AI Services Roll-Up

The two paths

One large law firm has reportedly committed roughly 500 million dollars to building its own software rather than buying it. Brad Gerstner does not like the bet: "what else are they going to do? The competition is coming straight at them. But I don't think it's a high-probability bet. Am I confident that a law firm that gets up every day thinking about legal briefs is going to write killer legal software to compete with OpenAI and Anthropic? Unlikely."¹

The question underneath the number is who captures AI's value inside professional-services businesses such as law and accounting, and Gerstner frames it as a choice between two paths. The skeptical path is the law firm's own approach: building competitive software in-house. The historical base rate for services firms with weak internal software competency spending hundreds of millions building their own tooling is poor, and while the counter-argument that software-building-software changes the calculus is real, it does not change his underlying probability estimate.

The favored path is acquisition: an investment firm buying accounting companies outright and staffing them with strong engineers working in close partnership with a frontier AI lab to drive productivity gains directly. "A Thrive Holdings buying a Kirkland & Ellis and saying 'now we're going to AI-turbocharge it' seems a more likely outcome. You'll see a lot of that out of private equity firms, and take-privates done company by company."

Why it matters

The thesis locates AI's value capture in services at the ownership layer rather than the tooling layer: the durable edge is not a proprietary legal or accounting application competing directly with frontier labs and losing, it is owning the existing book of business and client relationships and layering frontier-model productivity on top of them, margin expansion on cash flows that already exist rather than a software moonshot that has to out-build OpenAI or Anthropic from scratch. Capital acquires the incumbent and disrupts it from the inside, rather than a startup disrupting the incumbent from the outside.

The rival trade

The strongest challenge to this thesis in circulation is not a critique of it but a different private-capital allocator making the opposite bet with equal conviction. Robert F. Smith argues that AI will let enterprise software eat services rather than the reverse, and is running that conversion across roughly ninety owned software companies.² Where the services roll-up thesis buys the services firm and applies frontier-model productivity to its existing cash flows, betting the durable asset is the book of business and the client relationships, the rival trade buys the software vendor and deploys agents directly inside workflows it already owns, betting the durable asset is the workflow and data substrate itself. Both are private capital chasing the same services-margin pool from opposite ends, and both are argued by someone whose fund is positioned to benefit from being right. They are not strictly incompatible, since a services firm with genuinely irreproducible client relationships and a vendor with genuinely irreproducible workflow data could each capture part of the pool, but the two views make opposite predictions about which asset actually appreciates, and the disagreement is testable over time. The specific weak point the rival trade exposes is that an argument for why a law firm cannot out-build a frontier lab is not the same as an argument that the law firm's revenue is safe from the software vendor that already owns its underlying workflow, a different attacker entirely.

Open questions

The roll-up strategy is still early, with productivity claims from at least one prominent acquirer not yet showing up as durable, measured returns. Roll-ups have historically struggled to realize cross-firm synergies even before adding an AI-turbocharging layer on top, and the entire premise assumes the productivity gain survives ordinary partnership-culture friction during integration. And the build-versus-buy debate is genuinely probabilistic rather than settled: a services firm with unusually strong technology partners could still beat the historical base rate, so the underlying bet is a probability judgment rather than a certainty in either direction.