The Specificity Gap Just Got Expensive

Five hundred million dollars buys a lot of vendor licenses. The world's highest-grossing law firm is spending it on something else.

Kirkland & Ellis confirmed this week it will commit $500M to build its own AI platform — a system the firm describes, in marketing language that should be treated as marketing language, as the 'collective intelligence of its lawyers.' A nine-figure capital commitment proves strategic intent, not execution. The internal build could ship and quietly fail. I don't know whether it will, and that's not the interesting part.

The interesting part is that a partnership whose entire business is converting partner-hours into fees has decided the vendor stack isn't good enough to bet those hours on.

This is a trajectory story, not a one-off.

Before the cheque

Eighteen months ago the consensus among Am Law general counsel was that enterprise AI procurement would consolidate around two or three foundation-model vendors plus a thin layer of legal-specific wrappers. The pilots were promising. The demos were dazzling. The unit economics were unexamined.

Then the bills came in. Axios this week called it 'sticker shock' and coined toxenmaxxing for the dynamic. Enterprise teams optimizing for capability demos discovered, when the invoices arrived, that they had also been optimizing for token consumption. That's one diagnosis: a cost problem.

There's a second diagnosis Axios doesn't fully name. The vendor platforms aren't only expensive. They're generic in a market that pays for specificity. A law firm's value isn't access to frontier reasoning. It's access to how this firm has handled this kind of deal, this kind of regulator, this kind of partner across two decades of partner-track memory. General-purpose models don't have that. Wrappers don't either, not at the resolution that matters.

Call it the specificity gap.

The specificity gap

The specificity gap is the distance between what a foundation model can do on a benchmark and what a knowledge business actually sells. For consumer software, the gap is small and closing. For elite professional services, the gap is structural — the firm's IP is the residue of thousands of partner decisions, most of it unwritten, almost none of it in any vendor's training corpus.

Fortune ran a thoughtful piece this week framing AI's productivity puzzle through the pre-internet Solow paradox: individuals are faster, the macro numbers haven't moved, history says be patient. As a general claim about technology diffusion, that's defensible. As a guide for how a Kirkland partner should plan a 2027 budget, it isn't. Law firm economics don't run on consumer-internet diffusion curves. They run on quarterly partner draws and a billable-hour model that punishes any tool whose unit costs are visible to clients.

Kirkland's $500M is what you spend when you've concluded the specificity gap won't close on a timeline that matches your business cycle, and you'd rather own the asset than rent it from a vendor whose roadmap doesn't prioritize you.

Eighteen months out

By the end of 2027, expect at least three more Am Law 20 firms to announce comparable in-house builds, and expect a wave of mid-market firms to quietly pause vendor procurement while they wait to see whether Kirkland's product actually works. The first wave was pilots. The second wave is capital allocation as admission. The third wave — the one that breaks the vendor narrative — is the procurement freeze.

If I'm wrong, it'll be because a vendor (probably one of the legal-specific players, possibly an unbundled offering from a hyperscaler) ships a customization layer that closes the specificity gap cheaply enough to make $500M look like vanity spend. That's possible. I don't think it happens in eighteen months.

What I'm sure of: every general counsel still budgeting AI spend on the assumption that capability gaps will close on the vendor's clock should read the Kirkland announcement twice. Once as news. Once as a procurement warning written in nine zeros.