Loading...
#206 - How lawyers are using AI in 2026
This episode explores how artificial intelligence is disrupting the legal industry and why lawyers have both an opportunity and an obligation to stay ahead of these trends. Through conversations with lawyers, general counsels, and legal-tech leaders, it examines how AI is changing legal workflows, reducing repetitive work, and creating more time for judgment, strategy, and client service. The episode also challenges lawyers to think beyond augmentation and consider how AI might fundamentally reinvent legal work. Finally, it addresses the risks of overreliance on AI, including hallucinations and mistakes in legal documents.
I really think it's a unique moment.
I mean, professionally, I don't know if we're ever going to see this kind of disruption—and I mean that in the positive way—ever coming before in the legal industry. And certainly, going forward, if you don't get in front of a lot of these trends, unfortunately, you might get left behind.
So I not only see it as an opportunity, but also as an obligation.
That was Sujit Rahman, whose name was among the most impactful general counsels of 2025, for his work as chief legal officer of TMM Labs.
As I mentioned, there's an obligation that we have to stay ahead of AI as a legal profession.
You don't need me or anyone else to tell you that AI is important for lawyers or that it makes things more efficient. That's not what this podcast is about.
Sujit is one of the eight lawyers and CEOs who I spoke with for this podcast, which is about how those at the cutting edge of legal AI are getting ahead and how you can get the most out of it in 2026.
Welcome to the Law of Code podcast. I'm Jacob Robinson, and by the end of this episode, I promise that you'll understand how the lawyers at the cutting edge are actually using AI today.
It should be the most helpful legal podcast on this subject that exists. That's the standard with every episode that I publish. And if it's not, let me know how I can improve and I'll make it better for you.
That's all I care about here: you. I want this podcast to be helpful for you, and not just because you're so good-looking, I swear.
This podcast has four main themes when it comes to AI.
We'll start by talking about mindset shifts. We'll talk about how lawyers are using AI tools today, with real-world examples, risks to be mindful of, and how these tools actually work, as well as how you can better leverage AI.
Some of the tools are so helpful that I hope you'll actually pause the podcast at some points to start using them yourself.
I think it's an amazing time to be a lawyer. That was a sentiment shared by everyone who I interviewed, including Scott Stevenson, the CEO of legal AI contract-drafting tool Spellbook.
Here's Scott on why he thinks it's the best time to be a lawyer.
I think it's genuinely the best time to be a lawyer that there's ever been, for a few reasons.
>
One of the reasons I co-founded Spellbook was because almost every lawyer I knew privately admitted to me over a beer that practicing law wasn't quite what I expected. The amount of time I'm spending in front of Microsoft Word, especially on the transactional commercial side.
>
You know, a number of very passionate lawyers had become disillusioned with the job. And then, after AI came along broadly, so many of these people were like, “I actually love my job now. This is a lifesaver. I'm actually really excited for the rest of my career because I'm not going to have to spend hours and hours a day just copying and pasting in Word.”
>
Now, there's so much about practicing law that I love, and same with everyone I interviewed here. They see AI as something that doesn't mean less work. It means fewer mistakes, faster client communications, and a more consistent work product.
Unfortunately, what comes to mind when you first hear the words “AI” and “lawyers,” for many, is hallucinations. We're a risk-sensitive profession, after all.
Here's A.C. Paracero, the founder and managing partner at Raines LLP, to explain what most lawyers think about when it comes to AI.
The first thing to say is, when lawyers think about AI, the image that I think most readily comes to mind for folks are these scare stories about litigators getting sanctioned for citing hallucinated cases, including the famous example where Sullivan & Cromwell apparently did this in a bankruptcy case.
>
On the one hand, I find myself sort of frustrated by that because this so clearly, to me, is human error and not AI error. How could you ever submit a brief to a federal judge without having checked the citations?
>
Whether or not the AI is drafting it, even if you have a junior associate drafting it, you don't just pass off the work product.
>
But the deeper story is that because AI output looks so polished and convincing, it presents a real temptation to turn your brain off and abdicate your mental labor to the tool.
>
To more directly answer the question, I think the right way of using AI is the exact opposite of that. It's to do all of the mental struggle and work and cognitive labor—everything you would have done beforehand—and then use that as the input in the AI.
>
That makes the tool incredibly powerful. And I think if you're doing it right, you should actually be more exhausted at the end of the day, having used AI all day, than you did before, because you're spending more time on the quality of the thinking and less time on formatting citations or all the stuff that can be more easily automated away.
A.C. made some great points. I think the most important is how a lawyer's time is spent differently when they're using a tool like AI.
More time is spent on judgment, strategy, and thinking, and less on the repetitive tasks that can be automated with AI.
If you use AI or you have clients who do, you should probably know about the presenting sponsor of today's episode, Altitude.
Not only does Altitude offer a command center for payments, you can use it to manage entities across jurisdictions from one spot instead of a bunch of different accounts everywhere.
They also offer unlimited credit cards and elevated cash back for AI spend.
So, if you're listening to this and you have clients that use digital assets and want to transact not only in stablecoins but also in fiat on one place, it's worth taking a few minutes to visit altitude.xyz/law to learn more.
If you are using AI in your legal work, like almost everyone is these days, it's important to remember why.
A more efficient system, right? That's sort of talked about as the goal, but I don't think that's it. A more efficient system is the byproduct. The goal should always be the same: to better serve your clients.
David Wang, the chief innovation officer at the law firm Fenwick, explains:
It's undeniable that AI improves efficiency. I always think of the efficiency improvement as a byproduct.
>
What you really want to do is use AI to improve the quality of our services to our customer. And if that is happening, I kind of don't care what else is happening. Does that make sense?
>
Because there's this fear of lawyers out there, and that's like, “Oh, if I use AI to do X, Y, and Z, it's going to take away some of my billable hours.”
>
Very much like the fear that I had as an associate that automation was going to take away. But if you throw all that away and truly have the courage to embrace that, ultimately, both from a professional perspective—what is the soul of the lawyer? What is the most important thing that lawyers do? Serve a client.
>
And from a commercial perspective, as a law business, what's the most important thing to do? Serve your clients, right?
>
And then you throw everything else away and you just say, “Whatever it's going to take, and whatever I can do to make a better-quality service for my clients, is what we're going to do.”
>
Then you arrive at the inescapable conclusion that you must use AI in this way.
>
And by the way, it also tells you when using AI doesn't make sense, right? Because if you try to use AI and it doesn't improve the quality of your services, then you don't need to use it. You can do it manually.
>
It's not like AI is the answer to everything, right? So that's, to me, the most important pillar of it.
It's such a good reminder that what we're doing here is practicing law. We're not trying to be as quick as possible. We're trying to do the best work we can.
The same thing applies to any lawyer, obviously, whether in private practice or in-house. Quality is what people want. AI doesn't change that, but it does change how you get there.
Many people see AI as augmenting our work, something we turn to when necessary. But not Molly Abraham.
Molly is the general counsel at Coinbase and an engineer-turned-lawyer who has a pretty forward-looking approach to AI.
Here's my back-and-forth with Molly on why many people are thinking about AI all wrong.
Those who are thinking about AI as augmenting their current workflows and efficiencies are thinking about it all wrong, in my opinion.
>
I think it really has an opportunity to upend what we're doing. But one of the things that I found is that myself and my team had all of these great ideas of how we would automate certain workflows, how we could shift people from working on something more high-value and strategic, and automate the majority of their role.
>
And we were getting stuck. The number of times I've been in a terminal window and reading code and trying to get my token key to work—and our amazing engineers at Coinbase are always willing to hop on the boat with me—but I was hitting a lot of brick walls.
>
So I don't have an expectation that my team all of a sudden all becomes engineers. Just like we go to outside counsel for specialized expertise, this is my specialized expertise: how to build these agents for envisioning.
>
I think it is critical, though, that it is embedded in the legal team. And here's why: I think that for an engineer to be able to be successful in supporting a legal team, you also have to deeply understand how the legal team works and what their workflow is like.
>
What is our red line? The first time I showed an engineer, “I'm called. I'll hate it. Can you help me figure out how to automate something around redlines?” Just the concept blew them away. It was unfamiliar.
>
And so I think it's critical that not only do you have someone with the technical capability to build these agents, but they're embedded in working with the teams so that they understand how they work. Otherwise, we won't reinvent how we work in the first place.
And I want to touch on one thing you mentioned there: the idea that these won't necessarily just augment, but they'll upend how we do legal work.
The word “augment” has come up so many times throughout this episode, talking with lawyers about how they're using AI. Everyone likes to say that they're using it to augment their work.
Walk me through why you see that as being the wrong way to look at it.
I think we can really revisit whether the workflows make sense in the first place and which workflows need to be done by humans.
>
Which ones, when done by humans, are just creating coordination tasks? I think we can't underestimate what we can replace with AI.
>
I'll give you an example. My superpower, I would have told you six months ago, was drafting. I can draft for complex legal topics for a non-legal audience and make it make sense.
>
That's the thing that I was known for, and I would always have a heavy pen in editing my team's work.
>
And I asked myself, “Can one of these agents do this?” I said, “I thought they can't. There's no way. This is my true, sort of unique superpower.”
>
And I haven't edited any of my writing now, or edited my team's work. I've gotten unsolicited feedback from other cross-functional partners: “Wow, Molly's writing has really improved.”
>
Don't talk. It's the AI. And that's okay.
>
I have to get comfortable with the idea that something I considered an N-of-1 attribute of my own, something I could automate, I could train, I could improve.
>
And what I've filled my space with is more coaching of the team, more strategic and innovative lawyering, all of these different things.
>
Having an open mind, I think, is the first step in the process.
I think, as lawyers—or in any profession—we'd all like to believe all AI can do is augment what we're doing, because it's self-preservation.
But I think that if you really, really want self-preservation, you have to have a more open mind now, because someone else will. They'll find a way to really, really upend workflows.
What's the biggest opportunity in AI that most lawyers aren't thinking about or talking about enough? What would you say stands out to you there?
I think that some people still have a tendency to think the same way I did about my own writing: “Oh, AI can't replace my tasks.”
>
I think there's a little bit of self-preservation there. The biggest opportunity is for those who ask themselves, “How can I replace 99% of what I do with AI?”
>
And then have trust that we have so much else to do, so much more opportunity.
>
A lot of people have gotten over that very little bit of fear. And I think those who look for a way to truly replace themselves with AI are going to be those to have the biggest and brightest careers.
When you're saying that, it just made me a little nervous, right? The idea of replacing yourself with AI makes people a little uncomfortable, but you need to lean into that because other people are.
You're freeing up your time to do the things I think you're uniquely suited to do. And one of those things that you've mentioned—strategy—I think decision-making is only going to grow in importance now because of this extra leverage that we have with AI.
How do you sharpen your decision-making skills?
I think it goes back to something you just said, which is effectively making yourself dispensable.
>
Part of the reason I have an easy time thinking this way is it's how I've always managed my team. The more I think there are a lot of lawyers who might view themselves as, “I need to be indispensable. Everything will fall apart without me.”
>
I have the opposite view. I am routinely trying to find ways to make myself dispensable and to say, “If I got hit by a bus or won the lottery tomorrow, the team would operate just fine.”
>
I think you have to think the same way with AI. And when you do, what has inevitably happened to me is, as I continue to say, “I don't need to go to this particular critical meeting. Someone on my team is now fully capable of doing that.”
>
And that's a good thing. That's because I've created a phenomenal bunch of leaders. I'm also going to create a phenomenal bunch of agents.
>
And that is only going to give me more space to make decisions, to make those strategic calls, and to help piece everything together in a way that's really impactful.
>
So I do think that dispensability mindset is critical as a leader generally and also with respect to AI.
AI can sort of hurt the ego a little bit. It can be difficult to admit that a large language model can do something as well as you can.
When I was working as a securities and corporate lawyer, I would pride myself on not making mistakes. And, of course, I never made a single one. I never sent an email that said, “See attached,” without attaching the relevant document.
Or maybe I did. I forgot. Anyways, I doubt you've ever made any mistakes in your life, but other lawyers definitely do.
The team at Spellbook actually dug into contracts on EDGAR, the SEC's filing system, and found mistakes in a ridiculous number of documents. The percentages might be shocking. Maybe to non-lawyers, I think lawyers will understand having read all these documents, but the percentages are quite high.
Here's Scott Stephenson, the CEO of Spellbook, on their findings.
We launched a report a few weeks ago called “Human Solutions AI,” where we scanned thousands of agreements that were submitted to the SEC's EDGAR database, and we used Spellbook to determine how many of these had objective mistakes in them.
>
We're not looking at whether something was negotiated poorly or anything subjective. We're looking at, objectively, how many of these contracts have bad section references or conflicting terms that can't fully reconcile with each other and things like that.
>
We ran about 50,000 to 60,000 pages of contracts through Spellbook from EDGAR, and surprisingly, we found that 60% of contracts filed in EDGAR had mistakes.
His audio cut out a bit at the end there, but Scott said 60% of contracts filed in EDGAR had objective mistakes—mistakes that were caught by Spellbook's AI tool.
That's about 33,000 contracts with mistakes.
Contract drafting is one way that Molly Abraham is using AI in her role as general counsel. She's also built a writing agent that drafts and refines internal legal documents, translating legal analysis into something a business audience—the people she works with in-house—can actually use.
The result is that instead of documents coming to her from, say, the business...
So, for a team or a counterpart, she doesn't have to spend as much time redlining everything that comes toward her desk. Here's Molly to explain how she's built that.
Molly: So I created and trained an agent that understands Frame's legal lens: how to translate a legally drafted document to a more business-facing audience. Most of my initial drafts now come from this agent. My team runs all of their drafts through the agent.
I went from spending a huge chunk of my time reviewing documents before they would go to a more senior audience to, at most, reviewing them and maybe having one redline, which is something I never would have predicted a year ago.
Host: Another example, just on that point: how does it work? Let's say you get the contract and get the redline back from the counterparty. Do you send that to the agent? Does the agent automatically pull it? What does the workflow look like there?
Molly: Sure. So this agent is really about writing an internal document as opposed to modifying a contract. Modifying the contract and dealing with redlines turns out to be quite technically complex, so once I hire this person, I'll get back to you on that one.
I have an agent that I built in LibreChat, and one of the things that we've done really well at Coinbase is that it will actually route to the LLM model that is most efficient in terms of total usage. That's part of how we're able to increase our AI usage without increasing our AI spend at the same rate.
I will give it ideas. I'll say, “I really need to draft a document about these particular topics. Here's two or three other notes that happen to have been created, or check this Slack channel,” and it will pull all of it together. It will draft the document.
Alternatively, I can give it a draft document and say, “Please clean this up to make it sound more like my voice.”
One of my strongest tenets as a writer is Mark Twain's: “If I had had time, I would have been a shorter letter.”
I used to get a lot of long documents from my team, and I think it's really important to keep it short and concise. I think that actually takes way more time.
I'll get the output as a Google Doc. I might have one or two more tweaks. I might go back and forth with the agent and say, “I didn't want to frame the recommendation this way. Please edit accordingly,” or, “Please additionally pull information from this document.”
It's able to read both across Google Docs and across Slack if I point it in that direction, draft the entire document, and it's really fast.
What Molly just described—a system where you can bring in an external document and compare it to a playbook or a baseline—is one of the most consistent uses I heard from lawyers who are maximizing their use of AI.
They're starting with a precedent or a guidelines document on a repetitive document and having AI draft based on that. It's one of the highest-benefit use cases of AI today.
Now, why is AI so good at building off a precedent as opposed to starting from scratch?
That's because of how an LLM works. Let's take three minutes now to walk through how most LLMs work.
Here's Samson Azar, a partner at K. H. O. L. E., Gordon & Reindel, who I'm thrilled to share is a sponsor of the Law of Code podcast. Instead of responding on the team, Samson, Louis Cohen, and the team are actually joining the episodes. In this clip from an interview we did, Sam makes a really important point that is often forgotten, which is about how the popular LLM chatbots and legal AI tools actually work today.
Sam: As we discuss it, this conversation is really about LLM-based solutions. Those are basically predicting what's the most likely response or what's the most likely answer based on this information.
So that is not really a judgment thing. That's really just collating data and presenting it.
When people say “AI” in 2026, they almost always mean large language models, or LLMs. That's what Claude and ChatGPT are. There are other kinds of AI, but the current wave, and what we're mostly talking about today, are LLMs.
As Samson said, LLMs are prediction machines. Here's how they work in five steps.
First, you give it a prompt. That's whatever you type in the box. It could be, “Draft an NDA between two software companies.” That's your prompt. You hit Enter.
What happens next is the model breaks your prompt into tokens. The rough rule of thumb is that one token is about three-quarters of a word in English. So a short word like “draft,” “the,” or “contract” is usually one token. A longer, less common word, like “indemnification,” might get split into three or four tokens.
So you've got your prompt. You've typed in, “Draft an NDA between two software companies.” That goes to the model, which now breaks it into tokens.
So you've input your prompt, and the model's broken it into tokens. What happens now?
The next step—and this is the third step—is that the model predicts the next token.
Based on your prompt, plus everything it learned during training—and we'll talk about what training means in a bit—the model runs the math on every possible next token and picks one.
The model will look at your prompt and say, “Okay, given everything I've seen before, what's the most likely token to come next?”
So if your prompt is, “Draft an NDA between two software companies,” it will start that based on how it's answered similar questions before, find prior examples, and use those to predict what happens next.
That's why these tools can produce something called hallucinations. It sounds confident, it looks correct, but it's wrong.
It's because models aren't picking the next true token. They're not exercising judgment on what's real. They're predicting what the next token should be.
Once they predict what token should come next, they add that token to the text and then begin predicting the next one, and the next one, and the next one. Every token it generates becomes part of what it looks at for the next prediction.
So the model doesn't sit back, think about your whole question, plan a response, and then write it like we do. If someone asks me a question, I think about what my answer is going to be and then I begin speaking.
What the model does is guess one token at a time, over and over and over. So it'll look back at the whole thing, and it'll add a new word.
“Given the lawyer filed the motion,” what's the most likely next token? Maybe it picks “to.” So now the text reads, “The lawyer filed the motion to.”
It keeps going, and this is the final step. It keeps going until it decides it's done. The model has learned when to stop, usually when it hits a natural end or a length limit.
So what looks like a thoughtful, complete response is really just thousands of these one-token predictions stitched together in real time.
That's why the model is not actually thinking ahead, like you and I do when we plan a response. It's predicting one token at a time.
So that's the brief outline, at a very high level, of how an LLM works.
I mentioned training earlier, and training is such a big part of this because there's a reason answers actually feel so coherent, and a reason answers have been getting better since ChatGPT launched a few years ago.
That answer is training.
Training is the reason that AI feels like it's thinking, even though it is just predicting. Training is how a model learns to be better at predicting. It's why ChatGPT has improved so much since it was first released a few years ago.
The model starts out with billions of internal settings called weights, which you can think of as tuning knobs. At the start, every one of those knobs is set to a random number.
It gets fed an enormous pile of text: books, court decisions, websites, code—basically the entire searchable internet, and a lot of data.
What it does to train is look at a chunk of text with the next word hidden, and it tries to predict what that next word will be. It checks the answer. It does that trillions of times.
Every time it guesses wrong, it slightly adjusts its weights—its tuning knobs—not to make that type of mistake again. Every time it guesses right, it reinforces whatever it just did.
So we've got LLMs. They're prediction machines. They get trained by tuning their answers compared to what the actual answer was.
After trillions of times, the LLMs start getting real patterns. They start getting smarter. That's what the weights are. They're just the accumulated result of trillions of tiny adjustments.
Two models can feel similar in the same chat box, but the quality of what comes out is entirely dependent on three things: first, what they were trained on; second, how much they were trained; and third, how their weights got tuned along the way.
That's why customized tools, like the ones Molly and others are building, can outperform a general-purpose chatbot. It really depends on the underlying training that's happened.
Of course, no lawyer wants to be, or should be, the one who lets their work product or sensitive client information be used by LLMs for training. That could mean the LLM could regurgitate the sensitive information it's been trained on to someone who prompts for it.
That's why companies like Harvey, Luminance, and Spellbook don't train their models on lawyers' data.
Here's a snippet from my conversation with Scott Stephenson, CEO of Spellbook, where he explained how that works.
Host: I know you've negotiated agreements with OpenAI and Anthropic for zero data retention. So that means customer data included in requests and responses with the LLMs only exists in memory to process the request. They're not trained on that afterward.
You also use cloud providers with data centers located in the U.S. for storing and processing customer data. You're going to have a trust center on your website, which I thought was a fantastic idea.
When people go in and see all these vendors and see who you're working with, what are some questions you get about security from lawyers, and how does your team think about that?
Scott: I think the number one thing that lawyers care about is that we're not training AI models on their data, because once you train models on private legal data, the model could spit that client data back out again. That's like the worst possible thing that could happen.
The problem with training is actually that it's not a very good approach. We did training and fine-tuning of AI models in the early days when we launched, back in 2022 and 2023.
There are a bunch of reasons why training on legal data is not good, but one of them is that it can allow that data to be regurgitated in all sorts of weird ways.
That's the main thing customers want to know. We do not train on any customer data whatsoever, and that protects that risk.
We're HIPAA compliant, so we sell to hospitals and healthcare providers. We sell to one of the largest consumer banks in the world. So we have a very high bar for security and privacy.
People want to see those controls for SOC 2 and HIPAA, standard controls put in place. They want to understand where their data goes and doesn't.
Generally, no one wants the data to be sent to the model providers for them to use it for training, which we ensure through zero-data-retention agreements with all of our model providers.
Host: Scott, just on that point, can I ask you about that? I don't know the details of how that works inside a node. What would happen if I were using Spellbook, and the text is sent to Anthropic, OpenAI, or one of those third-party LLM providers? Is that data processed through the LLM and then deleted after the answer is provided?
Scott: Yep, that's right. So it's just a round trip, and then they do not store the data for any other purpose after that.
They do have a window of time where the data may be privately stored for moderation purposes, like detecting abuse and people doing things that they shouldn't. But other than that, the data is deleted and not used for training.
Now, if you're not using a tool like Spellbook but a general-purpose model like Claude or ChatGPT, and you're not turning off the training data, what happens with the information you input?
Even if you are turning off the training tab, you're not necessarily always protected. Eric Dylis is an attorney and programmer. He explains how there are other risk factors for sensitive client information—not just model training—that lawyers should be aware of.
Eric: You start with the example of what happens when you submit a prompt with your data.
When you submit a prompt, or you start a workflow with AI that uses a model that's not on your computer—that is remotely hosted on a server not in your room, which is just about everybody—the way that these AI service providers, OpenAI and Anthropic, now really hammer the fact that they're not using your inputs to train their models. They're great for privacy for that reason.
That is just a sliver of the problem.
Another big part of the problem is that whenever you submit that prompt or start that session, your input is getting logged on their servers.
Separately from the model, it's not getting trained into the model; it's getting logged on their infrastructure for a minimum of 30 days for safety, quality assurance, and legal requirements that are intentionally very opaque.
They're storing your inputs, they're storing your information, your metadata, anything and everything that is attached or accessible to what you provide or expose.
It's maybe not the same exact risk factor as model training because models will have a more difficult time regurgitating it. But your information is still on someone else's server, subject to potential exploits and all kinds of potential issues.
That logging also occurs on the way out from the model. So it's not just your submission; it's also what the model provides back to you.
As Eric outlined there, sensitive client information is always at risk if you're using a server located elsewhere or not on your laptop, since that information must travel on the internet rails and is capable of being intercepted or stored.
Eric created CamoText. It's an offline desktop application that detects and redacts personally identifiable information, or PII, and sensitive text before sending data to AI models.
It's pretty cool. It's worth checking out. It's called CamoText.
Another thing you should know about Eric is that I'm in a fantasy football league with him, and I won the league in 2025. Sorry, I just couldn't help myself.
CamoText is a fully offline app that works on any laptop. Here's Eric explaining how it works.
Eric: So you load a file, document, text, or a folder full of them. Plan to use the settings you want or not, hit “Anonymize,” and a combination of custom algorithms, NLP, and so forth detect sensitive terms, names, and common patterns like emails, GPS coordinates, and e-signature IDs—anything that is reasonably identifying—and replace it with a labeled and hashed placeholder in new output text.
What that output allows you to do is supply it as context or as part of a prompt to an AI model for analysis, summarization, and all the common use cases that lawyers and other professionals use AI for. It preserves confidentiality in that way.
Now, Eric explained how he ensures client confidentiality when using AI. That's something that's a huge risk for lawyers.
We've seen cases where clients have put privileged information into a consumer AI tool and lost the ability to protect it in court. Well, at least we've seen it when non-lawyers have.
The clearest example is *United States v. Heppner*, a case that was decided in the Southern District of New York in February 2026. I covered that decision at length in episode 186 of *Law of Code* with Mike Katz.
Basically, Judge Rakoff held that communications between the defendant and Claude were protected by neither the attorney-client privilege nor the work-product doctrine, because privilege never was created in the first place.
Conversations with somebody who's not a lawyer cannot be attorney-client privileged. Sometimes there are other forms of privilege that apply if they're a doctor, for example, but Claude is not a doctor.
However, in a different case, *Warner v. Gibb Bralco*, ChatGPT outputs were actually protected as attorney work product, because a self-represented litigant had prepared them using a public AI chatbot, and the court found that work-product protection applied.
So it is unclear where the line is, and that makes it more important than anything to minimize the risks here.
I asked Molly Abraham, General Counsel at Coinbase, more about the legal risks of AI specifically and how she thinks about guardrails when it comes to using AI with her team.
Host: How do you think about implementing guardrails when it comes to using AI with your team?
Molly: This is great and such a perfect opportunity to talk about how I started to think about AI.
Molly: Because I'll be honest, when I started to think about this two years ago, I started with the risks. I didn't yet see the opportunity, and it's evolved my thinking in terms of what every legal leader needs to take into account.
So the first and foremost is: you have to protect the company.
And part of the guardrails here include—I don't know if you remember—a couple years ago, a major service provider all of a sudden sent out an email notifying folks, “Oh, your data may be used in LLM training.” That was a big red flag. It was confidential company data that they were going to use to train their own LLM to be able to effectively use that with clients.
And the very first thing that we did is we scrubbed across all of our key vendor agreements to ensure that we were protecting the company and protecting our data.
I do think, first and foremost, understanding how the company's data is being used is really critical. The legal leader at any particular company right now needs to start with: How do I protect the company with respect to AI?
From there, you can go to how to enable the company with respect to AI. How do we have—for example, we've created a fast-pass procurement process for key AI tools, because we can't go through a super-lengthy process by the time you do, we're the new model.
And so how can we ensure that all of the important guardrails are taking place, but we can move faster? Because it's important to enable the company.
How do you scale your team with AI, which we talked about? And then how do you be a super user yourself?
But it starts with that protection piece. It's where I started a couple of years ago, and I think it's something that every GC should be looking out for.
Host: And given all the work that you're doing now, I think it's fair to say that you're at the cutting edge of legal. Where are the guardrails you're thinking about now?
Molly: I think it's a couple of fold.
One is how other companies could potentially use your data. Terms get updated all the time. There are LLM providers, and you really have to keep an eye out to make sure that your data is not being shared, not being stored, and not being used as a training model if you don't intend it to be.
So I would say that's the external data piece, and there's the internal data piece.
Because as the company adopts more AI, if you don't have good data controls internally, all of a sudden employees might be able to access all sorts of information if you haven't set it up correctly.
For example, it was very important to us as a public company that, in order to enable my team, who's working on an M&A deal, to use AI, anything related to M&A needs to be very carefully cabined.
And so, again, it's less about AI and more about the framework that you set up. Data integrity is just so critical.
For us to feel comfortable rolling out broad AI tools for a variety of teams, including those working on really sensitive information—whether that's M&A, whether that's employee-related personal information, whether it's comp data—we had to have the right guardrails and restrictions set up such that we felt comfortable with those teams using those tools and there not being data leakage.
So I think it's really, really about data framework.
Host: I want to highlight the last thing Molly said there: data framework, data controls, integrity. That's something that's top of mind for most lawyers today, as it should be. It's certainly top of mind for David Wang.
He's the chief innovation officer at Cooley. David's job is literally to determine the firm's technology strategy, so there aren't many people more qualified to talk about this.
He has a similar emphasis as Molly: getting the data sorted before and while using AI. Here's David to talk about how he stays on top of all the data that a law firm like Cooley has.
David Wang: In the data domain, that means really getting ourselves organized, right?
For example, we talk about our AI principles, right? But don't you hear out there, “What's behind those principles?” So, a tremendous amount of work.
Really get on top of our data so that we can truly follow our clients' instructions in terms of how they want their data to be handled.
But what they don't want to happen to their data requires massive investment infrastructure in our data lake, right, in our systems and processes, such that the work that we're doing on the software, on the AI side, can be supported and accelerated.
Host: Like you, like me, David understands the importance of data when it comes to implementing AI. There's an overwhelming amount of data to navigate, and many lawyers and law firms aren't thinking about all this. But they should be, just like they should be thinking about how to best leverage AI in the legal practice.
And that's something that we're going to talk about now.
I want to start with the most basic way that most of us interact with LLMs: via the chatbot window. Zac Shapiro, whose post on the cloud-native law firm was read nearly eight million times, explains the first lesson he gives lawyers on how to better interact with and better leverage AI today.
Zac Shapiro: It starts with the prompt—not treating it like a Google search, but instead as an associate or a genie who needs really clear instructions.
Anyone who's seen the movie *Obsession* recently—a good horror movie—knows how important it is to give clear instructions.
When people put prompts in AI, the natural thing to do is to write one or two sentences, which is wrong. The right mental model for how to start, instead of a Google search, is how you would brief an associate across the table from you on an assignment that you want them to do.
It's not quite the same as talking to an associate, but the starting point is: Okay, imagine you're going to delegate this piece of work to a human. What would you say to them?
And now imagine if, instead of a human, this was a genie from a fairy tale. In all of the genie stories, the problem with the genie is they take your wish too literally.
You wish for a million dollars and it falls out of the sky and hits you on the head. You wish to be the smartest person on earth and everyone else disappears.
Imagine that anything you're not super-clear and specific about is going to be used against you. Then do that briefing again for the genie, where you're forcing yourself to sit there for at least a couple of minutes at a time, filling all of that time with more and more of what ultimately ends up being context for the LLM that allows you to get more elite output.
I think that's really step one in getting good at talking to the AI.
Host: Now, as Zac said, we need to be more patient when we use AI. And I'm as guilty as anyone for typing in something like, “Milk safe, drink two days past expiry.”
Prompts need to be focused. They need to be tailored and convey everything necessary for the LLM to provide the right answer.
Or do they?
What was most telling to me across all the conversations I had is that everyone has a different approach, and that seemed to work okay for them. I think what does matter, though, regardless of the specifics of how you prompt, is the context you give the LLM.
You can either give it that context in the prompt or actually train the model itself. Here's Zac again with more on where an LLM will pull an answer from and how to train it to get you better answers. It's where he says the edge is for lawyers to differentiate themselves.
Zac Shapiro: The edge is in the instruction that you put into the AI about what output you want.
In order to do that well, just going back to what an LLM is: It's trained on the corpus of the internet. If you are vague at all, if you leave any room for interpretation, the LLM is going to fill that gap with the median of the internet, which is what we've all come to recognize as AI slop.
It's the em dashes and the training specificity, I think, from Reddit, which it overly weights as a source.
And so, in order to not get that, you need to very specifically describe exactly what you want out of the model, which requires you first to be able to imagine it.
So I'm spending more of my time thinking about, okay, what really is the right work product for this client? And then how do I express that in a way that leaves absolutely no room for vagueness or ambiguity that can be weaponized against me by the LLM?
That working-net muscle has also, I think, made me a better thinker and a better lawyer.
People say that the best way to learn is to teach, and working with AI is just constantly teaching the model all these unstated assumptions that go into your work that you don't think about. Now it has to come to the fore, and you think about it and you have to explain it to the model.
Host: As Zac just explained, we need to be thinking about the unstated assumptions when we use AI. That can be basic things, like which party you're representing, or something more complicated, like the negotiating power or leverage you have when making edits.
Here's Aaron Kelly, who's the general counsel at EdenNode and the founder of Little Labs, to add more on the ways you can teach the LLMs you work with to improve their outputs.
Host: How do you think about teaching? And then what do you mean when you say teaching the model? What does that look like in practice?
Aaron Kelly: Well, there's two ways.
You can do it through prompt engineering. You can just say, “Hey, here's a document that we have that we use for reviewing these agreements, and I want you to take it and apply the playbook to it. And here's the context.”
Now, you can also train the model to do that too, meaning it's already inherent to it, so you don't have the prompt every time. That's a little bit more difficult. There's ways to do that through what they call fine-tuning.
That's really what I focus on: teaching the model, “Hey, this is what is good, this is what is bad, based on my preference.”
It's really hard. That's probably why we don't see these general legal models out there.
And going back to the Harvey question—why maybe what's working on good work or good net—it's because so much of what we do is subjective.
One lawyer could argue that this is a good change versus another, which we've all gotten—the red lines back from a lawyer that has changed the font size, the font type, and it's a preference thing.
Well, again, you can teach those things, but it's like going back to why you do it. Because you want the model to behave like you want it to.
You're not replacing your judgment. You're basically giving yourself a partner that can help you in looking at an issue's body. You want to have something that can look at every single angle when you're looking at a deal or a contract.
That's what a good model should do. And if you teach it how to do that, you're going to have a lot better time.
Host: So Aaron talked about training the model yourself. You can also get models like Claude to train themselves.
Here's Zac Shapiro again to explain how he improves his AI with the Skills feature on Claude, which lets you package a set of instructions, reference materials, and examples that Claude will load whenever a relevant task comes up.
He also explains how he leverages different models to check each other before he even sees a document.
Zac Shapiro: Claude can get pretty good at checking its own work. Skills are really helpful for that.
Knowing how to word skills and word prompts for self-review is really important for that.
Opus especially has a personality. I know that sounds weird, but it is very true, and I've tested it rigorously. The way that you phrase things is really important in terms of how thorough it's going to be at checking its own work.
Being earnest and appealing to, I guess what you would call emotions, for whatever reason tends to work better than just getting a list of things for it to detect that I can randomly decide to ignore.
Even more powerful than that, I found that different frontier models check each other's work more effectively than any of them check their own work.
And so you can really scale that up. There are complicated drafting assignments where I'll have Claude work do—I’ll pass it back and forth between Claude and ChatGPT eight or ten times before I ever take a look at it and wait for the different models to agree that this is ready for my review.
Host: There are all sorts of tips and tricks to have AI check its own work. That's a significant amount of leverage—a game-changing amount of leverage—which is why lawyers like Zac are benefiting from it.
Another lawyer who runs his own firm, Michael Showalter, founder of Showalter PLLC, has created an AI-native appellate and complex litigation boutique firm.
Michael's current AI stack includes Claude and Claude Cowork, and I think it's obvious he's really getting the most out of it. He's written a lot of viewed articles that typically and previously had taken 150 hours in just 15 hours.
This is one you should definitely listen closely to for an idea or two that you can implement in your practice. Here's a clip from my conversation with Michael.
Host: Could you walk through your AI stack?
Michael Showalter: Yeah, I use Claude Cowork and ChatGPT Codex. My primary workflow is through Claude Cowork. It's quite remarkable.
There's the chatbot, which most people are familiar with, which is quite powerful. Cowork is a desktop app for those who are not familiar with it that plugs into my laptop.
It's connected to the folders that I granted access to. It's connected to my Chrome browser. It's connected to my calendar. It's connected to my email.
So it is constantly doing agentic things, such as drafting a first draft of an email that I might go and review. It might be letting me know that I had an email come through that I haven't responded to.
Every day, it is going onto my Chrome browser and doing internet research or filling out a form for me to review.
When I draft briefs in the chatbot, I walk Claude in the chatbot through the legal arguments that we want to make. And once it's ready to draft the first draft, it just drops a Word document to a folder that I then go mark up with line edits and comments, like I would within an associate.
Then I go back to the chatbot and I say to Claude, “My revisions are in the document,” and then it versions up and implements my revisions. And I go back to my folder on my desktop.
It's quite a remarkable process.
It's also just constantly scanning. These are called scheduled tasks.
I have Claude scanning for mentions of my large yard of cool that we were just talking about. I have Claude scanning for business development opportunities, for any developments at the state bar associations on their postures toward AI, and all sorts of other things.
As far as other tools, I use a couple of legal research connectors. One of them is called mid-page. The other's called ding-duff. And they are connected to my Claude.
So if I have a legal research question, I will tell Claude, “Here's my legal research question. What's the answer?” And then it will use those tools to research the case law and present me the answer as if it's a junior associate.
I have not entered a Boolean search like I used to into Westlaw with the ANDs and ORs while I have operated my practice. I've been talking to Claude like I would talk to a junior associate, and then it sends me the app—having research done.
Host: I really wanted Michael to walk through his AI stack because I think it's such a great snapshot of what an ultra-forward AI-based legal practice actually looks like.
He's running Claude Cowork as his primary tool. It's plugged into files, browsers, calendar, and email, with scheduled tasks scanning for business development opportunities in the background, along with legal research connectors and linked case law.
He mentioned two tools that you might not be familiar with: mid-page and ding-duff. These are free Model Context Protocol connectors and PC connectors, which are basically bridges that link Claude directly to primary legal research databases, so it can look up real case law in real time rather than relying on what happened to be in its training data.
Remember, these are just prediction machines. So if you go with something like those tools, they can actually check what exists rather than try to predict what exists.
While we're on the topic of what's under the hood, I thought this was a good moment to quickly talk about open-source LLMs.
The things in the topic are warning us about are saying it's a bit dangerous, and we've seen things with the U.S. government maybe wanting to stop open-source LLMs.
There's a lot of talk about them now, but also a lot of confusion about what open-source LLMs actually means. Here's Aaron Kelly to explain what open source actually means and why it's kind of a misnomer when it comes to LLMs.
Aaron Kelly: Open source is kind of like a marketing term here when it comes to LLMs. It's not so much open source as it is open weights.
The weights are what the model has been built and trained on. It's everything that the model knows, basically.
And I guess the biggest thing I can say is…
Erin: What’s open weights is you can change the weights. You can customize the weights. It can be called post-training. It’s the recipe. Open weights are like the recipe for how the LLM was trained and how it operates. That’s the best way I can put it.
Host: Now, let’s pause on Erin’s point for a second. A lot of people have told me that they pause this podcast often. These are dense episodes, but I appreciate you sticking with me because I think this is so important.
When people say “open source” and they mean “open weights,” the weights are those tuning knobs we talked about earlier—the ones that help improve the results and make them more accurate.
Now, lawyers like Michael and Zach are using frontier models like Claude in the AI stack. I asked Michael how long he thought it would take a lawyer who’s familiar with Claude to set up a system that made a dramatic difference in their efficiency and legal practice.
His answer surprised me. It was less than I expected. I think it’ll be less than you expected.
Mike: It would take probably half an hour to have the tools set up in the way that—the infrastructure set up in a way that would dramatically, off the bat, increase your productivity and allow you to improve the quality and reduce the time it takes to do your legal work.
The skill that it takes to actually guide Claude through the process is something that has to be developed and learned. Especially for most lawyers, they’re a little bit hesitant about using AI.
One thing that trips a lot of people up is they’ll get some output from the AI that they don’t like for whatever reason. It’s very rarely hallucinating in the strict sense. It’s been months since I’ve seen Claude give me a case that just didn’t exist. That just doesn’t happen anymore.
But it will still say things that are wrong or unintelligent, and that happens quite frequently. For many lawyers, they get these responses from Claude and they think, “This is a tool that I cannot trust,” and therefore, “I cannot use it.”
And they are right that they cannot trust it in the sense that they must verify the accuracy of its assertions and the soundness of its reasoning. But the conclusion does not follow. Does that mean that they can’t use Claude?
The key is not getting flustered by some of the output that is unintelligent or wrong. I am constantly telling my LLMs, “That makes no sense. Please correct this. Please fix this. Why did you say this? This is a problem. We need to fix this right now so it never happens again in the future.”
If you do that, it’s quite good at correcting, and you actually don’t need to drop the thing. You just need to work with it to get the result that you want.
Mike: Getting to the place where you can leverage AI to the extent that I am is not a 30-minute project, for sure—at least not for most people.
But just getting the infrastructure set up, getting Claude connected to the folders that you want to connect it into—for example, if someone wanted to have Claude help them brainstorm a legal issue, creating a folder of 10 or 12 articles on the subject and getting Claude connected to that folder and reading the thing—that would take Claude about five minutes, and the setup time is very brief.
Host: It’s shorter than we think. That initial setup—connecting Claude to folders, wiring up basic infrastructure—if that’s a 30-minute exercise, I think that puts you a step forward toward the thing that actually separates lawyers who get 10x results.
Another alternative is to use a ready-made system. If you don’t want to do as much of this yourself, I spoke to the CEO of Spellbook, Scott Stephenson. Their program, I think, is $299 a month for lawyers.
They have an AI copilot and contract-review system designed specifically for transactional lawyers and legal teams. It operates as an add-in that sits directly inside Microsoft Word.
Here’s Scott Stephenson to explain what makes the tool unique.
Scott Stephenson: I think this is something that’s unique to an AI tool like Spellbook as opposed to general LLMs like Claude or ChatGPT.
The problem we’ve had in legal, and one of the problems in legal and transactional work, is that this data has been so opaque and very asymmetrical. Google is going to have all the data in the world, or a big law firm is going to have tons of data. But if you’re a small guy, you’re going to be at a data disadvantage, and you’re not going to know whether you’re getting screwed over or not.
And so we thought it just would democratize negotiation. Making this data more transparent would be more fair and more efficient.
Around the models, I think the other challenge is that if you’re just using ChatGPT to negotiate an agreement, it will tell you, “Oh, this is not standard,” or, “Oh, this is standard.” But what’s that really based on?
It’s mainly based on public company contracts and data that was in the dataset on the internet. And that’s not necessarily—if you’re a private company and you’re a healthcare company, is that really what’s market? Is what ChatGPT is telling you is market really standard? It doesn’t really have the knowledge to know that.
So we think grounding AI in this data, and then being able to filter based on the jurisdiction, the industry, and the deal size, is really important to get accurate results.
Host: So, as Scott just said, Spellbook’s market feature democratizes real deal data, so you’re not just trusting ChatGPT’s guess or prediction based on whatever public contracts happen to be in its training set.
I asked him a follow-up question, though: How do they actually find these market contracts to determine what’s market? What about private-party data?
Here’s the back-and-forth I had with Scott to better understand how they determine what’s market at Spellbook.
Scott Stephenson: First off, what we capture is aggregate statistical data. The market data that we store, at the end of the day, is: This is the average price per square foot for a commercial lease in New York City, or these are the average late-payment terms for a SaaS agreement in California.
One of the secrets to what we do is we store this really high-level statistical information, which has no PII associated with it. This allows a lot of our customers to be comfortable with a give-to-get model.
In our give-to-get model, lawyers basically say, “Hey, we will provide the statistical mathematical data, and in exchange, we’ll get access.”
That is one way that it works.
Some customers will pay extra. They’ll say, “You know what? We don’t really want to contribute to the pool. We just want access to the pool,” and we’ll charge those customers.
And then, for customers with really big contract flows—customers the size of Google—we’ll actually allow siloed data. That’s something we’re launching, where you just want it to be your own data and nothing else, and we’ll be able to do that, too.
That’s how it works today.
Host: Yeah, that’s a good way to do it. I remember everyone does something similar in law school, where you can contribute notes to a group, and then you can all benefit from that. I think bringing that to the legal profession is just going to be a big win for everyone.
But I think one distinction is the quantitative data versus the qualitative data. The quality of the stock side of contracts is so important. And, like you said, you’re sort of pulling terms that relate to more—and maybe I’m misinterpreting that.
It sounded like when you referenced cost per square foot of a lease and stuff, that’s the quantitative side. What are you doing for the qualitative side?
Scott Stephenson: For the qualitative side, what we’ll do is identify the common variations of a term.
Let’s just take a simple one, like late-payment terms in a SaaS agreement. When is your service term in a SaaS agreement? Late-fee interest? There are different variations to that term that you could have.
We would look at, “Okay, here are the major variations of this term.” And then we would say, “20% of people use this variation, 30% of people use that variation,” and bucket it into the major types of terms that exist.
Host: Does that make sense?
Scott Stephenson: Yeah.
Host: So, from my understanding, if you’re looking at a term—if you’re a lawyer working on a contract and you’re working on a term such as confidentiality—you would see some of the variations of that term as you’re drafting. And then, like a software developer, you could choose which one you wanted to implement in your code, or in your case.
Scott Stephenson: Yeah, that’s right.
Host: So how Spellbook actually builds its market data is similar to how other legal AI companies do it as well.
On the quantitative side, it aggregates statistical benchmarks: average lease prices, late-payment terms. For the qualitative side, it maps out common variations of clauses.
It’s sort of like what a lot of lawyers would do when they pull out sections of contracts that they want to keep in their contract bank. A lot of lawyers I know do something like that on the commercial side.
I also spoke to the CEO of an AI-powered litigation platform, Justin McCallon. He’s building Strong Suit to transform how litigators handle legal research.
Here’s a quick outline from Justin of how that platform works. And remember, I’m not sponsored by any of these AI companies. I’m just sharing so you’re aware of what exists.
Host: How does your product work in terms of the legal research and where it’s pulling cases from? And what context window does it have? What’s going on on the back end?
Justin McCallon: Let me break those into different parts.
As far as what we’re using for data, we worked to form partnerships to cover all precedential U.S. cases. There are about 11 million of them.
We took each of those cases and said, “Okay, let’s have an agent go through the case and understand the key facts, the key holdings, what was the key analysis”—about 20 different pieces of information for each case.
We spent a lot of time really shaping that agent to be recursive and iterative in how it determined that, and it did a lot to ensure that it was going to be accurate.
For example, if we made a holding, the agent had to point back to text in the case to say, “This is what I derived this from.” Then a secondary agent would have to check and make sure that was real.
So we have this great corpus of knowledge. Because of the way we built our tool, we knew that this was going to be done in the AI era. We didn’t use the AI era, so we can use things like RAG.
Host: Sorry, Justin, what is RAG?
Justin McCallon: Retrieval-augmented generation. What we’re trying to do is basically say, “Hey, this case might have some similarities to your case.” These other cases might have other similarities or be important for other reasons—maybe they’re cited a lot, or they’re in the right courts, or whatever.
We have different ways of pulling the right cases for your matter to ensure that we’re finding the most accurate and relevant ones that are worth citing.
Then we run a lot of evaluations on that. We’ll say, “Okay, let’s go find 200 cases,” and say, “Over the back of the case, hold those out from the briefs.” What were the cases cited in those briefs?
Then we have the AI say, “Go take the same fact pattern and find the cases that you think are most relevant.” We keep optimizing until we have a lot of overlap with what’s actually submitted.
We can do things like that to make our retrieval ability very strong.
Host: Now, you don’t need to buy one of those packages. I think it’s good to know what exists. Sometimes they’ll suit what you’re looking for.
You can also build using AI yourself. Sujit Raman, the chief legal officer at TRM Labs, has built a system for monitoring regulatory updates that are relevant to his day-to-day work.
This describes how he’s created an agent that actually helps him and his team do their jobs on a daily basis.
Here’s Sujit.
Sujit Raman: My job is to protect the company. That’s a very broad remit. So there are a number of sub-outcomes within that broader outcome that I’m responsible for.
I think about other things where I need advice or updates on the regulatory structure that’s out there. My responsibility is to create an agent that will essentially give me updates at whatever cadence I want about what’s happening in the crypto industry, the cybersecurity industry, broader privacy, broader AI. There’s so much happening on a regulatory front.
I need something to pull all of that together. I don’t want to be spending money on outside counsel for this, and I want something that’s uniform—something I can look at every single week.
What do I do? I create an agent for that.
A member of my team has created an agent that tracks all the global regulatory developments in each of the areas that I talked about. And you can create skills for that, right?
I’m looking for privacy. I’m looking for litigation involving crypto or digital assets. I’m looking for AI-related regulations in Europe, in Asia, and in the United States. Break it down federally and by state.
This is all stuff you can code. It’s extraordinary because it’s the kind of thing that, even as recently as a year ago, you’d have to pay money for an outside firm to do for you. And even then, it’s still just a client alert.
Now it’s something you can run in-house using the idea of the AI cortex. So that’s one very small example.
The other one that I like to talk about because it’s really opened my eyes is the commercial side of the house. There’s so much that we have coming in in terms of NDAs, vendor agreements, reseller agreements, and so on.
Those are the kinds of things that historically take a lot of manpower. Some of it’s pretty routine, but you still have to spend time doing it. And humans make errors.
If your team has adopted an agent that’s been coded with your company’s playbook, with all the different provisions and fallback provisions and the general orientation of us as a company—some companies are large financial institutions, and we’re a smaller, early-stage, venture-backed company—you’ve got different perspectives, right? And you code all of that in.
Our contractual review SLA’s efficiency has increased by something like 1,000% using these tools. You’re not doing your job unless you’re adopting this kind of technology.
Host: Sujit has written about what he calls the “AI cortex” way of thinking about AI, and it’s a theme in how they think about it at TRM Labs.
Those are two examples of that in action. He has an agent that tracks global regulatory developments across privacy, crypto, AI, and cybersecurity, broken down by jurisdiction.
On the commercial side, they’ve got an agent that has the TRM playbook and handles NDAs, vendor agreements, and simple, repetitive contracts. He said that increased their contract-review efficiency by roughly 1,000%.
I think Molly Abraham, the general counsel at Coinbase, has seen similar benefits. She’s taken a similar approach to commercial work. She’s actually hiring a senior software engineer to join her legal team.
Yes, hiring a senior software engineer to join her legal team to expand their AI automation capabilities.
I asked her about the goal of this hire because I think that explains how people who are taking these steps are thinking about what you can do.
Here’s my back-and-forth with Molly about the highest-priority use case.
Host: The highest-priority use case for this hire is to build an end-to-end agent workflow that automates negotiations with counterparties by evaluating markups against the playbook.
Why is this the top priority as opposed to regulatory filings or many other things that I know Coinbase has going on?
Molly Abraham: This, to me, is one of the top priorities for a couple of reasons.
One, I know exactly what it is I want. I have done RFP after RFP for different vendors. I have worked with engineers internally. No one has been able to craft this to the extent where we can actually implement it.
What I’m envisioning is, we have a huge number of interested institutional clients who all want to onboard onto Coinbase because we are the most trusted platform. Part of that process for some clients does include—they’ve got a legal team, and their legal team feels the need to…
And this—to me—feels like something where we could be faster responding to our clients. We've already done the thoughtful kind of data analysis and work on the backend. I just need to somehow make it work technically, and I am stuck.
And so this is why it feels like the right place to start, because it's a defined use case. It's an immediate way in. It unlocks revenue because these are clients who want to come on to Queen's trade, and there's just so many of them that it takes time.
And so this, to me, is a great first win and something that is just right for automation. And then those lawyers who were previously doing that can focus more on important conversations with those clients—negotiations and those live conversations where an agent is not about to replace a human at any time.
Molly just explained, I think, a principle that applies to any in-house team where you have inbound and outbound sales. Having a playbook that your AI marks up against doesn't just save legal time; it actually directly unlocks revenue, and it can free up time to focus on things like judgment.
Well-trained junior lawyers can also offer similar benefits, but what does the future of AI mean for legal training as a whole, really, as juniors may no longer be a necessary cost center if AI can replace those tasks in a faster, better, and more consistent manner?
What will happen to the next generation of lawyers? I asked all my guests that question.
Here's Sujit Raman, the chief legal officer of TRM Labs, who said his biggest worry is the development—or lack thereof—of judgment.
Yeah, I mean, what I do worry about is the absence of the development of judgment, because you can talk about efficiency, you can talk about billable hours, you know, all that is what it is.
But you can be an associate who bills 2,500, 3,000 hours a year. But if all they're doing is a document review or some other kind of mechanical work, they're not actually developing as lawyers, right?
I think where the junior associate really benefits from the old system, just the kind of billable-hour system, is the presumption that they're getting exposed to not only the senior associate, but the partner—the person who's tried the case, the person who's done the investigation.
If you're sitting at their elbow and sort of watching them, you're on the client phone calls, you're in court. You know, maybe you're carrying the briefcase, but you're seeing how the more senior lawyer interacts with the bench, interacts with the jury.
If you're on a deal and you're watching the senior corporate lawyer quarterback the transaction with seven different people on the conference call, that's where you develop the judgment and the skill.
My concern about where we are in this particular moment technologically is that, with all the benefits of automation, particularly for junior lawyers, they might still miss out on that very critical—if you want to call it the apprenticeship, if you want to call it the guild kind of training.
I don't think we've found an adequate replacement for that, and there's going to be potentially a generation of young lawyers who miss out on that. And I do worry about that because I don't know if there's a good technological solution for that.
Yeah, the only solution for them, Sujit, is going to be to listen to a lot of good podcasts.
I think that's a great place to start.
Now, I couldn't help but leave my little joke in there, but that shouldn't take away from the seriousness of this discussion, because Sujit just raised what might be the most important long-term question in this whole conversation:
What happens to the development of lawyer judgment when AI takes over the tasks that used to teach them?
When I was starting out, like every other lawyer, I sort of loathed the brutal due-diligence assignments or work that I thought was mundane. But looking back, it was pretty clear to me—and a lot of lawyers told me at the time—you learn a lot from that.
You learn what to look for, when to triple-check something, and questions to ask off the bat. Those are really invaluable skills.
And I think it's important that we talk about how lawyers can actually use AI to improve their training today.
Here's David Wang, the chief innovation officer, to explain how they are training junior lawyers differently in the age of AI.
So, one of the critical components that we have in our AI training program is to say to young lawyers: What we want you to do is fight back against the cognitive surrender.
So you have a partner email you something and it's like Greek, right? You're like, “What is this person talking about?”
What you can do is take that and plug it into Legora, right, and you can get an answer. And, by the way, Legora's answer will probably be better than yours if you just came out of law school, right?
And so there you go. There's an instant answer. You're done, and shift that back to the partner. But wait—why do we have you? You know, I can read the partner. I couldn't just type that email into Legora.
So this is, ironically, the fear of junior attorneys, but it's also something that we see junior attorneys do all the time: the betrayal.
And so the training is like, okay, now instead of that, what I want you to do is take that and put it into Legora. And this time, what the prompt and what you're saying is—and this is not like prompt engineering or anything—it's really like a mindset.
It's about remembering who you are. It sounds a little Disney, really. You know, don't worry, I won't do the dance for more on it.
But, like, then take—you’re a lawyer. You're supposed to understand this, and you are in the process of training yourself.
And now, because of the acceleration of everything—and previously you were all training on the job, and, by the way, passing that on to clients, in terms of the billable, which clients don't like—instead, what you do is take that email and then say, “Hey, explain point by point every single legal doctrine that is in here, and all the things that are implied in these instructions, and give me a full list of those two things.”
And then you get a primer, like I think, right? And then now what you're going to do is read and reply.
Take that whole thing, put it back in, and ask again about every single thing that you don't understand. And then keep doing that until you understand every single thing in there. Now there's nothing else that you do not understand.
And then you take all of that—the outline of what your response is—yourself. And then you can use it to generate the response email so that it can be grammatically perfect and all of the stuff.
And then what you do is reply to the partner: “Here's what I think,” right? Because now it is what you think. It's no longer what the AI thinks, right? It's what I think.
And, by the way, here are my questions. Can you let me know if this is correct? And, by the way, can we discuss this?
That is the act of learning for a young attorney. This is what they need to be engaging in. That's the most important thing.
Now that ties into what Sujit said before about the biggest risk facing junior lawyers. And I thought David's step-by-step training method was really important, because it gives a guideline not only for junior lawyers, but also senior lawyers to think about how we're teaching the next generation of lawyers.
I also spoke with Molly Abraham, the general counsel at Queen's, about how junior lawyers will navigate the future.
Because in her job posting—which should still be live by the time you're listening to this—she was hiring for a software engineer for her legal team. She actually describes building an agent platform where today's junior lawyers become tomorrow's agent builders.
Today's junior lawyers become tomorrow's agent builders.
I asked Molly about that and what the future might look like for the next generation of lawyers.
Interviewer: So, in the job description, it describes building an agent platform where today's junior lawyers become tomorrow's agent builders. How do you see today's junior lawyers, or tomorrow's junior lawyers, navigating that future? Do they still look like lawyers, or do they look more like agent monitors and managers? What do you think that'll look like?
Molly Abraham: I think it's going to be a combination of things. And let me start with how I think we need to train junior lawyers to be ready for this moment.
I'm an alumna of the University of Chicago Law School, as is Paul, and we both have spent a lot of time with the school talking about how they're approaching AI.
And it's interesting because they're doing a combination of completely embracing AI, wanting students to do certain assignments with AI, because that's what's coming for them in the job market, while also doing certain settings where we're not going to have AI.
So, for example, that first-year curriculum in the classroom, because they want to train them in a way that they're going to be able to think like a strategic lawyer and use AI efficiently.
So it's embracing it while also learning some of the core concepts that are critical to being able to do with AI.
I think about training junior lawyers the same way. I do think we're going to have members of the legal team who are primarily agent builders or agent designers, and that's where I'm really excited for this engineer to join, to help figure out the framework.
But I can see a world where a number of my really, really talented paralegals right now could actually be designing agents because they most deeply understand their workflows, and they could be designing the agents to actually be able to replace some of those workflows.
And so, as they get better and better at that and they learn from it, they could be the next era of agent builders.
Now, that said, the thing I worry about is we can't just stop training junior folks in terms of how to do the important legal work, or eventually we will lose it over time.
And so that is the million-dollar question—or I guess trillion-dollar question, if you think about infrastructure size—that I don't have an answer to. But I think it's going to be really important for us all to figure out.
And the way I think about it, too, is that the onus is going to be more on the junior lawyers to educate themselves, because now, with AI, you have the tools to do that.
You can ask questions that typically you would have asked a senior partner, and you can get some more answers. So it almost will become, “Hey, you're on your own, and you have to keep up, and that's on you.”
The tools are there, but if you don't take advantage of them, then it's going to be really difficult.
It is that I think we have to empower folks. And it's interesting: We have a lot of self-starters on our team who have become incredibly AI-proficient, but we also owe them great training in this space.
We still owe them career development and mentoring, even if some of the substantive questions they can get a faster answer to.
And I also think this is a great opportunity to be able to be a manager earlier in your career. To say, “Okay, this is how I run a function, and it's no longer just me. I can have three agent employees as well. And here are the different tasks they're doing. Here's how I'm efficiently operating a team.”
That's going to be something where they get to fast-forward ten years in their career, which could also be really powerful for those who take advantage of it.
So, as Molly said, anyone who's at the forefront of work can figure out where the best opportunity is to automate themselves.
It's an exciting time. It's also a risky time.
And it's easy to draw parallels to another time that felt so similar: the early 2000s, when the world and the legal profession first got online with the internet.
But there are fundamental differences between the internet and AI, and it comes down to leverage.
Here's Sujit Raman with a great explanation of why the leverage of AI is many times higher than the internet.
The difference I would draw between the early days of the internet and the current moment is, on the one hand, you're absolutely right. I mean, the internet revolutionized the access to information that people had.
And when it comes to lawyers, I mean, think about something like Westlaw or LexisNexis. Just the whole legal research process became so different, right? It's so much easier. And that only made folks more efficient or better.
And yet, I mean, you look at some of the Supreme Court opinions from the 1940s, or the briefs that were filed by Thurgood Marshall or whatever, right? They were extraordinarily well-written briefs. And they did it in an age before there was an internet.
And so, in that sense, I think the profession or the capabilities transcend the time or technology.
I think where AI is different is this idea of the cortex, right? The idea of melding human capabilities with machine capabilities and thinking about what the interaction looks like between the two.
And that's where I think, for lawyers who don't get on that bandwagon, so to speak, you really will be left behind, because the capabilities of the machine are so much more supercharged than what any other human can accomplish.
So the way I like to think about how AI applies to the legal industry in particular is you're not eliminating lawyers. You're not eliminating skill sets. You're not doing anything like that.
What you're doing is actually creating leverage. One lawyer can do the work of five lawyers or ten lawyers if they know how to test the agent and create that cortex—that human-machine interaction—the right way.
And so, for people who are able to do that, you're not only a better practitioner, you're more efficient. If you're in private practice, your clients appreciate it because you're more cost-effective.
And that's the kind of melding of man and machine that I think is going to happen. And for those who aren't able to master that, unfortunately, they may be left behind.
I think that's different from the days of the internet, where you could still be a phenomenal lawyer even without Westlaw, even without LexisNexis, as long as your work product was fine.
In this day and age, I just don't think it's just about the work product. The work product will be much better, much quicker, much more considered by the people who are leveraging the cortex rather than the people who are not, if that makes sense.
Now, he's talking about the cortex and using AI not just to augment your work, but to be part of that work.
We do a similar thing with the internet, although it doesn't feel like that today. But the benefits have been huge, I'd say, for lawyers.
So, back in 1978, before the internet was widely used, before we had tools like Microsoft Word, there were about 464,000 lawyers in the United States.
Do you know how many lawyers there are today?
Well, according to the American Bar Association, we're at an all-time high. There are 1,370,000 lawyers in the United States.
So, while the internet made lawyers more efficient, that actually led to more work for lawyers. And I think the same will be true for the next generation that uses AI.
But how do you stay ahead when AI can allow a lawyer to do the work of five to ten junior associates?
Here's Sujit on how he's encouraged his team of lawyers to adopt a programmer's mindset—a builder's mindset—for their legal work.
Interviewer: What are some ways that you've found it most effective to encourage that type of building?
Sujit Raman: Well, I think part of it is the attitude. You know, it really is like folks: We're here to work together, and we're here to build careers together.
And I've been very transparent with my team that part of my job and part of my obligation is to make sure that everyone is set up for the next fifteen, twenty, twenty-five years in their career.
And I feel like I'm doing them a disservice unless I push them a little bit, right? I feel an obligation for the people on my team to make sure that they're well-positioned for the future.
So I have that conversation, and it's a very candid and transparent conversation. We're not here to scare you or put you on some kind of schedule.
It's much more about, let me find ways to empower you so that your job is more fun. The more of the stuff that can be automated is in fact automated, so you can focus on the brain work—the part where humans can actually make a difference.
And every single person on my team has embraced that, because I think it's in all of our interests, our collective self-interest, to figure that out.
So I do think tone matters. And I think really being hands-on in a constructive way—you know, during our one-on-ones, as we talk about, “What are you building? How can I be helpful? Show me what you've got.”
You know, this is something that I think you might improve on, or, “Wow, I never even thought of that. Keep doing that. That's incredible.”
I think it does take that level of hands-on, one-on-one integration. But that's the job. That's what general counsel should be doing, right?
So I find it empowering, and I hope that you will work with finding it empowering as well.
Sujit makes an important point there that we've kind of tied together.
How do you get a team of lawyers who, based on a lot of the lawyers that I know—you and I know—aren't all programmers, how do you get a team of lawyers to adopt AI, to actually want to build AI tools?
And hopefully this podcast can help with that. I think that's important because today, with these tools, one lawyer can do the work of five or ten if they know how to task an AI agent correctly. That's a gap that can't just be closed by being smart, because the distinction might be the most obvious in the one area that I think most clients use to determine whether their lawyers grade or not: speed.
How quickly you can get answers to clients and internal teammates. Imagine something that takes one lawyer 150 hours and takes you 15. Speed matters. That's why lawyers like Molly Abraham are using the extra time they've saved by using AI to build more AI.
Honestly, we're using that extra time to experiment in AI and to find other use cases, because right now that experimentation does require a pretty heavy lift to figure out, okay.
So when I wrote my agent writer that would both write documents in a kind of playlist format and also edit similar documents, it probably took me six or seven hours to build. And so it was a mix of strategic training of the agent with other work that I had done, with instructions about why I think it's important to translate a legal concept in a particular way—all of these different things.
Again, it was honestly about 50% of the time with sheer debugging.
And so right now, every incremental hour we save, I would say, on AI—or, frankly, dollars that we save, because we're taking the first draft of a memo instead of asking outside counsel to do it, and instead going to them with that draft asking for advice—we're reinvesting in AI.
The team is very focused on all of the different ways we can do it.
I think the SLAs for lawyers are about to be blown out of the water. I have got to say, because now that anyone can go and ask Gemini a question—including folks trying to ask Gemini legal questions who are not lawyers—the expectation of how quickly you can advise those clients and get back to them is really changing.
So I think we will not run out of things to do.
But my hope for the use of AI at companies is this is not about job replacement. This is about shipping faster, growing the economy more quickly, and just bringing more products to customers.
Now, it's not about job replacement. And maybe it's job displacement. Jobs probably do look a lot different in 20 years, much like they look different today than they did 20 years ago.
Here's Samson Enzer, a partner at Cahill Gordon & Reindel, and Ryan Dettmer, who I'm so grateful to have as a sponsor of *Law of Code*, with more on how the internet changed law and why we don't necessarily know what AI will do.
Many people are concerned, “Oh my, you know, jobs are going to go away because of AI.” But, you know, the computer did not lead to less jobs—or, if we want to talk about lawyers, the computer did not lead to fewer lawyers.
Westlaw, Lexis—these internet-based tools that allow you to research. You know, the judge I clerked for, when he did legal research, you went to the library, pulled a book.
When I did legal research as an associate, you would go on a computer and you could do a Boolean search and look at every case ever published.
Westlaw and Lexis, that tool, did not lead to fewer lawyers. It led to an arms race of more.
I don't know. None of us knows how the efficiency gains and productivity gains from AI will affect the legal industry or many other industries. Unclear.
Samson's point echoes what's been consistent throughout the podcast. The past tells us that big technological leaps in law and efficiency have never led to fewer lawyers. They've led to more.
And nobody actually knows how AI is going to reshape the profession. I didn't make this podcast because I know the answer to that. I made it because I don't know the answer.
But I thought that speaking to those at the cutting edge could offer us a look into what the future might hold.
There's a line I love from William Gibson. He's a science fiction writer who said, “The future is here. It's just not evenly distributed.”
That's exactly where we are with legal AI today. The lawyers you heard from are living in the future. Meanwhile, a lot of the profession is still deciding whether it's okay to use ChatGPT.
Thank you for listening all the way through. I think this is such an important topic.
My goal with this podcast is to provide the content that I wish I had when I was working as a lawyer. So if you have any feedback or topics to see, please email me at Jacob@lawofcode.fm. I'll read everything, which isn't hard because I don't really get much, but that'll change. I'm determined to make this the Acquired meets the Huberman Lab, but for law.
Thank you to the presenting sponsor of this podcast, Altitude. I wanted to work with them because I thought you should know about their product. Basically, Altitude is building the financial operating system of the future, and lawyers that I know refer their clients to Altitude on paid referrals just because it's a better solution.
Altitude has a simple strategy: solve many hard problems for businesses relating to payments. That's why, despite launching publicly in December 2025, they've already processed over $350 million in payments.
I use Altitude for my international business. If you're interested in learning more, you can visit altitude.xyz/4ds-law. You could refer your clients to them. You could also do a quick demo to meet the team and understand the product.
Again, I'm Jacob Robinson. This podcast is part of my mission to help people understand the legal layer of emerging tech so that we can live in a world with transparent and well-understood rules that are applied fairly to all.
Part of the reason I'm able to do that is thanks to the other sponsors of this episode: the Solana Policy Institute, which is advocating for transparent, decentralized technologies like Solana in Washington and around the world, and educating policymakers and doing really important work there.
So thank you to the Hyperliquid Policy Center, which is pushing for the future of on-chain finance.
I'm really grateful for both of their support, as well as Cahill, Gordon & Reindel.
Now, I'm going to launch a newsletter at lawofcode.fm, and you can sign up for free. It will be very short, probably just five bullets once a week.
Upcoming episodes, if you are still listening here, include how the law applies to AI agents, crypto and Canada, and anything that you reach out to me and say, “Jacob, this would be interesting.”
Thanks for joining me. See you next time.