The M&A route to AI deployment: buy the customer, own the outcome

The acquisition play. Prathik Naidu, founding partner at Long Lake, argues the hardest part of enterprise AI is the vendor-customer wall, not the model. Security reviews, legacy integrations, training, pricing, and value capture drain outside vendors. His answer is to buy the customer. Long Lake, an applied AI holding company with $7.5B from Elad Gil, General Catalyst, Alpha Wave, and D1, has acquired 40 services businesses and embeds AI inside them. Its $6.3B Amex GBT take-private shows the scale.

Why services. Naidu contrasts software and services. Software keeps about $80 of every $100 of new revenue; services keeps about $20 because delivery and overhead scale with headcount. AI breaks that link, so labor costs stop rising in lockstep with revenue. The examples are concrete: an HOA community manager can run only so many associations; an architect, only so many hospitals. AI is meant to give these businesses software-like margins, instead of growth that only scales with headcount.

Ownership effects. Buying the firm instead of selling to it gives the AI team direct access to people with decades of judgment, real tasks and edge cases for evaluation, aligned incentives, and customer trust. Naidu calls this a compounding flywheel. The argument is coherent, but it renames the classic M&A integration problem. Tacit knowledge and internal politics do not disappear; ownership means absorbing them into the P&L instead of negotiating across a contract.

The J-curve. The talk describes a familiar pattern: invest in platform, workflows, and training; hit a valley of hard work; then usage climbs and capacity appears. The next step is the one most AI projects skip. Capacity must be converted into growth, client or employee NPS, and higher revenue per employee. That is where management teams and incentives matter. None of this is measured for the 40 companies.

Missing evidence. No deployment timelines, revenue per employee changes, client retention numbers, or margin figures appear in the talk. The most cited number is the fund's own capital. That makes this an investment thesis, not an engineering case study. The J-curve is promised, not shown. Engineering leaders should treat the compounding language as a hypothesis and ask for the same evidence they would demand from any vendor.

What to watch. The real signal is market structure. If M&A becomes a credible way to deploy AI at scale, enterprises will face competitors that own the outcome end to end rather than vendors that sell a tool. Teams should watch whether Long Lake's portfolio publishes comparable results. If it does, the buy-versus-build decision around AI changes. If it does not, this remains a capital story.