platforms.
Ab (31:46)
Yeah.
Mmm.
Hmm. Do you think that I mean, both of them have also raised and I know you talk about Harvey, I talk about both Harvey Lagora, they're my sort of anthropic open AI, sort of equivalent in legal tech world. But they also have both raised a lot of money. So that's a big acquisition, right. So as you think about the multiples required to return what probably the you know, what the VCs need and the investors demand.
Helen Fan (32:11)
Yeah.
Mm-hmm.
Ab (32:23)
And of course, and I don't know, I'm happy to have either of the founders on the the pod to talk about this. But I wonder, you know, I imagine they have the conviction that they are going to be bigger companies than what a acquisition would get them today. Right. Generally, you know, there are both very mission driven companies and they want to, suspect, fulfill the mission.
Helen Fan (32:27)
Hahaha!
Hmm.
Yeah. ⁓
Ab (32:49)
But do you think an acquisition of one or the other happens in the next 12 months?
Helen Fan (32:55)
maybe not very near future, I think maybe, yeah, eventually, because I see last week, you know, SpaceX announced a deal to potentially acquire cursor. Yeah, so, so, and before that, I think OpenAI tried to buy cursor also. So every major AI lab wants to own the full stack, they have the motivation to acquire these companies.
Ab (32:59)
Hmm, eventual. Yeah.
Mm-hmm. Yeah.
Yeah. Yes. Yeah.
Mm-hmm.
Helen Fan (33:22)
And
also I would say ⁓ something like the app store. ⁓ Maybe one day, in iPhone, we have app stores and maybe one day Claude can become the app store. So all.
Ab (33:26)
Mm-hmm.
Mm-hmm.
Yeah. Well, OpenAI has tried that, right? They have
their sort of plugin store. also had their, they did have an app store, I guess, once upon a time, and they had GPTs and all of these other things. But I do think as you're talking, I think probably the biggest signal is, you know, both companies and Anthropic and OpenAI have now developed and are creating their own applications. No longer is it just API and chatbot. You know, you've got Codex.
Helen Fan (33:41)
Yeah. Yeah.
Mm-hmm.
Ab (34:03)
And then you've got cowork and just the general Anthropic app. So they are starting to take space on the user's desktop too, to sort of start competing with, okay, this is what we want you to open to start your day.
Helen Fan (34:13)
Mm.
Yeah, so that might be very dangerous. Much more dangerous than for the companies now. So scary.
Ab (34:22)
Yeah.
You
know, so, but it's a, you know, it's a fast moving ⁓ field. Okay. Do you mind if we change gears just a little bit? I'm conscious of time. So you wear multiple hats and I know you also run these, these sort of seminars, round tables in Silicon Valley. So you get the opportunity to speak to lots and lots of people. I wonder what's your...
Helen Fan (34:31)
Thanks.
Ab (34:48)
What's your take on the state of Big Law versus AI native firms? So my friend and colleague, Matt, he tracks AI native firms through this site, which we'll link in the comments. And I see more and more firms being added there on a sort daily, weekly basis. So certainly there are more and more AI native firms. YC before sort of put in a request for startups for people to explicitly build competitors to Big Law?
So we have sort of signals from, you know, one of the biggest incubators. Yeah, what do you think? Where are we with AI? And do you think it's real competition from AI native firms towards the big law firms?
Helen Fan (35:31)
Yeah, so like a few, a few days ago, I just attended a Stanford Future Law Week. I was one of the guest speakers in Dazza's workshop. So I saw this dynamic between the Big Law and AI Native Law firms playing out in real time, like Big Law firms and AI Latest startups. So it's very, I it became clear that
They are not competing with each other, but they are each racing to fix their own weakness. So, AI-native law firms have speed. They can build fast, deploy fast, and iterate fast. But they lack deep legal expertise. They struggle with complex cases, and clients don't trust them yet. So trust takes years to build. And also, on the other hand, big law has trust.
Ab (36:23)
Hmm.
Helen Fan (36:28)
They have the talents, pipelines. They have the relationships with very good clients. But they adopt AI extremely slowly. here's the uncomfortable truth. Their billable hour model works against efficiency. And if AI makes you faster, you bill fewer hours. So that's a structural problem. And also, I mentioned in my post before, so
Ab (36:31)
Hmm.
Helen Fan (36:55)
In big law you have a very long term pilot phases. You have a very complex piloting issues. have a committee supervising all the things. So for example, maybe this year we have Opus 4.7 and we raised this issue. We wanted to buy it and we just raised this issue to the committee.
Ab (36:59)
Hmm.
Helen Fan (37:24)
Maybe finally they approve it. But at that point, the latest model becomes another one. So AI changes very fast. And I don't think the efficiency in big law firms can guarantee anything. So I think both sides are racing,
Ab (37:35)
Hmm.
Yeah.
Helen Fan (37:43)
They are not, they're still in a different level. But maybe one day.
Ab (37:43)
Yeah.
Do you think if
we project ahead, there is a convergence in the services they can offer so they do become competitors in a sense? I know they're not today, but do you think that's trajectory they're on to become competitors, or it's much more the AI native firms are taking on the kind of work and the scope of work which maybe the big law firms either don't want or aren't suitable for?
Helen Fan (38:15)
Yeah, that's a good question. And I think it's also very inspiring. Because one thing I know some big law firms are acquiring ⁓ very leading legal tech companies now. ⁓ And on the other hand, I can see some AI native law firms like Norm AI, they hired many very great talents from big law firms. So maybe I think in both ways, they can
Ab (38:26)
Hmm.
Hmm.
Helen Fan (38:42)
kind of as you said, converge. ⁓ But yeah, I think it's not a common thing, at least for this stage, because it's hard to tell whether it is AI native law firms, because there are many buzzwords around this concept. So this is an AI native law firm. Maybe it's just to use AI every day. Lawyers use AI every day.
Ab (38:45)
Hmm, maybe. Yeah, yeah.
Yeah.
Mm.
Helen Fan (39:08)
And they call it a native law firms.
Ab (39:10)
Yeah, well, I guess from your perspective, what do you think is the one of the characteristics of a native law firm then?
Helen Fan (39:19)
I think at least you should ask AI to help you to organize the client intakes part. And also it can help you generate all the first draft of the memos or other deliverables your clients ask for. So that's the basic idea. ⁓ That means you heavily relying on
Ab (39:26)
Mm-hmm.
Mmm.
Hmm.
Helen Fan (39:44)
AI to control the your all the works on your desk. And also agents. So if AI native law firms doesn't use AI agents now, I would say it's not AI native law firms. Yeah, because agent AI is already a very I think that if you still use it as a chatbot,
Ab (39:45)
Mmm.
Hmm.
Hmm. Yeah
Helen Fan (40:08)
It's not a good example to say, okay, I'm leading in this area.
Ab (40:19)
it sounds like there's a lot of a lot of open question, but things are moving fast. We'll see if there's acquisitions of the big players by the Frontier Labs at some point in the future. And we'll see what happens with AI native law firms. And before we wrap up, I have a
a couple of quickfire questions. You don't have to give me quickfire answers, but I would love to get your take on this. I guess to start with, what's one AI tool that you would miss if it disappeared, if you couldn't use it anymore?
Helen Fan (40:46)
Claude, definitely. ⁓
Ab (40:48)
Claude, yeah. Are
you a cowork user? Are you a Claude code user or all of the above? OK.
Helen Fan (40:52)
Cowork and Claude code both. Yeah, so
Claude is my thinking partner, obviously. So not just for legal work, for everything, strategy, writing, debugging agent rules, all the things.
Ab (41:01)
Hmm. Hmm.
All the things Okay,
cool. What task do you think will be fully automated within two years that you do today? I specific tasks for you that you do day by day, maybe in like six months or a year or two years, you will just be looking back and thinking, my God, I can't believe I used to do that myself.
Helen Fan (41:27)
Yeah, I think standard contract review, because now people can use Legora or Harvey to do the contract review, not every, but you can not use it like you cannot see it in every law firm or in every legal department. But I think within two years, most law firms and most legal departments, when they do the standard contract review, they will use AI.
Ab (41:35)
Mm-hmm.
Helen Fan (41:51)
It can do the review against their own playbooks. So the AI reads the contract and compels it to your standards. And the flags were different. And the human just looks at the flags. So that should be fully automated.
Ab (41:59)
Hmm.
Yeah, perfect. What's one of the biggest mistakes you think teams make when they're thinking about implementing AI or leveraging AI day by day?
Helen Fan (42:20)
I think the biggest mistake is buying tools without redesigning your workflows, kind of in between the level one and level two. Because sometimes people just bought AI onto the old process and wonder why it doesn't transform anything.
That's step one thinking, while we needs step two thinking.
Ab (42:45)
Yeah. Yeah.
Yeah. Yeah. Couldn't agree with you more. Could not agree with you more. And then last question.
How do you think about where to begin? So I guess maybe think about it this way. If someone who has done nothing with AI
Helen Fan (43:00)
I
Mm-hmm.
Ab (43:04)
And they're like, wow, after listening to this, Helen has been able to do so much with AI. I need to start. Where can they begin? Yeah.
Helen Fan (43:11)
Oh, okay, I got it.
So for those people, have to say, they need to change their mindset about AI. Because I think for this group of people, they usually think AI is completely unsafe. If you actually read Anthropics privacy terms or open AI privacy terms, can see there is a provision that can help you opt out of data training.
Ab (43:25)
Hmm.
Helen Fan (43:39)
⁓ So this compliance question is more nuanced than people think. So avoiding AI entirely is not a safety strategy. It's just a falling behind. And that's one thing I have to tell those people because they didn't try it before. So when you have their basic confidence about the security side, then I think the next step is to do something like to incorporate AI into your workflow.
And the starting point is something I call it orchestration layer. For example, in a law firm, you can set up your agent to help you to do all the client intake works, just to, you know, share the context between you and your clients to help you to filter the cases that you think that is really valuable. And when you are in a legal department,
Ab (44:11)
Mm-hmm.
Helen Fan (44:30)
you also need to build up the orchestration layer as a first step. Why? Because only in this way you can connect the business side and the legal departments. So whenever your business team have some issues, they can discuss their questions with agents first. And if that is a low risk task, agents can directly deliver the results. But if that is kind of middle,
Ab (44:49)
Yeah.
Helen Fan (44:58)
like the high risk task, then the agent can connect to someone in the legal department and just move things forward. So I would say orchestration layer is really important in both law firms and in legal departments. And in my framework, you can see we can go through from the workflow and to the data and to the infra.
Ab (45:05)
Hmm.
Hmm.
Helen Fan (45:25)
So the orchestration layer is the proper type of the infra. So all the data will kind of lie in your infra, lie in your orchestration layer. And gradually, as many, years maybe, and then you will have a very mature infra that based on your orchestration layer. So that's my thinking about it.
Ab (45:30)
Hmm.
Yeah, perfect. Well, Helen, thanks so much for coming on the podcast. was wonderful speaking to you. I look forward to following the rest of your experiments and see how they go. link to your LinkedIn and a lot of your writings for people to read up and connect with you. Well, yeah, thanks so much for coming on. It was a great conversation.
Helen Fan (46:06)
Thank you, I really enjoy it.
Ab (46:08)
Thanks.