TL;DR: If you're an equity partner, contribute your knowledge to your firm's AI build. That's what equity means. Skim past that line and you'll misread everything under it.

Now the question nobody's asking. A Legora executive just laid out the endgame on a podcast: partner knowledge encoded into agentic skills becomes firm IP, and once that's done, partners stop being portable. Fine. Except most of this encoding is happening inside platforms firms don't own, in formats they can't export. A moat on rented land is a lease.

So, two assignments. Firms: if you left your vendor in eighteen months, what exactly would you take with you? Partners: the layers that encode well are the same layers every firm already shares. What's left is your differentiation and your distribution. Go build those.

Before anything else, so nobody misreads me halfway through: if you are an equity partner, you should contribute your knowledge to your firm's AI efforts. Fully. That is what equity means. You own an institution; you are not renting a desk from it. Nothing in this issue argues otherwise, and if you quote me saying otherwise, you skimmed. What this issue is about is a different pair of questions. First: does the thing your firm is building actually belong to your firm? Second: what should you be building for yourself at the same time?

Now.

Legal AI’s strategy has been laid bare and you need to understand it because it is not unique to them.

This week, Legora's VP of Legal Innovation, Kyle Poe, went on Jacob Robinson's Law of Code podcast and described the endgame of legal AI more candidly than I've heard from anyone at a platform company. Read what he said before I tell you what I think it means.

On what firms are building:

"The ability to take the knowledge and experience of your lawyers, of your partners in particular, and encode them into reusable skills that will become the firm's IP... That is going to be the equivalent of what source code is to Apple . . . these agentic skills will be to law firms in the future."

On how that changes lateral hiring:

"The sophisticated firms are thinking... what data moats do partners out there have that we might be able to bring into our firm if we bring those partners in?"

On what happens after the encoding is done:

"If a partner walks out the door and goes across the street, they're not taking that IP with them... The firm that they left has a copy of all that knowledge and experience in the form of these agentic skills that remains behind."

There's a detail in there most listeners will blow past. Poe estimates it takes a departing partner about six months to rebuild their system at a new firm, and he says that lag "gives a window of opportunity for the firm they left to patch over the client relationship."

Stop right there. Sit with that. The copy of you that stays behind gets deployed against you.

His prediction: lateral movement accelerates now, then locks down as encoding matures, because portability collapses once your differentiation stays behind when you go.

Poe is nobody's fringe voice. Legora raised $550 million at a $5.55 billion valuation in March, serves 800-plus customers across 50-plus markets, and counts a meaningful slice of the AmLaw 100 as clients. He was a Morgan Lewis partner for a decade before this job. When someone in that seat describes the strategy out loud, believe him.

The take everyone will have

The easy reading is partner versus firm. They're encoding you to compete with you if you leave.

I've hit on this before. In Issue #286 I asked on what terms a partner would ever agree to be assimilated, and I stand by it. Firms will have to answer it eventually. But first . . .

Let’s put the partner and the firm on the same side of the table, not across from on another. Sure, the lateral market is crazy but there are more partners happy to stay at their firm than there are shopping. So let’s bundle the partners and the firm together and ask a question almost nobody is asking. Would the moat Poe describes belong to the firm at all?

Rented land

Follow what is really happening here, not the marketing speak.

A firm encodes its partners' judgment, playbooks, and knowledge into agentic skills. Those skills live in a legal AI platform's (a vendor) proprietary format. Not the firms. They run on the vendor’s infrastructure, alongside the encoded judgment of dozens of other firms doing the same thing on the same vendor platform. Look at how legal AI platforms brag about the number of firms they are working with. Then think of all the usage patterns, which workflows/playbooks fire, in what sequence, corrected and reviewed by whom. This all accrues to the vendor layer. Forever.

Poe says agentic skills will be to law firms what source code is to Apple.

Fine. But perhaps the better way of saying it is all of the ways of working these legal AI vendors siphon off are the source code to them competing with you.

I made the underlying point in Issue #289: "we don't train on your data" and "we don't learn your methods" are two different promises, and vendors only make the first one loudly. Harvey's security page, to its credit, is clear that inputs, outputs, and documents don't train the underlying models. Ask any vendor about the second promise, the one covering workflows, patterns, and the shape of how elite legal work actually gets done, and watch how the conversation changes.

And I'm not singling out Legora here. This is the incentive component of the entire vendor platform layer, and the platforms keep telling us so themselves. On July 1, Palantir CEO Alex Karp went on CNBC and accused the frontier labs of "stealing [their customers'] weights and alpha" while enterprises pay for "tokens that create no value." His targets were OpenAI, Anthropic, and Google.

Twenty-seven days earlier, Palantir's own press release had announced a platform that "embeds Kirkland's institutional knowledge" from its thousand-lawyer funds practice. But Palantir is model agnostic and Kirkland appears to understand the game. Bloomberg Law reports that the firm's $500 million build is deliberately model-agnostic to avoid vendor lock-in, with its own engineers and on-prem GPU clusters showing up in the job postings. Whatever you think of the bet, they're at least trying to own the land under it.

One more turn of the screw. Legora itself runs "mostly on Claude." And as I showed in Issue #270, the leaked Claude guidelines define a principal hierarchy that puts the model provider above the operator and the operator above the user, with the provider functioning, in the document's own words, like "a silent regulatory body or franchisor." The vendor selling your firm a moat is itself a tenant.

Even Debevoise, whose client-facing STAAR platform I consider one of the smartest strategic moves in BigLaw right now, built STAAR 2.0 on Legora's Portal architecture. Thousands of hours of the firm's AI-governance judgment, made queryable by clients like Blackstone and GSK, running on a platform the firm doesn't own. That's not a criticism of the tool. It's a description of where the tool lives.

The sovereignty test

None of this argues against encoding. The firms that encode well are going to pull ahead, and they should get on with it. The argument is about where the encoding lives.

Picture a firm that ran its encoding on open-weight models it controls. Weights the firm owns. Skill libraries in portable, inspectable formats. Evals the firm designs and runs, which is what I meant by owning the exam in Issue #271. Infrastructure the firm governs.

In that world, Poe's source-code story comes true. The moat attaches to the firm as there is no legal AI vendor. And the conversation with partners about contribution becomes what it should have been all along: a properly constrained surface area for bargaining. Two parties negotiating over one asset on real terms, with no third party holding the board.

Fantasy? The pieces exist today. OpenAI now ships open-weight models you can run on your own hardware. Microsoft and Mistral announced a partnership in July aimed at regulated industries, promising frontier AI deployable in fully disconnected environments. Their phrase, not mine: "frontier AI they can control." LexisNexis put Mistral inside Lexis+ Protégé in France and sold it as sovereignty. Luminance's CEO gives interviews about why owning your models beats "rented intelligence." Open skill formats are portable across thirty-plus tools. Shawn Curran, who ran legal technology at Travers Smith, said it best: think "like Coca-Cola preferring to keep its recipe internally and have optionality on water suppliers."

So I went looking for the law firm running its encoding on weights it owns.

I could not find a single named firm doing it in production.

You know who I did find? A general counsel. In Issue #285 I wrote about a GC who built a free, open-source, self-hosted legal AI platform for in-house teams, designed so legal departments handle more of the substantive work before it ever reaches outside counsel. Read that again slowly. The clients are going sovereign while the firms sign leases.

If you sit on an executive committee, one diligence question earns its keep this quarter: if we left this vendor in eighteen months, what exactly would we take with us? If the honest answer is "prompts, and a negotiation over data export," you don't have a moat. You have a lease, and your landlord just told a podcast what the property is worth.

What's actually in the encoding

Now the partner side. And here I want to lower the temperature, because the panic take is as wrong as the grievance take. And I sort of started it so allow me to be more thoughtful.

Be precise about what encoding captures. I think of a great partner's value in four layers.

Gravitas. Reading the boardroom. Knowing when to speak, when to sit in silence, how to shape a conversation through influence both direct and indirect. The art of positioning. This does not compress into a skill file. It exists live, in the room, or not at all.

Foresight. Seeing seven steps ahead because you've sat on the other side of the table, in litigation or on deals, and you know the breadcrumbs when you see them. Instinct built from adversarial reps. The training data for this doesn't exist in any document management system, a point I made in Issue #271 when I described the heuristics that matter most as "accumulated, often undocumented ways experienced lawyers decide what matters, what doesn't, and how far to push."

Judgment heuristics. Rules of the road. Playbooks. How we handle this fact pattern, this regulator, this counterparty. Partially encodable, and this is what the platforms are actually after.

Production. Drafting, diligence, markup, synthesis. Fully encodable. Already commoditizing.

Notice something about the two layers the platforms can reach. Judgment heuristics and production are the most homogeneous layers in BigLaw. We can debate everyone's unicorn status all day, but at the level of playbooks and production, one sophisticated practice looks a great deal like another. Same website language. Same conference talking points. But for the logo, you might never know who you were dealing with. I wrote that in Issue #283 about firm marketing. It applies at least as well to the encoded layers.

So the encoded moat is less proprietary than the platforms claim. And the encoded loss is less existential than nervous partners fear. What's actually at stake in the encoding is smaller than both sides think.

Your attention belongs on what the encoding can't reach.

Grinders get encoded. Finders get paid.

Poe reaches for the old taxonomy himself in the episode: grinders, minders, finders. He's right to. Encoding automates the grinder and compresses the minder. Every quarter, it raises the price of the finder.

I said my piece about contribution at the top, and I meant it. Contribute the playbooks. Contribute the rules of the road. Honestly, that material is less unique than any of us likes to admit, and your firm should hold it. Your firm, in turn, should be able to tell you where its encoding actually lives. The eighteen-month question cuts both directions.

Contribution is one thing. Self-erasure is another, and the difference between them is what you build for yourself while the encoding happens. Because when the encodable layers homogenize, and they will, across every firm, on the same short list of platforms, what remains personally yours is differentiation and distribution.

Most partners misunderstand both words, badly.

Partners think differentiation means credentials. The Chambers band. "Deep experience in complex matters," a sentence that appears on every bio at every firm and therefore differentiates nobody. Actual differentiation is a sharply defined practice genome pointed at a specific market position: the clients you're built for, the work you're best at, the handful of signals that precede your best matters (Issue #287).

Partners think distribution means the referral network and two panels a year. Serendipity with a calendar. Actual distribution is owned, systematic access to demand before it becomes a matter. Presence at the pre-matter layer, built on architecture, the way investment banks built coverage around their best bankers: "Not their intelligence. Not their relationships alone. The architecture behind them that made them impossible to replace without starting over." The research backs this up. Activator-type business development behaviors correlate with up to 32% higher partner revenue generation.

The market is already pricing all of this, whether or not partners understand what's being priced. Lateral partner hiring in the Am Law 200 jumped roughly 20% last year to 4,152 moves. Twenty million dollars is the new benchmark package and the top of the market has crossed forty. The American Lawyer just coined a name for the partners commanding those numbers: "franchise value" laterals. Even Debevoise, lockstep's standard-bearer, added a discretionary bonus pool in May to compete for exactly this kind of partner.

Poe reads the lateral frenzy as firms racing to acquire data moats before the window closes. Maybe. But look at who's commanding forty million dollars. It isn't the partners with the best playbooks. Those are being encoded as we speak. It's the partners with markets that know their name.

So contribute your playbooks like the owner you are. Then spend this year building the two things no platform will ever hold for you: a differentiation that's actually yours, and a distribution architecture pointed at demand. Do both and you're worth more after the encoding than before it.

Do neither and you have a problem. But the AI initiative didn't cause it.

My close

The platforms have told you what they're playing for. It was said out loud, on the record, this week. Take them at their word.

Firms: encode, and run the eighteen-month test before you call the result a moat. Land you own compounds. Land you rent appraises to someone else's balance sheet.

Partners: I put your assignment in the first paragraph, and the second half of it is the part nobody else can do for you.

A copy of you is going to exist either way. The open question is what the original does next.

Talk soon,

Josh

If you work in fintech or finance, you already have too many tabs open and not enough time.

Fintech Takes is the free newsletter senior leaders actually read. Each week, I break down the trends, deals, and regulatory moves shaping the industry — and explain why they matter — in plain English.

No filler, no PR spin, and no “insights” you already saw on LinkedIn eight times this week. Just clear analysis and the occasional bad joke to make it go down easier.

Get context you can actually use. Subscribe free and see what’s coming before everyone else.

Sign Up Free

To read previous editions, click here.

DISCLAIMER: None of this is legal advice. This newsletter is strictly educational and is not legal advice or a solicitation to buy or sell any assets or to make any legal decisions. Please /be careful and do your own research.