The Long Lake playbook
A three-year-old AI holding company just closed a $6.3 billion take-private of Amex GBT. How the model works, what it claims, what is still unproven, and what private equity should take from it.

In shortOn Sept 29, Long Lake closed its $6.3 billion take-private of Amex GBT, the biggest test yet of the AI roll-up. It buys healthy services businesses, puts engineers inside them for years and uses the time AI frees up to grow. Its results are self-reported, and the part a fund can't copy is that it never sells.
On Sept 29, a company founded in 2023 completed the acquisition of a company founded in 1915. Long Lake, which describes itself as the world's first AI holding company, closed its all-cash take-private of American Express Global Business Travel at a value of about $6.3 billion. The money came from Long Lake's existing investors and Koch Equity Development, along with $2.5 billion of committed debt from JPMorgan, Bank of America, Citi and MUFG. Holders of about 69% of the shares, including American Express, Expedia, the Qatar Investment Authority and BlackRock, had agreed to support the deal.
Amex GBT reported $2.7 billion of revenue and $182 million of operating profit in 2025, and it manages travel for about 17,500 businesses in 140 countries. Buyout funds usually compete for companies like that. Until this year, AI startups didn't.
For private equity, that makes Long Lake a competitor. It bids for the same kind of company, borrows from the same banks, and then runs the business in its own way. The rest of this piece looks at how it runs them, how much of that is proven, and which part a fund would struggle to copy.
Who Long Lake is
Alexander Taubman founded Long Lake in 2023 with backing from General Catalyst, and investors including Alpha Wave, Elad Gil, D1 and Thrive joined along the way. In interviews this year, Taubman has described a group of close to 40 companies in five industries: homeowners association management, HR services, specialty tax, infrastructure services such as architecture and engineering, and now corporate travel. With Amex GBT, the group employs roughly 30,000 people, and Taubman expects revenue above $4 billion next year.
The company reports big gains. Long Lake says its first group of acquired businesses doubled EBITDA in under two years. General Catalyst's chief executive wrote in May that Long Lake had doubled free cash flow in its HOA management businesses in two years without cutting headcount. Taubman has said those HOA businesses went from roughly zero to 5% annual volume growth before the acquisition to more than 20% afterward, and that the average employee saves more than eight hours a week with its platform.
None of these numbers are audited. They come from management and investors, and from the early, smaller acquisitions. What makes the model worth studying anyway is how openly Taubman has described it: how Long Lake chooses companies, what it builds inside them and what it does with the gains.
How it chooses companies
Long Lake picks the industry first. It keeps a ranked list of 15 to 20 industries and, by Taubman's account, screens four to five thousand deals a year against it. The reason is economic: the platform work only pays off if it can be reused across many companies in the same field.
Within those industries it looks for services where failure is expensive, customers rarely leave, net revenue retention runs above 100%, management is experienced, and the work will stay valuable as AI improves. Taubman separates high-quality services that AI makes better from lower-quality ones that AI can simply replace, and he only wants the first kind. Amex GBT fits: its customer retention is in the high nineties, and its relationships with its largest customers average about 15 years.
The companies also have to be healthy already. Taubman has put the bar at EBITDA margins of 20% to 25% or better, with management teams worth keeping. That makes Long Lake very different from a turnaround investor. Its bet is that AI makes a good company better, and that a profitable business can reinvest what it gains.
One platform, and engineers on site
Every Long Lake company runs on a single AI platform the company calls Nexus. Taubman describes it as model-agnostic software that sits between the AI models and a business's data, skills and workflows. By his estimate, 70% to 80% of it is shared across industries. The other 20% to 30% is the work of fitting it to a particular company.
That split explains why the model gets cheaper as it grows. Taubman has said early acquisitions took more than a year to show results, and that new ones now see an effect within days, because most of what a company needs already exists.
The expensive part is the people. Taubman has said about 70% of Long Lake's engineers sit alongside frontline staff in the operating companies, sometimes for two years. When the person building a tool sits next to the person who needs it, a fix takes days. Long Lake also treats a tool nobody uses as the engineers' failure. In Taubman's words, if it has to force someone to use the platform, the platform isn't good enough. In the HOA businesses, replies that used to take 45 minutes to an hour now take five to ten, written in each manager's own voice.
What it does with the time AI frees up
Most AI business cases count the hours saved and then cut staff. Long Lake says it does the opposite. It uses the freed hours to serve customers better and win more of them. Taubman has said its oldest companies are adding jobs, paying people more because each person produces more, and keeping them.
We think this is the most important part of the model. In a competitive market, cost savings tend to disappear as rivals adopt the same tools and cut prices, a pattern we looked at in beyond cost: AI as a revenue engine. Better service from people customers already trust is much harder for a rival to copy. If doubling free cash flow without cutting staff holds up, that's what the growth path looks like in the numbers.
Why it never sells
Long Lake says it has never sold a company and doesn't plan to. Taubman describes transforming an operating business as a two-to-five-year job, and asks why anyone would spend years building the best business in an industry only to sell it. He has said he wants to keep growing these businesses for decades.
Owning companies indefinitely is what pays for everything above. It covers two years of engineers on site, a platform paid off over decades, and the patience to let people adopt the tools at their own pace. A private equity fund, which has to sell, can't make those trade-offs in the same way. We come back to that at the end.
Others trying the same thing
Long Lake is the most visible example, and there are others. General Catalyst committed $1.5 billion to what it calls AI-enabled roll-ups. It mapped 70 service categories and found ten where today's AI can automate 30% to 70% of the work, aiming to at least double the EBITDA margins of the companies it buys. Thrive Holdings' accounting platform, now called Current, passed 50 firms this summer, with more than 2,000 employees and over $500 million of revenue. It works with engineers placed by OpenAI and reported that its tax product prepared 7,000 returns last season while cutting preparation time by about a third.
October brought two more signs of where this is going. On Oct 29 Thrive Holdings agreed to buy ADT in a $15.3 billion deal, a second large company bought to be run with AI. And on Oct 19 the Federal Trade Commission and 12 states sued to block a playground equipment deal, pointing to the buyer's three decades of acquisitions. That deal wasn't AI-driven, but it shows regulators are willing to judge a roll-up by its whole history.
What hasn't been proven
The results are self-reported. None of the headline figures have been independently audited, and they come from early, smaller acquisitions that suited the model.
Scale is also a different problem. An HOA management firm is a long way from a global travel company with $2.7 billion of revenue, decades of old systems and operations in 140 countries. Amex GBT is the first real test of whether the approach travels.
AI can create work as well as remove it. Research from Stanford's Social Media Lab and BetterUp Labs found that 40% of surveyed employees had received low-quality AI-generated work they had to fix, at a cost of nearly two hours each time. A model built on productivity has to measure the net effect, and tool usage alone won't show it.
The people are hard to find. Engineers who can sit with a frontline team for two years and build tools that people choose to use are scarce and expensive.
And the debt is real. The Amex GBT deal carries $2.5 billion of it. AI doesn't change what leverage does to a business when results disappoint.
What private equity can borrow
Most of this a private equity firm can do. Picking industries before companies, buying healthy businesses, building one platform and reusing it, putting engineers in the field and taking the gains as growth all fit a buy-and-build strategy. Funds already know how to source, finance and integrate add-ons, which made up roughly three quarters of all U.S. buyouts in the first half of 2026.
Where a fund runs into trouble is the sale. Whatever AI capability it builds during the hold has to go with the company to the next owner. In practice that means a platform layer the portfolio company owns, that runs on any model and that a buyer's diligence team can inspect, so the capability doesn't leave with the sponsor's engineers. We go through what that takes, and how a firm can build it without turning into a software company, in building an AI roll-up platform.
If Long Lake's numbers hold at Amex GBT's scale, it will have shown that AI can make an old, trusted services business grow again. Over the next two years, Amex GBT is where we'll find out.


