In the AI economy, three business models have emerged as especially compelling:

• AI Roll-ups buy service businesses, then deploy AI across them.
• Service as a Software companies, which we previously wrote a primer on, provide software solutions that produce outcomes.
• AI-Native Services build or operate a service business whose employees use AI to deliver the work more efficiently.

Each offers a different way to capture the value created as AI begins performing work that was once done entirely by people.Each comes with distinct advantages, limitations, and risks, and all three have the potential to produce massive companies. In this piece, we will provide a primer on how we’re thinking about AI roll-ups and AI-Native services:

A Primer on AI Roll-Ups

Buy the service businesses, then deploy AI across them.

AI roll-ups have emerged from a fairly simple observation: the economic opportunity created by AI may be much larger in services than in software.

For the last twenty years, venture capital largely followed a software playbook. Find an industry with inefficient workflows, build software that makes those workflows better, and sell that software to the companies doing the work.

AI creates another possibility.

• Instead of selling the accounting firm better software, buy the accounting firm.
• Instead of selling the call center an AI agent, own the call center.
• Instead of selling an MSP software that makes each technician more productive, acquire the MSP and capture the productivity improvement yourself.

General Catalyst has been one of the most aggressive proponents of this idea. Marc Bhargava, who leads much of the firm's work here, has framed the opportunity around the relative size of the two markets: roughly $16 trillion of global services revenue versus roughly $1 trillion of software revenue.

The firm's argument is that SaaS captured only a small portion of the economic activity inside many service industries because these businesses often don't buy much software, are fragmented across thousands of operators, and are difficult to penetrate through a conventional enterprise sales motion.

So they changed the question, rather than asking, "How do we sell AI into this industry?"

Ask, "What if we owned the industry participants through which the AI gets deployed?"

Imagine that you build an AI system that can make an accountant twice as productive. If you sell the system to an accounting firm for $20,000 per year, you capture the software budget. The accounting firm captures most of the economic benefit created by the software.

But if you own the accounting firm, the economics are different. The productivity gain appears directly in your P&L. An accountant who could previously serve 50 clients might now serve 100. Revenue per employee rises. Margins can expand. Capacity increases without requiring labor to increase proportionally.

You are no longer selling the productivity improvement….You own the productivity improvement!

There is another problem AI roll-ups are designed to solve: distribution.

Some of the industries most susceptible to AI are precisely the industries where selling new technology is hardest.

There may be thousands of small operators, little internal technical talent, tiny software budgets, old systems, and strong resistance to changing workflows.

Bhargava has described this explicitly. General Catalyst evaluated roughly 70 service industries and focused on sectors where it believed 30% or more of the underlying work could ultimately be automated. Yet building the technology was only part of the problem, getting that technology into the hands of customers could be equally difficult.

The answer was to acquire the distribution rather than build it.

A normal AI startup builds a product and then spends years acquiring customers. An AI roll-up can acquire a company and receive the customers, employees, workflows, historical data, contracts, and industry knowledge at the same time.

Crescendo is an example. The AI customer-experience company acquired PartnerHero, an established customer-operations outsourcing company. With the acquisition came more than 200 customers and a combined business with over $50 million in ARR. Crescendo could then deploy its autonomous customer-support technology across an installed base that otherwise would have taken years to acquire organically. General Catalyst later reported that Crescendo was automating as much as 80% of customer-support work and achieving margins of roughly 60–65%.

Long Lake is applying the model to property and community management. General Catalyst says the company has acquired more than 18 businesses; in HOA management, its technology has produced 25–30% productivity improvements while its AI-powered sales operation increased new-customer pipeline roughly tenfold.

Thrive Capital has pushed the idea even further through Thrive Holdings.

In December 2025, OpenAI took an ownership stake in Thrive Holdings and agreed to embed researchers, product people, and engineers directly inside its businesses, initially focusing on accounting and IT services. Joshua Kushner described the underlying philosophy as moving technological transformation from "the outside in" to "the inside out." His argument was that ownership creates the right incentives while giving AI teams direct access to domain experts and real-world data.

That structure is already producing an interesting example inside Crete, Thrive's accounting platform. OpenAI and Thrive engineers built tax agents alongside accountants working across Crete's network of more than 30 firms. During the 2026 tax season, the system processed 7,000 returns. OpenAI reported that one senior accountant who had spent approximately 180 hours preparing returns the previous year spent only 15 hours doing so after the system was deployed, allowing her to spend more time advising clients and taking on additional work.

That example reveals another, less obvious component of the thesis.

Buying the business does not merely buy revenue. It buys the environment in which the AI learns.

• The accountants correct the tax agent.
• Those corrections become data.
• The data becomes evaluations.
• The evaluations improve the agent.
• The improved agent makes the accountants more productive.

The more accounting firms the company acquires, the more practitioners and workflows enter the system, producing more feedback that can improve the technology across the entire network.

The Three Challenges of AI-Rollups.

Most objections to AI roll-ups are really objections to roll-ups. Integration is hard, acquisitions distract management, employees resist change, customers dislike being handled by a system instead of a person. All true, all true since the 1960s, and all priced in by anyone who has done this before.

The three challenges worth taking seriously are the ones specific to the AI part of the thesis.

They stack, and each one has to clear before the next matters:

The AI-Roll-ups Challenge #1: The productivity gain may never reach the financials

It is easy to look at a workflow and conclude 40% of the labor could be automated. But eliminating 40% of the work does not eliminate 40% of the employee. The remainder tends to be judgment, customer contact, exception handling, and regulatory accountability, the parts that don't compress.

So you land in an awkward middle state: AI reduces the work humans perform without removing the humans. If an accountant goes from eight hours to five, you capture nothing unless those three hours become more customers served, fewer people employed, or a higher-value advisory product sold. Otherwise the improvement is real and invisible, it shows up in how the work feels and nowhere in the P&L.

This is where process variation does its damage, too. AI works on standardized workflows, and acquired businesses that look identical from outside run very differently inside. Every difference is an exception, every exception needs a human, and enough exceptions turn the automation layer into an expensive suggestion.

The question is not how much work AI can automate. It is how much cost actually comes out, or how much revenue actually goes in.

The AI Roll-ups Challenge #2: If the gain is real, it may not stay yours.

Suppose it works. Accountants are 40% more productive. Now ask where that capability came from. If it came from OpenAI, Anthropic, or Google, every independent firm in the market gets the same lift on roughly the same schedule, and independents don't have an acquisition premium to earn back, so they can pass the savings to customers faster than you can.

AI adoption can improve a business without creating a moat. In a fragmented, price-competitive services market, a productivity gain that everyone receives becomes a price cut that everyone gives.

The same dynamic runs on the buy side. The more credible the thesis becomes, the more capital chases it, and the more you pay for the businesses required to execute it. Buy $10M of EBITDA for $80M and grow it to $15M and you've created real value. Pay $120M for the same business because four well-capitalized buyers are working the sector and you haven't. The technology didn't get worse; the entry price did.

So the thesis cannot be "we buy businesses and introduce them to AI." It has to be something the model access doesn't confer, proprietary workflow data, real distribution, purchasing scale, a structural cost advantage, with AI sitting on top of it.

The AI Roll-ups #3: If it stays yours, the market may not pay for it

Roll-ups implicitly hope for services economics at software multiples.

But acquirers value what actually drives revenue. If revenue still depends on thousands of employees, local operations, relationship-based retention, and a continuous acquisition pipeline, buyers will underwrite it as a services business no matter how good the margins get.

Using software to improve a services business is not the same as becoming a software business. The first can still produce an excellent company. It just may not clear 12x on the way out, let alone 20x.

You can get the automation right, keep the advantage, expand margins exactly as modeled, and still miss the return, because the return was underwritten on multiple expansion that the buyer was never going to grant. Buying under one valuation framework and exiting under another is a decision made at entry, not at exit.

The risk you run is that AI Rollups business may be valued on exit at a much lower multiple than you anticipated.

All three challenges collapse into one question: are these businesses actually better after they're acquired, or has the company simply assembled assets?

Does revenue per employee rise post-integration? Do margins improve structurally, or only because the mix shifted? Does organic growth accelerate, or is all growth purchased? Does each integration take less time than the last? Does the technology get better as the network grows, and would a buyer agree that it does?

Roll-ups can appear to work for years, because acquisitions themselves produce growth. Revenue rises because revenue was purchased. EBITDA rises because EBITDA was purchased. The headline numbers look healthy right up until the acquisition engine stops, and then you find out whether there was a machine underneath it.

For the most part, I believe there is a real place for AI roll-ups.

I don't think they are a gimmick, but I also don't think they are a cheat code.

They are probably best understood as a new type of business sitting somewhere between private equity, a technology company, and an operating company. The model makes sense because ownership gives you something software vendors do not have: the ability to force adoption, redesign the workflow, capture the full productivity gain, and use the underlying business as a training environment for the technology.

But none of that guarantees an exceptional outcome.

My base case is that we will see a wide distribution. Some AI roll-ups will become genuinely great businesses because they find an industry where the work is highly automatable, the businesses are easy to standardize, the technology gets better as the network grows, and the acquisitions are made at sensible prices. Others will still become perfectly good businesses, better margins, better operations, better customer experiences, but ultimately look like well-run services companies rather than software companies. And some will fail because they overpay, overestimate what AI can automate, or discover that integrating twenty businesses is much harder than putting twenty businesses into a spreadsheet.

That dispersion is not evidence that the model doesn't work. It is what we should expect from a new business model.

Zooming Out: Considering AI Roll-ups in Context

The mistake would be treating AI roll-up as the investment thesis itself.

A roll-up is only the structure. The thesis still has to be proven at the company level.

• Can you buy the businesses at the right price?
• Can you actually change their economics?
• Can you repeat that improvement across acquisitions?
• Does the advantage compound rather than disappear?
• And when the acquisition engine eventually slows, is there a meaningfully better business underneath it?

I suspect the answer, across the category, will be: sometimes.

And that's enough for AI roll-ups to matter.

They don't need to replace SaaS or become the dominant model for applying AI to services. They only need to work exceptionally well in the industries where ownership, automation, and consolidation reinforce one another. In those industries, I think they will.

A Primer on AI-Native Services

Build the service business from scratch around what humans and AI are each best at.

AI roll-ups start with an existing business and try to transform it. AI-Native services start with a blank sheet of paper.

Instead of asking:

How do we take this accounting firm and automate 40% of what its employees currently do?

The question becomes:

If we were inventing an accounting firm today, knowing everything AI can now do, how would we design it?

How many accountants would we hire? What would they actually spend their time doing? Which work would never touch a human, and which decisions require judgment? What information would the system collect from every engagement?

Those questions lead somewhere very different from adding AI to an existing services company.

Every traditional services firm is partly a product of the technological constraints that existed when it was created.

Accounting firms employ armies of junior accountants because humans had to reconcile transactions. Law firms employ large associate classes because humans had to read, research, draft, and review. Insurance brokerages employ people to collect information, fill out applications, talk to carriers, compare policies, and follow up. Consultancies employ analysts to gather data, build models, and prepare presentations.

None of these structures emerged because they were the theoretically perfect way to solve the customer's problem. They emerged because human labor was the available technology.

When technology changes, there is no reason to assume the business structure should stay the same.

Forerunner Ventures framed this well in 2024: legacy service providers carry several structural disadvantages when adopting AI. Their economics are tied to billable human labor, their org charts were built around that labor, their technical capability tends to be weak, and their cultures evolved around a slower cadence. A new entrant inherits none of it. It decides from day one which work belongs to software and which belongs to people.

A traditional law firm implementing AI asks how to make its associates more productive. A new firm can ask why the associate role exists in its current form at all. A traditional brokerage asks how AI can help brokers process submissions faster. A new brokerage asks what the smallest amount of human judgment is that still delivers an exceptional insurance-buying experience.

Harper became the broker

Harper did not build AI software and sell it to insurance brokers. The founders explored that approach and abandoned it. They became the broker instead.

Harper owns the customer relationship and designed its workflows around AI from the beginning. AI reads applications, routes submissions, follows up with underwriters, processes quotes, and matches risks against a broad carrier network.

Humans stay where expertise, relationships, and judgment matter. Emergence, an investor in the company, says Harper has served thousands of businesses since its 2024 founding and built its workflows "from the ground up around AI."

The point is not that Harper uses AI. Everyone will use AI.

The point is that the organization itself was designed assuming AI exists.

Emergence describes Harper as running with an unusually small core team while operating what would historically have required several separate systems: CRM, support infrastructure, transaction software, call-center agents, carrier matching, coding agents. Humans are concentrated on the judgment-heavy part of the workflow rather than distributed across the organization doing work software can increasingly handle.

That is what a first-principles services company looks like. You don't take the traditional employee-to-customer ratio and improve it by 20%. You ask why the ratio exists.

The human/AI division of labor gets cleaner

There is a tendency to think about AI-Native services as services businesses with fewer people. That framing is incomplete. The more interesting question is what kind of people remain, and what they spend their time doing.

Take an AI-native accounting firm. The accountant should not spend three hours importing transactions and moving numbers between systems; the machine does that. The accountant investigates unusual transactions, understands the economics of the client's business, explains what is happening, advises management, and takes responsibility when something is wrong.

AI moves humans up the value chain.

And because you are building from scratch, you hire for that job from the beginning. A legacy firm might have 1,000 people whose skills, compensation, career progression, and identity were built around the old workflow. A startup can hire 100 people selected specifically to thrive in the new one.

You aren't teaching an old organization a new production function. You're hiring people for the new one.

Rebuild the workflow

When technology enters an existing company, the instinct is to automate the process that already exists. Step one used to be performed by a person; now AI performs step one. Step two used to be performed by a person; now AI helps with step two. And so forth.

Sometimes the better answer is that steps one through six should not exist.

Emergence makes this point explicitly in its AI-native services playbook. Domain experts are invaluable because they understand how the work gets done, but they naturally tend to recreate the existing process. Product and engineering have to absorb that expertise while still asking what the new AI-Native process should be from first principles, rather than encoding the historical workflow into software.

The first automobile was not a faster horse carriage. The spreadsheet was not a faster paper ledger. The internet was not a catalog sent through the mail. New technologies create the most value once companies stop imitating the system they replaced.

Hanover Park sells the outcome

Fund administration has historically required large teams of accountants performing reconciliations, reporting, treasury, and other repetitive operational work.

Hanover Park was built as an AI-native fund administrator rather than as software sold to fund administrators. It combines a proprietary ledger, AI automation, and human fund accountants to deliver the finished service — what Emergence describes as the speed and economics of software with the coverage and accountability of a full-service administrator.

Again, the interesting part is the organizational design. The customer does not decide when to use AI. The accountant does not buy a tool and redesign their own organization around it. Hanover owns the problem.

That lets the company continuously adjust the ratio between software and people behind the scenes while the customer keeps buying the same thing: the outcome.

You don't have to remove the human

This is where I would distinguish AI-Native services from the purest version of Service as Software.

Sequoia's Julien Bek argues that the real opportunity is selling the work rather than the tool, moving from "copilots" toward "autopilots." His observation is that businesses spend far more on the work itself than on the software used to perform it.

The economic observation is right. But I don't think every great company has to end with the machine performing 100% of the service. Some categories will. Others may settle at 80/20, or 60/40, or even 30/70.

What matters is whether that combination creates a structurally superior production system. A law firm where each attorney responsibly handles five times as many matters is an extraordinary business even if lawyers never disappear. A brokerage where the machine does the administrative work and brokers concentrate entirely on judgment and relationships is extraordinary. A consulting firm where five people produce what historically required fifty may still look like a consulting firm to the customer.

That is fine. The objective isn't to remove humans. It's to remove the historical relationship between output and human labor.

The Two Challenges of AI-Native Services.

The AI-Native Service Challenge #1: Sometimes starting from scratch is the wrong starting point

There are also industries where the purity of starting AI-native simply does not make economic sense.

Some service businesses sit behind regulatory approvals, licenses, certifications, payer relationships, government contracts, carrier relationships, or other forms of infrastructure that can take years to build.

In those markets, the incumbent business may contain something much more valuable than its existing workflows.

It may contain permission to operate.

If entering an industry organically requires years of licensing, regulatory approval, contracting, or relationship building, then starting with a blank sheet of paper has a very real cost.

You may design the perfect AI-native operating system and still be unable to put meaningful volume through it or, in certain cases, even begin to operate. In those situations, acquisition can become the more rational path.

The company is not necessarily buying the legacy business because it wants the legacy operating model.

It may be buying the regulatory infrastructure, customer relationships, licenses, contracts, and credibility required to participate in the market at all.

Then the problem becomes:

Buy access to the market, but rebuild the operating system underneath it.

In industries with low barriers to entry, I generally prefer starting from first principles and building the service exactly as it should exist today.

But in industries where the right to operate has taken an incumbent ten or twenty years to assemble, acquiring that infrastructure may be far more sensible than recreating it from scratch.

The question is therefore not always:

Should we build or buy?

It is:

Which parts of the old business are genuine legacy baggage, and which parts represent assets that would be prohibitively slow or expensive to recreate?

If most of the value sits in the latter, buying first and rebuilding second may be the more rational approach.

The AI–Native Services Challenge #2: Margin Compression Due to Competitive Pressures

Offshoring was the last major change in the production function of professional services, and the arbitrage was enormous. TCS and Infosys became giants, and both still operate in the low-to-mid twenties: TCS reported a 25% operating margin for FY26, its highest in four years, and Infosys ran at roughly 21% over the same period (company results, FY ending March 2026). Two decades of a structural cost advantage did not translate into the provider keeping the advantage.

Why not? Because once a cheaper way of producing the service became understood, the savings stopped belonging exclusively to the provider. Competitors entered. Customers negotiated. Indian wages rose. Prices adjusted. Some of the productivity gain stayed with the provider; a meaningful portion flowed through to the customer.

AI-native services may face the same mechanism with a much shorter fuse.

Suppose an AI-native accounting firm can initially operate at 70% gross margins because it needs a fraction of the labor of a traditional firm. What keeps those margins at 70%? Another AI-native firm can see the same opportunity, reach many of the same foundation models, and offer the service for less while still earning attractive margins. So can a third. Incumbents eventually adopt the technology too. The production advantage remains real, but the market price starts moving toward the new cost structure.

That may be the most important lesson from offshoring. A dramatic reduction in the cost of producing a service does not mean the provider captures the reduction. Competition determines who captures it.

If AI lets an accounting firm produce for $30 what historically cost $70, the question is not whether the company can earn extraordinary margins at the start. It is what prevents the market from turning a 70%-gross-margin AI accounting firm into a 45%-gross-margin AI accounting firm.

There may be good answers: brand, trust, customer experience, distribution, proprietary data, switching costs, specialization, first mover advantage, regulatory infrastructure, or an operating system that genuinely compounds with scale.

But if the answer is nothing, AI has created a much better services business rather than a fundamentally different kind of business. That can still be a very good outcome. It is simply a different one.

Zooming Out: AI-Native Services in Context

I find AI-Native services more compelling than roll-ups.

A roll-up begins with an artifact of the past. It buys the employees, processes, technology, customer contracts, pricing model, culture, and organizational structure that evolved under the old production function, and then tries to transform them. There are good reasons to do that: immediate distribution, revenue, data, industry expertise. But you inherit the constraints along with the assets.

An AI-Native services startup gets almost none of those advantages on day one. It has to earn every customer, build trust, develop the domain expertise, and construct the operating infrastructure itself. That makes the beginning harder. What emerges can be considerably cleaner.

No legacy org chart to defend. No existing workforce whose economics depend on the old workflow. No ten-year-old technology stack. No acquired culture. No customer contract saying you must deliver the service the old way. No step preserved because that is how it has always been done.

You start with one question:

Given the technology that exists today, what is the best possible way to produce this outcome?

Then build the people, software, pricing, workflow, and organization around the answer.

For decades, technology companies built tools for service businesses while the underlying architecture of those businesses remained largely unchanged. AI gives founders an opportunity to revisit the architecture itself.

Not: how do we build better software for the accounting firm?

And not: how do we buy accounting firms and make them more efficient?

But: what should an accounting firm actually be now?

I suspect some very large companies will come from answering that question correctly.

The Bottom Line

AI is expanding the venture opportunity beyond software by changing who captures the value created by automation:

• AI roll-ups acquire existing service businesses and transform their economics.
• Service as Software sells AI-delivered work into existing businesses.
• AI-native service companies rebuild the service provider itself, designing the organization from the beginning around what humans and AI each do best.

All three models can produce enormous companies, but the structure alone is not the thesis.

The real questions are whether AI materially improves the underlying economics, whether the company can retain that advantage, and whether the advantage compounds as the business grows. The winners will not simply add AI to a traditional service, they will rethink how the service should be delivered altogether.

We believe some of the defining companies of the next decade will look less like conventional software businesses and more like entirely new kinds of accounting firms, law firms, brokerages, consultancies, and other service providers. They will sell the same outcomes customers have always needed, but produce them through a fundamentally different combination of software and human expertise.

If you are building an AI-native service business, or thinking deeply about one, we would love to talk. Feel free to send us an email at hello@behindgeniusventures.com

‍