In June, I wrote about why AI training had moved inside the enterprise: the open web has run dry of the cognitive density frontier models now need, and the next phase of the Applied-AI race would be won on access to human tacit knowledge, not public data.
I said the leading tech vendors had already restructured to capture it. What I didn't have yet was a single, undeniable proof point of a professional services firm making the same bet with its own money.
Now I do. Kirkland & Ellis, the world's highest-grossing law firm, is spending roughly $500 million of its own revenue over the next three to four years building proprietary AI systems it will own outright.
A meaningful piece of that is a co-built platform with Palantir called the Fund Formation Engine, designed to run private-equity fund documentation, side-letter drafting, and obligation tracking across Kirkland's fund-formation practice.
Kirkland didn't pick Palantir for its model. It picked Palantir for its ontology-modeling expertise — the ability to represent a firm's decisions, obligations, and relationships as a structured, queryable graph rather than a pile of indexed documents.
That distinction is the whole story, and it's worth being precise about why.
Text Retrieval Was Never Going to Be Enough
Every enterprise AI deployment from 2023 to 2025 followed the same template: take a frontier model, bolt on a vector database, call it RAG.
It works fine for "find me the clause that looks like this one." It falls apart on "what did we decide, and why, and what does that obligate us to now" — because a flat document index has no model of decisions at all. It just has text.
The firms moving fastest past this limitation are building what I'd call a Proprietary Intelligence Foundry: a pipeline that captures the exhaust of daily work — email, meetings, negotiation threads, collaboration patterns — and turns it into a governed knowledge graph before any retrieval happens at all.
That graph becomes the substrate for an AI training curriculum, which fine-tunes a smaller, specialized model that reasons over the firm's own judgment instead of averaging across everyone else's.
The durable asset in this stack was never the model. It's the graph.
Two Kinds of Collapse, and Why the Second One Should Worry You More
Anyone paying attention to Applied-AI Initiative research has heard of model collapse — the well-documented risk that training generation after generation of models on synthetic data erodes the tails of the original distribution.
It's real, but it's also manageable: recent work shows that accumulating synthetic data alongside real data, rather than replacing it, keeps the degradation bounded.
The failure mode that should actually worry a firm like Kirkland is different. Call it knowledge collapse: the tendency of any shared, vendor-hosted AI model to regress everyone's output toward the statistical center of its training data.
If every firm in a market routes its differentiated judgment through the same off-the-shelf AI assistant, the assistant quietly erases the differentiation. For a law firm, a consultancy, or a bank whose entire pricing power rests on judgment competitors don't have, that's an existential risk dressed up as a productivity tool.
It's close to the exact language Kirkland used to explain why it built rather than bought: generic AI tools trained on broad market knowledge tend to converge on a lowest common denominator, and a firm charging $2,000 an hour cannot afford to sound like everyone else.
The Constraint Nobody Wants to Talk About
Here's the part that should temper any senior executive's enthusiasm before they green-light a similar build: none of this works if you assume your best people will happily spend un-billed hours training it.
Roughly 90 percent of legal revenue still runs through the billable hour, a compensation model essentially unchanged since the 1950s. Knowledge-sharing isn't factored into partner comp at most firms, which means the people with the most valuable judgment to capture have the least incentive to sit still and hand it over.
Client contracts are making this worse, not better — outside counsel guidelines increasingly bar firms from billing for AI-driven time savings outright, which means any ROI case for this technology has to be made through win rate, new revenue lines, or capacity redeployment. Never through hours saved.
That's a genuinely awkward position for a firm to be in: forbidden from billing for the efficiency the tool creates, while still needing to justify a nine-figure investment in it.
The firms getting this right aren't scheduling training sessions. They're mining signals their most skilled people already generate — email, meetings, document collaboration — without asking anyone to stop and teach the AI tool.
It's the same passive-capture principle I flagged in June, now with a name and a price tag attached.
The Takeaway for Every Other Knowledge-Intensive Industry
Kirkland is a law firm, but nothing about this architecture is legal-specific.
Any professional services business whose value is concentrated in a small number of senior people's judgment — consulting, investment banking, specialized engineering, high-end agency work — faces the identical structural choice: rent a generic AI assistant that averages your expertise into the market's, or build the unique graph that keeps it yours.
The interesting white space right now is that almost nobody in Management Consulting has made this bet publicly yet. Every case study I could find sits in legal or corporate finance. That's either a sign consulting firms haven't caught up, or a sign the smart ones are building quietly and haven't said so publicaly.
Either way, I'd bet on this thesis showing up in a consulting firm's own AI investment announcement within the next two to three quarters — and when it does, it won't be a coincidence that the forward-thinking firm making it competes on judgment, not headcount.
For the AI architecture, vendor landscape and sourcing, see my Intelligence Foundry research and analysis.
Reach out to learn more about our Applied-AI Initiative objectives.