The Rise of the AI Roll-Up

Last Wednesday, Grant Thornton agreed to buy CBIZ for $5 billion in cash... a deal, backed by private equity firm New Mountain Capital, that creates the fifth-largest accounting firm in America.

In May, a two-year-old company called Long Lake agreed to take American Express Global Business Travel private for $6.3 billion. Long Lake started out buying homeowner association managers.

And last week, Dwelly raised $170 million to keep buying letting agencies in the UK.

Three deals, three very different sizes, one playbook: buy service businesses, rebuild how the work gets done with AI you own, keep the margin. It has a name now... the AI roll-up... and in under two years it has gone from a Silicon Valley experiment to taking NYSE-listed companies private.

If you own, run or advise a professional services firm, this edition is about the buyers... who they are, how their machine works, and why one of them may knock on your door before you expect it.

Let me briefly go through the definition, because it matters.

An AI roll-up buys existing service firms... accounting practices, letting agencies, IT providers, call centres... and rebuilds their production around AI built in-house. Not software sold to the firm. Software installed by the owner, inside businesses it controls, with the savings kept rather than passed on.

Three parts make it work. Permanent capital, so nobody is forced to sell in year five. Embedded engineering, because the AI is built from inside the workflows, not bought off a shelf. And recurring service revenue underneath... tax returns, rent collection, support tickets... the boring annuity income that renews by law or by habit.

It is not the same thing as an AI-native firm, though the two meet in the middle. Corgi and Ominimo, from the insurance edition, built from scratch and acquired nobody. The roll-up starts from the other end: it buys the clients and the licences ready-made, then swaps the production line underneath them. And as I wrote in my AI-native firm playbook, acquisition is one of the routes to becoming AI-native in the first place... buy the relationships, rebuild the work. The roll-up is that route with permanent capital behind it.

The model has a famous ancestor. Constellation Software spent three decades quietly buying hundreds of small software companies and never selling them, and became one of the best-performing stocks in Canadian history. The AI generation is running the same machine on services... with a margin lever Constellation never had.

The best way to see the strategy is by size, because the same mechanic is now running at three different scales.

The small deals. Dwelly buys UK letting agencies... 17 so far, 15,000 properties, a top-10 UK letting agency assembled in a few years. The founders ran an agency themselves in Hull for 18 months before writing code, then built the system around what they saw. One staff member now manages 300+ properties against an industry norm of around 100. A traditional agency turns 10 to 15 per cent of fees into profit... EQT, which just led the $170 million round, says the most integrated agencies on Dwelly's platform do several times that.

The holding companies. One level up sit permanent vehicles built to buy service firms and keep them... no fund deadline, no forced exit. Thrive Capital... one of OpenAI's biggest backers... created Thrive Holdings last year with $1 billion for exactly this, and is reportedly raising $2 billion more from SoftBank, Altimeter and D1. Its flagship, Current, has bought nearly 50 local accounting firms, has reported revenue past $300 million, and was Accounting Today's fastest-growing firm of 2025. Thrive executives told Forbes they are committing another $1 billion to buying accounting practices.

Then, in December, OpenAI took an equity stake in Thrive Holdings itself. It embeds its own researchers and engineers inside the portfolio companies, and its stake grows as the businesses perform. Engineers from both companies built a tax-processing agent on OpenAI's Codex that now runs inside Current. In testing, Thrive says its agents reach up to 98 per cent data-entry accuracy.

General Catalyst has co-created roughly a dozen of these vehicles. Eudia buys legal services. Titan buys IT providers, and says its pilots automate 38 per cent of typical MSP tasks. Long Lake started with homeowner associations and hit $100 million of EBITDA in under two years. Crescendo, an AI-native call centre platform, runs gross margins of 60 to 65 per cent... roughly four times the industry norm... and acquired PartnerHero, a 2,800-person firm that should have been an acquirer itself. Lightspeed and a16z are entering. A former 8VC principal launched Kodiak to do the same for small businesses.

The mega-deals. In May, Long Lake agreed to take Amex GBT private for $6.3 billion... the strategy's largest deal, and proof it reaches listed companies. And last Wednesday, New Mountain-backed Grant Thornton agreed to buy CBIZ, the only publicly traded accounting services provider in America, for $5 billion in cash... a 54 per cent premium to where the shares had been trading. Grant Thornton had already announced a $1 billion investment in AI and technology. The announcement itself promises "AI-enabled service delivery." Build the platform once, then buy the clients to run through it.

One more detail from that deal, for readers of the insurance edition: CBIZ's benefits and insurance arm is being carved out as a separate company... backed by New Mountain. Even inside the biggest accounting deal in decades, the insurance brokerage gets kept as its own PE asset.

Start with the old engine, from the insurance edition: buy a small firm at 5 to 9 times earnings, fold it into a platform valued at 13 to 18 times, and the gap is profit by arithmetic. That engine has run for decades. It is why private equity sits behind seven of every ten US insurance agency deals.

The AI roll-up adds two new levers.

The first is margin. Traditional roll-ups cut costs around the edges... shared back office, better procurement. The AI version rebuilds the production itself. When one person can manage 300 properties instead of 100, or a call centre runs at four times industry margins, the acquired firm is not slightly more profitable. It runs on a different cost base under the same brand.

The second is data. Every tax return, claim and support ticket processed captures the judgment of the person doing it... what they flagged, what they fixed, what they let through. That is exactly the training material the models lack, and it compounds: each acquisition makes the platform smarter, which makes the next acquisition cheaper to integrate. Traditional private equity did financial engineering on fixed cash flows. The AI roll-up buys the cash flow and then changes what it costs to produce.

Which is why the buyers now include the AI labs themselves. OpenAI's equity in Thrive Holdings scales with portfolio performance... the lab now owns part of the machine, not only the models running inside it.

Roll-ups have a long history of destroying value, and this generation will not be exempt. Integration is brutal. Buying 50 accounting firms means 50 cultures, 50 half-migrated systems, and partners who can retire the moment the earn-out clears. The AI margin numbers are early, mostly self-reported, and produced by the people raising money on them. Crescendo's margins are reported, not audited. And a 54 per cent premium only pays off if the platform story is real.

My call: some of these will fail on integration, the way roll-ups always have. But the direction is not in doubt, because the two hardest parts of the old model... finding sellers and funding patience... have both flipped. The retirement cliff supplies the sellers. Permanent capital supplies the patience. What is genuinely new is the third ingredient: for the first time, the production line inside a services firm can actually be rebuilt, not just trimmed.

If you own a mid-sized firm... accounting, brokerage, agency, MSP... you are the inventory. That is not an insult. It is a seller's market being assembled around you, and the difference between selling as feedstock and selling as a platform piece is the work you do on your own operations in the next two years.

If you run a large incumbent, the GT/CBIZ deal is your preview. The question the Big Four now face is not whether to respond but whether the partnership structure lets them. Partnerships fund growth out of this year's profits. Their new competitors fund it out of permanent capital and price acquisitions on data value. That is not a fair fight, and the firms that changed structure early... Grant Thornton sold its majority stake in 2024... got a two-year head start.

And if you build in AI, notice where the labs are pointing. Embedded equity in roll-ups is the quietest way into the services economy, and the precedent is now set.

In the accounting edition, I wrote that the template... permanent capital plus an embedded AI lab... would be pointed at legal and insurance brokerage within twelve months. Legal is already moving: Eudia is buying legal services businesses with General Catalyst's capital behind it. The lab-embedded version, and insurance brokerage, are still on the clock... and I would now be surprised if either takes the full year.

A few more, loosely dated.

One clause in the Grant Thornton deal is worth watching. Until August 27, CBIZ is allowed to look for a better offer from someone else... that is written into the agreement. I do not expect a rival bid to appear. But if one does, my money says it comes from another PE-backed platform, not from the Big Four.

Within a year, one of the venture-built holding companies takes another listed company private. Amex GBT showed the model can reach that size.

And within two years, one of these holding companies lists or recapitalises at a valuation that forces the market to price "AI roll-up" as its own category... the way Constellation eventually forced software serial acquirers to be understood on their own terms.

I could be wrong on the timing of any of these. The direction feels settled.

Five editions in: legal showed the AI labs picking their law firms, accounting showed the first AI-powered buyers, the labs edition showed them coming for the work itself, and insurance showed what happens when they arrive in a market private equity already owns. This edition is about what those four stories share... the buyers, organised, funded, and now operating everywhere from letting agencies in Hull to the NYSE.

Subscribe and every edition lands in your inbox on Tuesday. If you are inside one of these firms... buyer, seller, or the person running the integration... reply on LinkedIn. I read everything, and the sharpest replies shape the next edition.

And if you are thinking about selling a services firm, or buying one... that is exactly what I do. I run Eilla, an AI-native M&A advisory. Message me.

See you next week.