AI Roll-Ups: The New Way to Build Companies

The startup world tends to romanticize starting from zero, but zero is also an expensive place to start. A new company needs a product, customers, distribution, operational knowledge and, usually, years of experimentation before it knows whether the market really wants what it is selling. An existing business has already solved many of those problems. A thirty-year-old accounting firm may have uninspiring software, but it may also have hundreds of customers who renew every year, employees who understand the edge cases of the industry, and years of accumulated knowledge about how the work actually gets done.

The weakness of many of these businesses is usually somewhere else. Their growth is tied closely to headcount. New customers create more emails, more documents, more reconciliations, more scheduling, more reporting and more people required to keep everything moving. Revenue rises, but organizational complexity rises with it. Eventually the company becomes a machine for turning additional revenue into additional payroll.

That is where AI changes the acquisition thesis. If the buyer can take an existing customer base and gradually reduce the amount of routine human work required to serve it, the same business can begin to produce very different margins and scale in a different way. Long Lake's proposed acquisition of Amex GBT is an unusually large example of this logic. Long Lake has explicitly described its strategy around applying AI to service businesses, and the Amex GBT transaction pairs that capability with a company that already possesses global distribution, customer relationships and a large operational footprint.

This is subtly different from buying a business and installing some new software. The real thesis is that ownership makes deeper transformation possible. A SaaS vendor can suggest a new workflow, but it cannot easily change the customer's job descriptions, reporting structure, incentive system or staffing model. An owner can. That distinction may explain why some investors are becoming interested in owning the service business rather than remaining one layer removed from it.

The temptation is to look at every old-fashioned industry and see an AI roll-up opportunity. That would be a mistake. A fragmented market and mediocre software do not automatically make a good target. What matters is whether the economics of the underlying work can actually change.

The strongest candidates tend to be businesses where a meaningful share of the product is information processing. Accounting and fund administration are good examples because enormous amounts of labor go into collecting documents, reconciling records, checking exceptions, preparing reports and communicating results. Similar patterns appear in insurance administration, compliance, legal support, payroll, benefits administration, corporate travel, property management and many forms of B2B outsourcing. The domain expertise may be sophisticated, but a surprising amount of the daily work consists of moving information through predictable stages and asking experienced humans to intervene when something falls outside the normal path.

Fragmentation also matters. An industry with thousands of small operators gives a roll-up multiple acquisition targets and makes it possible to improve the platform through repetition. Each acquired company brings another set of customers, workflows and edge cases. If those businesses can be moved onto a common operating system, the tenth acquisition should theoretically become easier to integrate than the first. That is where the model begins to compound.

Customer retention is equally important. A recurring accounting, compliance or administrative relationship is much more attractive than a business that must win every customer again each month. AI transformation takes time, so the acquired company needs enough underlying stability to survive while its operating model is being rebuilt. High switching costs can help as well, although they should not become an excuse for poor service.

What I would be much more cautious about are businesses whose value is primarily embodied in physical labor, individual reputation or highly bespoke human relationships. A plumbing company can certainly use AI for scheduling and customer service, but the plumber still has to travel to the house and fix the pipe. A boutique architecture practice may use AI extensively, yet its clients may still be buying the judgment and reputation of a particular architect. AI can improve those businesses, but improvement is different from fundamentally changing their unit economics.

The useful question therefore is not, “Can this industry use AI?” Almost every industry can. The better question is, “If the company doubles its revenue, can AI prevent its operating headcount from doubling with it?” If the answer is yes, there may be real roll-up leverage.

A conventional roll-up often ends up looking like a collection of boxes on an organizational chart. There is a holding company at the top and a series of acquired businesses underneath it. Each business still has managers, finance staff, operations teams, customer service employees and its own collection of software. Some functions may eventually be centralized, but the group frequently retains large amounts of duplication because integration is disruptive and local management wants to preserve autonomy.

An AI roll-up probably needs almost the opposite architecture. The central company should contain a relatively small group responsible for capital allocation, product and engineering, data infrastructure, AI systems, risk and operational design. Beneath that sits a shared operating layer used across the portfolio: a common customer model, workflow engine, knowledge base, permissions system, audit trail and eventually a set of AI agents capable of performing increasingly large portions of routine work. The acquired businesses remain close to their customers and retain the people who understand the industry, but they should not each rebuild the same administrative machinery.

This makes the headquarters less like a traditional corporate center and more like an internal technology platform. Its job is to discover a better way to perform a process once and then distribute that capability throughout the group. If one acquired company develops a better way to handle onboarding documents, reconcile accounts or respond to a certain type of customer request, that improvement should become available to every business on the platform. The economics start to resemble software: development costs are concentrated in the center, while the benefit can be distributed across an expanding revenue base.

That architecture also changes how acquisitions should be evaluated. A potential target is not valuable only for its EBITDA. It may also bring a new customer segment, a new dataset, a particularly strong workflow, a regulatory license or experienced people who understand a difficult part of the industry. Over time, the roll-up is effectively acquiring pieces of an industry and converting what it learns into shared infrastructure.

Talk about AI replacing workers tends to become abstract very quickly. In a roll-up, the mechanism is much easier to see because the buyer can examine an entire workflow and ask where people are actually spending their time.

Consider a service company that receives a customer request by email. An employee reads it, identifies the customer, finds the relevant documents, enters information into another system, checks a rule, perhaps asks a colleague a question, prepares a response, updates the CRM and then schedules a follow-up. None of those individual tasks looks particularly dramatic, but multiply the sequence across thousands of customers and dozens of employees and you have a large part of the company's payroll.

AI does not have to replace the entire job for the economics to change. It can read the incoming request, identify the customer, retrieve the relevant records, prepare the next action and draft the response. A human can then review cases where the confidence is low, the financial impact is high or the situation falls outside established rules. As the system becomes more reliable, the percentage of cases requiring intervention can gradually fall.

The organizational consequence is more important than the automation of any individual task. Instead of having fifty employees executing roughly the same workflow, a company may eventually have a much smaller group supervising the system, resolving exceptions and maintaining customer relationships. The remaining employees become more valuable because their time shifts toward the parts of the job where judgment actually matters.

This is also why the most realistic path is unlikely to involve waking up after an acquisition and firing half the company. Much of the early labor benefit may come through slower hiring, natural attrition and the ability to absorb new acquisitions without reproducing their full administrative headcount. If a roll-up acquires a fifth company and can place much of its back office onto infrastructure that already exists, it may not need to inherit every role that would normally be required to operate that business independently.

That difference compounds. A conventional services roll-up might become more efficient as it grows, but it still tends to accumulate people. An AI-native roll-up is trying to create a situation where every acquisition adds substantially more revenue than organizational complexity.

This is where I think some AI roll-up strategies could go wrong. There will be a strong temptation to raise a large pool of capital, announce an ambitious consolidation thesis and begin buying companies quickly. But if the underlying transformation engine has not been proven, scale simply makes the experiment more expensive.

The first acquisition should probably be treated as a laboratory rather than the beginning of an acquisition spree. The buyer needs to understand the business at a surprisingly detailed level: which workflows consume the most labor, where errors occur, which decisions require genuine expertise, how information moves between systems and which interactions customers actually value. Only then can the company determine whether AI is changing the core economics or merely making a few employees more productive.

A useful early milestone would be demonstrating that one business can grow materially without its operating headcount growing at the historical rate. Once that happens, the roll-up has something much more important than a collection of AI demos. It has evidence that its operating model works.

The second acquisition then tests another question: can that operating model travel? A workflow that works beautifully inside one company may depend on its data structure, management culture or customer mix. The roll-up becomes interesting when the same underlying platform can absorb a second and third business without being rebuilt each time.

Only after those two problems have been solved does aggressive acquisition really begin to make sense.

The obvious risk is simply overpaying. Once investors believe that AI can dramatically improve margins, it becomes very easy to price those future improvements into the acquisition before they exist. Amex GBT's proposed $6.3 billion transaction is instructive because the offer represents a substantial premium to the company's pre-announcement trading price. That does not mean the price is wrong, but it illustrates how much confidence can become embedded in the transformation thesis before the transformation has happened.

Integration is another danger. AI may make information easier to process, but it does not magically eliminate incompatible databases, regulatory obligations, employment contracts, unhappy employees or customers who dislike changes to a service they already know. A roll-up can easily acquire companies faster than it can absorb them, leaving the group with an impressive revenue number and a mess underneath.

There is also a technical risk that gets understated in financial discussions. A system that drafts marketing copy can occasionally be wrong without causing much damage. A system making decisions about client assets, insurance claims, medical administration or regulatory compliance operates under a completely different standard. The more AI moves from assisting employees to actually executing work, the more important permissions, audit logs, evaluation, exception handling and human escalation become. The operating platform has to be designed around failure, not around the assumption that the model will always be right.

The final risk is cultural. In many of the businesses most attractive for roll-ups, employees have spent decades building specialized knowledge. If management arrives and communicates that the objective is simply to eliminate headcount, much of that knowledge can leave before it has been captured. A better transformation probably starts by identifying the strongest operators and turning them into the people who teach the new system how the business works.

One question inevitably comes up: if everyone has access to the same foundation models, where is the defensibility?

The answer probably does not lie in owning a better chatbot. The defensibility comes from everything surrounding the model: the proprietary workflows, accumulated exceptions, customer history, integrations, evaluation data and operational knowledge created by running the business every day.

A roll-up that has processed millions of real transactions in a particular industry can eventually know something that a horizontal AI vendor does not. It knows which situations create problems, which exceptions matter, which customer requests tend to escalate and which decisions experienced employees repeatedly make. That knowledge can be encoded into the operating system and distributed across the portfolio.

This creates an unusual form of compounding. Acquisitions create scale, scale creates data, data improves the operating system, and a better operating system makes subsequent acquisitions more valuable. Whether that loop actually develops will separate real AI roll-ups from ordinary roll-ups that happen to use AI software.

For the past twenty years, the technology industry has largely separated the people who build software from the companies that use it. SaaS made that separation extremely efficient: build one product, sell it to thousands of companies and let each customer figure out how deeply it wants to change.

AI may push some entrepreneurs in the opposite direction. If the largest value comes from redesigning the workflow itself, owning the company gives you considerably more freedom than selling software to it. You can change the process, reorganize the team, centralize functions, alter incentives and decide exactly where humans remain in the loop.

That does not mean AI roll-ups will replace SaaS or traditional private equity. In many industries the operational improvement will be too small, the acquisitions too expensive or the integration too difficult. Some investors will undoubtedly discover that they have simply bought a collection of mediocre businesses and added an AI story on top.

But where the underlying work is information-heavy, repetitive and fragmented, the model deserves serious attention. The opportunity is no longer limited to building software for an industry. It may be possible to buy pieces of the industry itself, turn what those businesses know into software, and gradually build a company whose economics look very different from the businesses that went into it.

That may ultimately be the most interesting part of the AI roll-up idea. It is not really a new acquisition strategy. It is a new way of asking what a company should look like once intelligence itself becomes part of the infrastructure.