Kirkland & Ellis Bets $500M on Building Its Own Legal AI Instead of Buying It

Kirkland & Ellis — the world's highest-grossing law firm — is committing roughly $500 million over the next three to four years to build its own artificial-intelligence platform rather than rent one from the legal-tech vendors now courting Big Law. The Chicago-founded firm plans to start with about $100 million in 2026 and ramp from there, in what may be the most aggressive technology bet any law firm has ever placed.

The numbers behind the bet are what make it possible. Kirkland books on the order of $10.6 billion in annual revenue, a war chest that lets it absorb a half-billion-dollar build that smaller competitors could never justify. The custom platform is being shaped with input from some 250 Kirkland lawyers — including roughly 100 equity partners — and built by a cross-functional group of more than 180 technology professionals, with the firm advertising around 85 dedicated AI roles, from "AI Innovation Advisers" who translate legal workflows into scoped tooling to infrastructure directors who will run on-premise GPU clusters alongside Microsoft Azure.

That on-premise GPU footprint is the tell. It gives Kirkland the capacity to fine-tune open-source large language models on its own matter data behind its own firewall — a path it has hinted at without naming which base models it would tune. The firm says it will keep licensing third-party tools such as Harvey, Legora, Thomson Reuters' CoCounsel, and Lexis+ AI where they fit, but it wants to own — or hold the right to own — the core platform outright. Outside companies hired to help build it would be barred from reselling the technology to anyone else.

Kirkland's go-it-alone stance contrasts with rivals leaning on the AI labs directly: Freshfields, for instance, has worked with Anthropic's legal team to develop specialized tooling. Industry watchers frame Kirkland's move as the moment Big Law's old aversion to custom software fell away — and as a fresh way for the global elite firms to pull further ahead, since the scale required to bankroll a proprietary stack is itself a moat.

The wager is not risk-free. Generative models still fabricate case citations, misquote authorities, and invent nonexistent references — failures that have already drawn court sanctions against lawyers at firms including Sullivan & Cromwell. Owning the stack lets Kirkland keep privileged client data on its own infrastructure and tune for accuracy, but it also means the firm, not a vendor, owns the consequences when the system gets something wrong. For a partnership that sells judgment, that trade-off is the whole point.