“AI-native” became the hottest phrase in law this year.
In May I put 50 people in a room at Stanford to argue about what an AI-native law firm actually is, founders of VC-backed AI-native firms and legal leaders with decades behind them. That same week, I taught a session at Baker Botts for more than 30 chief compliance officers, most of them from the largest technology companies in the world. I asked how many had adopted agents in their daily work. Only three hands.
Two rooms in one week, and a question I could not put down - will we have “AI-native legal teams”? And if we will, what do they look like, and what is actually holding us back?
That question is why we went to New York. On July 17, we gathered 40+ senior in-house counsel, AI leads and senior associates from big law firms at Cooley’s Hudson Yards office in New York.
Who Was in the Room
Everyone told me a Friday night in Manhattan was a bad idea. I expected a half-empty room. Instead the room was full, several people came in from D.C. and Philadelphia just for this, and the conversation ran hot the entire time!
Half the room was General Counsel and legal ops leaders, from Amazon, American Express, BitGo, Equinor, and Foundation Medicine, among other industry leaders. A few were AI transformation leaders from top firms like Kirkland & Ellis, Paul Weiss, WilmerHale, and White & Case.
Three speakers, and I gave each of them the same mandatory question: what is an AI-native legal team, in your definition?
Stacy Lettie is Chief of Staff to the General Counsel at Organon, where the legal team runs to 200+ people globally. She practiced law for 20+ years before moving into legal operations full time, and she sits on the executive board of CLOC.
Marc Mandel is SVP and General Counsel at EXOS. Combining deep legal expertise with a technical development background, he builds his own tools and publishes them openly, and he spent the week before this event vibe coding a live demo for the room.
Matthew Wojtkowiak is Chief Legal Officer at ECI. This coming January it will be thirty years since he was admitted, and he brings some of the most forward-looking thinking on where AI goes next.
Nobody entirely agreed. But after two hours, a definition and three working paths started surfacing.
Part 1｜What AI-Native Actually Means?
AI-First, Then Human
Stacy Lettie spent the first half of her answer clearing away everything that does not count.
An AI-native legal team is not a team that bought AI tools. It is not a team with an AI committee, a pilot program, or a slide saying 70% of matters will be touched by AI by 2027. Those are things an AI-native team might have. None of them is the definition. A team that uses AI tools is AI-assisted, and at this point almost everyone is.
Her definition is structural: AI is assumed to be in the room before the work begins. Workflows, roles, metrics, and the decisions about what gets escalated to a human are designed with AI as a baseline assumption rather than an experiment.
The analogy she used is the one I keep coming back to. A native English speaker does not translate in their head before speaking. They think in English. An AI-native team thinks in AI first. The question stops being “could AI help here” and becomes “where do I still need a human in the loop.”
Three Conditions for AI-Native
She gave three conditions. An AI team is native when all three are fundamentally true.
- Data is treated as infrastructure. She said you have to invest in it before you do anything else. She called herself a huge data geek, and this is the drum she beats all the time.
- Workflows are designed around AI outputs, not AI inputs. Instead of asking what can I feed to AI, the team is trained to ask what decision needs to be made, and what does AI need to give me to inform it.
- Human judgment is reserved explicitly for what AI cannot do. Not because we are scared of AI, but because we are being deliberate about where the human in the loop is important, and what that means for every role.
An AI-native team draws those lines deliberately and revisits them on a cadence, instead of redrawing them by accident every time someone buys a new tool or a new vendor shows up with a demo. For example, in pharma legal, regulatory judgment and privilege decisions must stay with the human. Contract triage and invoice review do not have to.
This is also why I keep saying AI transformation is fundamentally a governance problem before it’s a technology problem. It does not ask what you bought. It asks whether anyone has actually made the decision. And as she put it, that decision gets made once, early, and it really shapes everything after.
AI-Adjacent vs. AI-Native
Then she turned to the reality, and said why: being honest is more valuable to you than me being inspirational.
Most legal departments have not done the redraw. They have added AI without touching anything else. The org charts look the same. The intake process is the same. The outside counsel relationships look the same. And somewhere off to the side there is a pilot program, and a champion, and a slide.
That is not AI-native. That is AI-adjacent. And the gap between those two things is exactly where most of the friction is.
Her line for what happens when you skip the redraw: if you layer AI over a broken process, you don't fix the process, you just get a faster, more confident version of the same broken process. That is not transformation, that's amplified dysfunction.
I would add one thing about why this keeps happening. Most teams get AI handed to them. Someone above them buys it, IT rolls it out, and legal is asked to find uses for a capable tool. That path only ever produces one question: what can we feed this thing? It is also why only three hands went up at my session at Baker Botts, in a room of compliance officers from companies that are not behind on technology at all.
And the image she closed on: most legal teams right now are building the plane while they are flying it. An AI-native team has already landed the plane, and is designing the next aircraft from the ground up. The difference is not skill. It is whether you were willing to stop first.
Part 2｜How to Get There? Three Places to Start
The definition tells you where you are going. It does not tell you what to do on Monday. Three things came out of the night that do.
One: Structured Intake
Of Stacy’s three conditions, this is where she said you actually start. Before anything else, you invest in clean matter data, consistent outside counsel classifications, and most importantly, structured intake.
Not because intake is interesting, but because without it, AI produces garbage. And she reframed what intake is for, and this is the part I think most teams have not caught up to.
Intake used to just be a front door. A ticket comes in and it gets routed to someone. Now there is a smaller signal underneath that senses the formal request and turns it into data. That data is informing bigger decisions. It is telling you what to insource, what outside counsel should own, and where to invest next.
Her test: if your intake exists and it is not giving you intelligence, then where is your AI strategy?
The same logic applies to metrics. Logging how many people log in is not useful data. The departments getting this right are building scorecards that look at the day-to-day legal work and tie it to business strategy and revenue enablement, not efficiency for efficiency's sake.
She was also clear about sequence, and admitted it is unpopular. She would rather see her department move slower and actually be able to trust the output, because she has watched too many colleagues roll a tool out globally and then have to pull the entire project back and redo it, because the data underneath was bad.
Two: Playbook Engine
If intake is about the data coming in, this one is about the judgment you already have and have never written down.
Marc Mandel skipped the talk track, pulling up his terminal right at the start.
He started with a screening skill his team actually runs today. When a business team wants a new AI product, they run the vendor’s terms through this skill before opening a ticket. It assumes the company has no ability to negotiate those terms, so it only looks at what would be disqualifying: confidentiality, no training on their data, output ownership, IP exposure.
It is forced to return one of three verdicts: this looks okay, this looks pretty good, or this looks pretty bad. And the requester has to paste that output into the ticket. The value is upstream of legal. When someone sees “this looks pretty bad,” they often go find a different tool, and the ticket never gets filed.
Those four checks may be simple enough that one person can just write them down. But most of what a legal team knows is not that tidy. So he built something more advanced: a skill called “playbook engine” that finds the rules for you.
He picked a deliberately unglamorous corpus. EXOS takes student interns from a lot of different places, and every regional site draws from a different set of local colleges, with the list changing as new locations open. So there is a constantly growing pile of educational affiliation agreements. Zero revenue, high volume, and people still redline them. He exported about fifty of them from the CLM, each in its own folder, with all of the interim drafts leading up to the signed copy.
That export is tremendously rich - what positions did I accept, what positions did I reject, what compromises did I seek along the way. All of that information is just built into the negotiating history.
Then he built the Contract Toaster to use it. Upload the playbook, put the new paper in, push down the lever, and the redlines pop out. One toaster can hold as many playbooks as you have contract types.
He also put a much bigger goal in front of the room. Right now most tools encode your negotiating positions in a proprietary format the vendor has every reason to make sticky. Switch tools and you re-enter everything. He asked whether anyone from a firm or a tech company wanted to work with him on an open format for legal playbooks instead. He has since published it as Open Playbook Format, open source and on GitHub, and he is forming a public working group.
This is the part of the night that stayed with me. We had just watched a general counsel run a live demo of something he coded himself, and then he asked us to help set a standard the industry might actually end up using. The room lit up.
For those interested in Marc’s Open Playbook Format and its GitHub repository, see this article.
Three: Co-Working Sessions
The first two are about the system. This one is about the people the system was supposed to free up.
Matthew Wojtkowiak opened with a question. Thirty years ago, firms did something he calls rotation: you come in as a junior associate, spend time in corporate, then litigation, then a couple more, and you figure out what you like while the firm figures out what you are. He asked how many people in the room had gone through one. Fewer than five hands.
Rotation did not disappear because of AI. It was gone long before. But it shows that the machinery for developing young lawyers has been coming apart for a while, and AI is now hitting what is left of it.
He worries about what people coming out of law school actually know now. His analogy was GPS, which he bought the moment it came on the market because he has always been directionally challenged. It gets him anywhere. He never learned to navigate. The question he left open: are we learning not to learn?
And the pressure runs both ways. Everything is available to everyone now, so a junior lawyer is expected to know everything. His own boss sends him a memo on any topic and asks what he thinks, and does it regularly. The same tool that removes the practice raises the bar.
Nobody is coming to fix it. Law firms were already bad at training ten years ago. Now they are worse, because there is one more thing they are not going to train on, and it happens to be the thing that will drive effectiveness.
For now, everyone is on a self-study plan. So he built his own. It is a standing appointment: “co-working session”, at least twice a week, with everybody on the team.
Three things happen in the room. They go over what tools each person is using. They go over what they think of the outputs, what they like and what they do not. And they go over all of it inside actual use cases, not demos.
That last part is what makes it work. His team negotiates roughly a hundred MSAs every three months and five hundred SOWs, arriving with varying degrees of edits. That volume is both the workload and the curriculum. They are using it to baseline their own experience, which means the team is building a shared sense of what good looks like at the same time as it clears the queue.
The training ground and the production line are the same line.
The co-working sessions are not a program. They are a fixture in how the team runs: at least twice a week, with everybody, on the calendar. I think this is a good experiment, because it changes the structure of the week rather than adding something on top of it. Most AI learning inside a legal department lives or dies on whether one person keeps pushing. Put it on the calendar and it stops being that person's job.
Part 3｜From the Open Floor
The formal part ended and nobody left.
One attendee put the same question to all three speakers: if a task is now mostly done by AI, are you still willing to pay for five hours of it?
One legal department in the room had redrawn its entire outside counsel panel the year before, made AI usage part of its regular one-on-ones with firms, and written the expectation into its outside counsel guidelines. It is spending more, not less. When they asked the firms why fees were rising if AI is doing the work, the answer was that the firms are making an enormous investment in technology, and the response from the buy side was not dismissive: I am not sure that they’re wrong. The shape is consistent across the room. Hours billed are down slightly, hourly rates are up sharply with increases in the 20 to 30% range on the table, and paralegal time is no longer being paid for at all.
Another answer was shorter. He sends less work out, and not because of budget: he deals with regional offices of tier-one firms, and those offices do not have the access to LLMs that he has, so he does not count on a discount and just takes more in house. The most useful line of the exchange was the reframe underneath it. Price per hour is the wrong unit. The conversation should be about the value you are getting for those dollars.
The second half of the evening gave us six sharp, tightly packed lightning talks.
Two of them stayed with me.
Bill Price, a three-time technology GC who now runs his own AI-native firm, estimates he has spent close to a billion dollars on outside counsel over his career, and his point was that young lawyers are not missing AI, they are missing relationship building and business development, because they grew up with a screen. He also said he never once hired a law firm in his career. He hired partners.
Nancy Peterson spent over 20 years at Allied Universal, left about a year ago, and concluded she could not serve on a board without genuinely understanding the technology, so she spent the year going into AI and did her first hackathon. Every conversation about the training gap that night pointed downward at juniors. What Nancy reminded me was that learning AI is not just a junior lawyer’s job. Some of the most experienced people in the room were learning alongside everyone else.
Thanks to Cooley LLP and Asian American Bar Association of New York for co-hosting this. And thank you to everyone who showed up on a Friday night, especially those who traveled in for it.
Come and join our next session by registering as a member of Legal Tech Frontier (legaltechfrontier.com) - the frontier community for AI transformation in law.