AI Roll-ups Are Buying Permission, Not Technology

The trade fits in a tweet. Buy an unglamorous service business with fragmented competition and a 10-15% EBITDA margin. Replace the spreadsheets, the phone tree and the manual back office with AI. Hold the cost base flat while the automated share of the work grows. A year later, run the same business at roughly 30%. People write it as an AI roll-up, an AI rollup or an AI-enabled roll-up. It is the same trade.

One clarification, because the label gets stretched. This piece is about buying service businesses that do not yet run on software. A buyer like Bending Spoons, which acquires companies that are already software and then cuts cost and raises prices, is running an older and different trade under the same word. We track that one separately, deal by deal, in Every Bending Spoons acquisition follows the same playbook.

The money behind it is not theoretical. General Catalyst has described screening around 70 subcategories of the services economy and finding roughly 10 where this works today, inside what its managing director Marc Bhargava sizes at $16T of annual services revenue. Long Lake has bought 30 or more non-tech companies and took Amex GBT private in a $6.3B deal, on the argument that somebody has to turn lab capex into GDP growth. Thrive Holdings raised $2B at a $12.5B valuation for its own roll-up. Hemant Taneja has said their HOA management business doubled free cash flow without cutting headcount.

The framing is deliberately anti private equity. Dylan Patel of SemiAnalysis puts it as a contrast in cost structures: the old playbook buys a company, squeezes it and discards it, while the new one spends heavily up front to modernise the systems and then drops the run rate permanently. His read on the capability is careful. AI cold calling, AI invoicing and AI accounting are not at critical mass, but they are close.

Take all of that as given. One question survives it: if the technology is what creates the margin, why does the deal require buying the company?

Every buyer in this market shops in the same aisle. The frontier models are available to all of them, at the same price, in the same week. So is every agent framework built on top. Nothing in the stack that automates an invoice queue is proprietary to the fund that owns the invoice queue.

That is the uncomfortable implication of Patel's own phrasing. A capability at critical mass is a capability everyone has. If AI invoicing and AI cold calling are good enough in 2026 to reprice a business, they are good enough for that business's competitor, and good enough for the business itself without any change of ownership.

A capability arriving at critical mass is a capability arriving at commodity. Commodities do not produce 15 points of margin on their own.

The engineering is real work and it is not trivial. It is also not rare, it is not defensible, and it is getting cheaper every quarter. If the technology were the moat, the winning move would be to sell the technology to all 400 businesses in the category rather than buy four of them. Nobody is doing that, and the reason is the interesting part.

Bhargava is direct about the target profile. Many of these industries are highly fragmented and old school. They do not buy technology at all, let alone AI technology. They are still working through software and cloud.

Long Lake co-founder Varun Shenoy makes the same point from the other side. He argues AI diffusion is bottlenecked by human curiosity rather than by the technology: people in these businesses have no need to use it, do not know how, and have heard of it from a friend or a cousin. His description of the fix is a description of change management, not of a product. Come in, fit the technology to what they already do, meet them where they are, and make them feel like they are driving.

Read those two together and the deal structure explains itself. You cannot sell software to a business that does not buy software. You cannot sell them the service either, because the service is what they already are. The only remaining way to change how the work gets done is to own the profit and loss and decide.

The equity cheque is not paying for the AI. It is paying for the authority to change a workflow that no vendor could have talked anyone into changing.

That is a genuine asset, and it explains the returns better than the technology does. It also shows exactly where the thesis is fragile. Permission bought with a balance sheet still has to survive contact with the people who do the work, the customers who call the same number they have called for fifteen years, and the specific reasons the legacy process exists. Ownership grants the right to change the workflow. It does not supply the knowledge of which parts can safely be changed.

The sceptics on this are worth more than the boosters, because they are specific.

Rohit Mittal's list of who a roll-up needs has circulated widely, and it is really a list of failure modes. Three people: a deal operator who can source and close without overpaying, an industry veteran who knows why the legacy processes exist and which ones can be touched, and an AI-native technical founder who can tell capability from hype. Most pitches have one or two. The technical founder assumes AI solves it. The deal operator treats it as financial engineering. The veteran knows both are wrong and gets overruled for not speaking VC.

His warning on the exit is sharper still. These are not innovation plays, so future multiple expansion is mostly a pipe dream. Paying up with venture money feels good and is not the same as knowing, in detail, how you will fix the company you just bought.

Chris Vasquez's recap of a founder meetup with Keith Rabois lands in the same place from experience rather than structure: most AI roll-ups will fail or produce lacklustre returns, because founders overestimate AI efficiency gains as a business model. Absent roughly a 50% cost reduction, you need a serious growth engine instead. Elsewhere the criticism is blunter, aimed at leveraged capital structures and at operators who have never shipped an AI product or completed a single integration. Someone joked about raising $300M to roll up the other AI roll-ups, and it travelled, which tells you where the mood sits.

Now read the trade from the seller's chair. The buyer is underwriting the gap between what the business earns today and what it earns once its workflows are automated. That gap is the entire return. It is computed on your business, and it is paid to them.

Selling into an AI roll-up prices your company on its pre-AI numbers and hands the buyer the upside you were going to create anyway.

If permission is the scarce input, an owner already has it. Nobody has to be acquired to get authority over their own workflows. What is missing is usually the sequence, so here is the one the funds are executing, run in-house.

1. Price your own gap first. Not "where could we add AI". List the tasks that consume the hours: the calls that go to voicemail, the quotes never followed up, the invoices chased by hand, the month-end close. Put a cost and a volume against each. That number is what a buyer would underwrite, which makes it the number worth knowing before anyone calls.

2. Start at the perimeter, not at the craft. The margin does not come from automating the thing you are good at. It comes from the surrounding work that nobody has time for at 6pm: intake, scheduling, follow-up, chasing, filing, reconciling. This is also the honest reading of Patel's list. Cold calling, invoicing and accounting are the perimeter, and that is precisely why they are close to critical mass.

3. Hold the cost base flat and let volume move. This is the lever the funds actually pull, and Taneja's version of it is the one to copy: free cash flow doubled without cutting headcount. Margin from capacity rather than from redundancies is both the more durable version and the only version your team will help you build.

4. Name an owner for each workflow. An AI workforce executes the work. It does not become accountable for the outcome. Someone by name still owns the result, decides what escalates and reviews the output on a cadence. We wrote the long form of this argument in Companies Need an Owner of the Problem.

5. Keep the data and the relationship. Shenoy's own framing is that the services layer sits at the top of the AI stack because it owns the end-customer relationship. That is exactly what transfers in a sale. It is also exactly what compounds if it does not.

The roll-up thesis is probably right about the destination. The services economy is where the capability lands next, most of these businesses will be run very differently in five years, and someone does have to convert the capex into output. None of that is in dispute.

What is in dispute is who captures it. The pitch presents acquisition as the mechanism that makes the transformation possible. It is really the mechanism that makes it possible for an outsider. The buyer needs to own the business because they cannot otherwise get permission to change it. You do not have that problem.

So the test before you take the call is a short one. If a buyer believes they can add 15 points of margin to your business in twelve months, using capabilities you can rent by the month, then they are not buying a capability. They are buying the twelve months, and they are buying them at a price set before the work is done.

Do the twelve months yourself and the offer does not go away. It just arrives priced on the business you actually built.

Selling is still a perfectly good outcome. It should be a decision about what you want to do next, not a transfer of the upside to whoever asked first.