This week Kirkland & Ellis, the highest-grossing law firm in the world, said it’s putting $500 million into building its own AI. And half of LinkedIn read it backwards.

Scroll the comments under Aaron Levie’s post and you get the same take on repeat: “they’re building their own model, worse than the frontier ones, what a waste,” “I’d do it for a fraction,” “they’ll roll it back in six months.” Everyone arguing about a decision Kirkland didn’t make.

Kirkland isn’t training a model. It’s building the layer that sits on top of the model: the one that puts 250 of its lawyers’ knowledge inside Claude, or GPT, or whichever, so the whole firm works the way its best partners work. That isn’t competing with Anthropic. It’s the opposite.

That distinction, model versus layer, is the thesis of this post. And it’s what separates the people who got the news from the people who repeated it wrong.

I’ll say what we do up front, because it’s exactly this. The piece that decides which of your knowledge is worth putting into that layer is AI Maestro, the discovery up front. The person who answers for how it ends up configured, and why, is our technology partner service. Kirkland does it with 180 technologists and half a billion dollars. The rest of this post is about why you need neither to capture the exact same value.

What Kirkland actually announced, minus the broken telephone

Strip the noise and here’s what’s left. Kirkland will spend $500 million of its own revenue ($10.6 billion last year) over three to four years, starting with $100 million in 2026, on a custom-built AI platform. It’s designed around input from about 250 of its lawyers, 100 of them partners, and built by more than 180 technologists, some in-house, some from outside firms. And those outside firms won’t be allowed to sell the same thing to competitors.

Notice what that paragraph doesn’t say. It doesn’t say “train a model.” It doesn’t say “compete with Anthropic or OpenAI.” It says platform, ecosystem, layer: software that sits on top of the models that already exist and wires them to the firm’s own knowledge. Kirkland has built its own technology before. This is the biggest bet it’s ever placed, but it’s the same play as always, just more expensive.

The firm’s chairman, Jon Ballis, put it in a line worth more than the $500 million: “Widely available AI tools are raising the floor for everyone, but we don’t get hired for the floor.” Translation: Harvey and Legora (the two legal platforms splitting the market, with 1,000-plus and roughly 700 firms each) hand anyone a solid floor. Kirkland doesn’t charge what it charges for the floor. It charges for the ceiling. And nobody sells the ceiling in a subscription.

One detail closes the point. Kirkland made sure its outside vendors can’t resell the platform to other firms. Compare that with Freshfields, whose Anthropic deal lets the resulting products be sold to rivals. One is buying a layer anyone can rent later. The other is building a layer no one else gets to touch.

It’s not a model. It’s a layer. And that difference is everything.

This is where half the thread went wrong, and it’s worth being precise, because the mistake is instructive.

Building a frontier model is what Anthropic, OpenAI, and Google do: training a network with hundreds of billions of parameters from scratch, thousands of GPUs, years of research. If Kirkland tried that, the comments would be right. It would be a worse model than the labs’, wildly expensive, and obsolete in six months.

But that’s not the play. The play is the layer above: take Claude or GPT as they ship and wrap them (the wrapper, people call it) with the firm’s precedents, its templates, the way it structures a deal, the accumulated judgment of its partners. The model is the engine. The layer is the car you build around it, the way Mac Gupta put it in that same thread. Swap the engine next year for a better one and the car is still yours.

None of this is new, and I wrote it three weeks ago: the model is already a commodity, the moat is the harness you build on top. Cursor, Devin, and Replit run the same three frontier models. Swap the model underneath and the products live. Swap the harness and they break.

Kirkland just put $500 million behind that same idea, in an industry that isn’t software. When the highest-grossing firm on earth spends like this, it isn’t betting on a better model. It’s betting that the model was never the moat.

The people who laughed at “a worse model for $500 million” hit a straw man. The straw man doesn’t exist. What exists is a firm that figured out, before almost everyone, where the value went once the models leveled off.

The $500 million is an extrapolation trap

Now the other half of the mistake, the one made by everyone who clapped without thinking.

The best objection in the thread didn’t come from a skeptic. It came from Catherine Kemnitz, who has run legal functions: Kirkland is an outlier, and drawing conclusions from what it does for the rest of the market is risky. She’s right. Kirkland grosses $10.6 billion and pays out millions per partner. At that scale, $500 million is a rational call: if it fails, the downside is contained; if it works, it becomes IP no competitor can buy. That fact set looks nothing like yours.

For 99% of companies, the Kirkland lesson is not “build your own platform.” Reading it that way is as wrong as thinking they’re training a model. The lesson is where the value sits: in the layer that connects your knowledge to a model that already exists. And for almost everyone, that layer is not a 180-engineer project. It’s workflow design on top of Claude.

I argued the flip side of this already: the most advanced AI companies buy their software, they don’t build it. Anthropic runs on Salesforce. The question was never build versus buy. It was which part of your knowledge deserves to be encoded, and who encodes it.

Copy the headline (“Kirkland builds, so will I”) without copying Kirkland’s balance sheet and you’re not buying a moat. You’re buying a cost sink with good internal press.

What IQ Source does about it

The two pieces this news exposes are, again, the two we work on. And neither one is “train a model” or “build a $500 million platform.”

AI Maestro is the discovery that decides what goes into the layer. Kirkland polled 250 lawyers to learn which knowledge was worth encoding. That’s discovery, just at half-a-billion scale. We do it in two months: we map the real processes in your operation, score each with an AI Opportunity Score, and end with a Go/No-Go gate that decides, process by process, what earns a place in the layer and what doesn’t. It’s what separates the knowledge that makes you different from the knowledge anyone downloads in a subscription.

Our technology partner service is the person who answers for that layer once it’s running. Kirkland has 180 technologists for that. A mid-sized B2B company doesn’t need 180. It needs one, with a name, who knows why the layer is wired the way it is, which model runs underneath, and what happens when that model changes next month. That role isn’t covered by a Harvey subscription or a committee.

Notice neither one is the expensive part of Kirkland. The expensive part, the $500 million, is building the platform. That’s exactly the part you don’t have to do. The part you do have to do, deciding what to encode and owning how it turned out, costs a fraction and is where the moat actually lives.

Before the week is out, ask your team one question. If you had to put the knowledge that makes your company different inside an AI tomorrow, would you know what that knowledge is and where it lives? If the answer is “in three people’s heads,” you already have your first project. It won’t cost you $500 million. It’ll cost you a decision.

Map which knowledge earns your AI layer