As part of Networked Conviction, the paid portion of my newsletter, I set out with the intention of revolving my thinking around three different buckets: (1) portfolio updates on the companies I’ve invested in, (2) my own version of Request For Startups, and (3) general investing ideas (frameworks, templates, etc.)

After two editions from my idea journal and one with a portfolio update, I’m finally doing my first request for startups. The insights around this idea come from existing portfolio companies that have already done portions of this, companies I’ve seen but didn’t invest in and what I thought should have been differently, as well as legitimate greenfield I see in the market for someone with passion and a plan to make a difference.

The request? AI rollups that don’t suck. The idea of a rollup has taken the world by storm, but the vast, vast majority of pitches I hear for AI rollups typically have a fundamental misunderstanding of how rollups work… or how AI works… or both. I recently articulated to a friend what the formula for an AI rollup that doesn’t suck ought to look like.

So here we go!

What Is An AI Rollup?

For those of you who don’t know (or, in other words, don’t have a friend in business school), an AI rollup is the idea that, rather than building “AI for accountants” and then selling it to accounting firms, you should just go buy accounting firms and force them to use AI to be more efficient.

I’ve been writing about rollups since 2023 and investing in them since 2019 (2017 if you count the first time I bought stock in Constellation Software). What strikes me as interesting about AI rollups is that its a perfect cornucopia of the things that attract all the smartest people I know, but most of them are doing all the wrong things.

First? The rollups.

Serial acquisition is a common playbook. Many companies have done it well. For Contrary Research, I helped write a full report that, in part, includes a whole overview about serial acquisition as a strategy. Be sure to check it out.

Dozens of companies, from [[Constellation Software]] to [[Danaher]] or the great [[Berkshire Hathaway]] itself have made their livelihoods being good at acquiring a large number of things. Constellation has made ~750 acquisitions, Danaher maybe 450, and Berkshire is ~100. These are in contrast to large private equity firms that will typically make ~1-3 new platform acquisitions a year and maybe a few bolt-ons.

A rollup has less to do with any one company acting as a “platform” that you bolt other pieces onto and more about a large number of similar looking businesses that, you believe, can be automated.

That’s where the AI comes in.

AI and, in particular, LLMs, are uniquely good at taking unstructured data and turning it into structured data. When you look at a huge swath of inefficiencies in various business workflows, the lack of data structuring is a huge part of the problem. Ineffective marketing campaigns, missing potential customer leads, massive overhead for clerical processes, repetitious data review; it’s all unstructured data that we throw humans at and pray to whatever God we worship that it works.

Leveraging AI can optimize marketing campaigns, identify high-quality customer leads, automate clerical processes, and remove the need for repeating data review. It’s perfect!

But here’s the rub. Small businesses, as [[Brent Beshore]] likes to say, are small for a reason. It’s not by choice. Typically, its because the people running them lack the knowledge, expertise, comfort, or ambition to drive efficiencies in the business. So when you build AI for these people and try to sell it, you run face first into that lack of knowledge, expertise, comfort, or ambition.

What’s the alternative? AI rollups!

Enter the AI Rollup

The core idea behind an AI rollup is that, by being the operator, you can (1) avoid a lot of the distribution friction inherent in selling technology to slower-moving incumbents, and (2) capture more of the margin by improving the underlying business vs. convincing them to pay you for that value creation.

In our Contrary Research piece, we laid out a bunch of examples of how some firms are starting to execute this playbook:

“[[Slow Ventures]] has formalized what it calls the “Growth Buyout” strategy. Will Quist, a partner at Slow, argues that in many legacy sectors the upside from selling software is “not worth the squeeze”. Sales cycles are long, pricing power belongs to entrenched oligopolies, and customers are lenient about paying. Slow’s answer is to buy a legacy operator outright, arm it with proprietary software, and let the improved cash flow finance the next acquisition. In another example, Thrive Capital created Thrive Holdings, a dedicated $1 billion roll-up vehicle to invest in everyday industries, in 2024. Its model is to help operate these businesses directly and use their cash flow to fund further acquisitions. Thrive has already backed Crete, an accounting platform, and Long Lake, a consolidator of homeowner association managers. General Catalyst (GC) announced its incubation of HATCo (Health Assurance Transformation Company) in October 2023, which acquired Ohio-based healthcare system Summa Health in January 2024. HATCo is a part of GC’s wider Creation Strategy, towards which the firm has raised $1.5 billion. The strategy is designed to partner with founders to build new companies from scratch, often in combination with an “AI-enabled roll-up model”. Other investors, including 8VC, Khosla Ventures, a16z, and Elad Gil have explored similar approaches across various industries. Most AI rollup coverage highlights the venture firms pitching the idea, which makes it easy to cast them as the modern Berkshire Hathaways. However, in most cases, these funds are writing ordinary equity checks and moving on; more marketing than execution. The actual heavy lifting of choosing markets, picking between buying, building, or selling, and then threading software into messy workflows to build an actual company sits with founders.”

As this AI rollup approach has become the “hot new thing” I’ve met dozens of people who want to do some version of this. But, despite the excitement about AI and the logic around an AI rollup, I believe there will be dramatic amounts of capital destroyed here, similar to ecommerce rollups before it where companies like [[Thrasio]], Tiny, OpenStore, etc. raised ~$15B in capital, almost all of which has gone or will go to zero.

Why? How can a category with so much hype, supercharged by such compelling trends like the silver tsunami of retiring business owners and the incredible capability of tools like AI; how could something like that fail? I boil it down to three issues.

The Personal Layer

Most small businesses are boring. Most service businesses are boring. What I see among the people pursuing this vision is people who are much more excited about the cutting edge of technology than they are about inventory turns in roofing tile or zoning codes for architecture services. The inherent truth in rolling up a business is that, if you’re successful, you’re going to be spending 10-20 years thinking about those businesses. And most hot-shot AI-native youngsters don’t want to do that.

Now, I completely empathize with that. If you paid me millions of dollars, I still wouldn’t want to be the CEO of Squatty Potty. There are things I just don’t want to spend my life doing. To each their own. But, first and foremost, as a founder you have to want the consequences of what you want, and one of the consequences is being obsessed with the category you’re invested in.

One of the logical fallacies I see founders fall into is that they believe their intellectual invigoration will come from the diversity, rather than the individual vertical. “We’ll invest in a broad swath of operating verticals and I’ll get to see a diverse basket of interesting problems to solve.” It’s the modern optionality bet that comes from doing banking / consulting out of undergrad.

But here’s the rub that brings me to my second point. Practically speaking, there is an inherently logical reason why a diverse basket of assets is better than a concentrated basket of in-kind assets. In fact, the more diversity of the underlying operating basket, the more likely there is failure lurking.

The Practical Layer

Practically speaking, these businesses are very messy. And for an AI rollup to be successful, it requires a very deliberate thesis that takes into account that messy practicality. I have met founders who want to acquire law firms who (1) don’t know what multiple of earnings law firms are typically bought at and (2) their hypothesis for driving efficiency is to “make all the lawyers use Claude.” I wouldn’t describe that as well-thought out.

But beyond just simple heuristics like “small businesses are hard,” the practicality goes to a deeper layer of understanding the underlying market dynamics. There is an ocean-sized graveyard of learnings in this regard across a broad swath of both small acquisitions and niche technology companies.

I’ll give you one example. There was once a company I spent time with that was building a marketplace for flatbed freight; basically delivering things too big to fit in storage container style freight. His hypothesis was there are buyers of freight shipment and providers of freight, but the pricing is opaque and confusing. Build a central market maker to bring transparency to pricing and efficiency to booking. Simple, right? Wrong.

Here’s the problem. Flatbed shipping, in particular, runs at a net profit margin of ~3%. Razor thin, even relative to other types of trucking which are already low margin. Why is it so low? Couple reasons: (1) the trucking itself requires bespoke equipment (e.g. step-decks, lowboys), (2) revenue per mile is inconsistent given the broad variance of carrying, (3) load matching is harder, and (4) driver costs are elevated given the special skills required. Hence, the incredibly low margins.

Guess what happens when you introduce higher visibility into price comparison? You drive that margin into the ground. There is enough competition that hides in that opacity that operators in the space want prices to be confusing and hard to determine. It’s the only way they make margin. So, obviously, that marketplace failed.

This is just one of countless examples of where what feels like the “low-hanging fruit” of efficiency is actually an existential threat to a category. People like to have lawyers they can blame, accountants they can panic to, therapists to listen. Automating away every human interaction within those categories is not the recipe for success that this year’s HBS graduating class believes it to be.

That doesn’t mean there aren’t opportunities in those categories. But after a failure to map personal preferences to the category you choose, the next most significant driver of failure will be a practical mismatch between the efficiency hypothesis and the underlying market reality.

That’s where the final element comes in; what is the actual product of an AI rollup?

What Is The Product?

There is framework for understanding where serial acquirers thrive. That framework comes in understanding what is the underlying product. In my perspective, there are three product playbooks around acquisitions more broadly that have some solid proof points:

• (1) Acquisition as the Product: (Berkshire Hathaway, Constellation, TransDigm) These are companies who have put all their thought and optimization around the top-of-funnel acquisition process. For Berkshire, its business quality. For Constellation, its business fit. But post-acquisition, they do almost nothing to the business themselves. They trust they’ve adhered to their acquisition product and then let time do the rest.
• (2) Margin Expansion as the Product: (KKR, Apollo, Bain Capital) This is the traditional PE playbook. You acquire a company for 10x earnings when earnings is Y and then, even if revenue doesn’t grow, you carve fat out of the business so that you increase Y by 20%, which allows you to sell it a few years later for the same multiple and make 20% on your money. Never mind that that harvesting usually ends up killing the company in the long-run.
• (3) Operations as the Product: (Danaher, Metropolis, Heico) These are companies that are not in the business of a huge volume of acquisitions; they’re focused on an incredibly high quality. The goal is to get to maximum scale with minimal capital. Companies that know that what they’re good at is operating a business, growing it, driving increasing profitability. They’re not harvesting the margin and they’re not just sitting and waiting. They know how to run a company better than the average bear.

None of these are necessarily bad “products” (although I would say bucket 2 is headed for extinction if it doesn’t have some kind of renaissance). The question is more a function of what type of acquisition is a good fit for each model.

One of the problems with the rising generation who want to do an AI rollup is they want the biggest basked. So the think they can do lots of different verticals. Fair enough; that would typically fall into Acquisition as the Product. Only problem is that almost completely leaves out the AI optimization thesis, which is a better fit for Operations as a Product.

Trying to run an AI optimization while also being an exceptional acquirer at the top of the funnel is extremely difficult. I’m honestly not sure there are any successful examples. Danaher is maybe the closest; they have almost a religion around operational improvements and a robust acquisition pipeline. But even then, their universe is super narrow focused on precision instruments. A wide set of verticals AND an AI play? I don’t think that will go well.

The Request For Startups: AI Rollups That Don’t Suck

So… what’s my request? AI rollups that don’t suck. What does that look like?

• (1) Founders With Earned Conviction: I’ve written plenty about clarity of thought and earned conviction. The biggest fear I have with any AI rollup is a founder / market mismatch. If there is not a clear story and vibe of complete love of a category, I struggle. Even more so than selling software to a vertical, this is OPERATING a vertical. It requires so much commitment to being the thing. And without genuine conviction and a fundamental earned insight, you’ll struggle no matter how fruitful the category is.
• (2) Deliberate Operating Principles: This is both defined by AI and completely independent of AI. First, there has to be an underlying operating playbook that you believe will work and that you’ve falsified as best you can with real operating data. Whatever vertical you’re acquiring into, why will what you’re doing be better than what generations of comparable operators have been doing for decades? In some cases, just being comfortable with ecommerce or digital scheduling could be a huge operational lift. Other times, it won’t scratch the surface. There needs to be an understanding of HOW these basket participants you want to scoop up are operating today and how they should operate in the future. Anyone who pitches me on this idea that hasn’t talked to at least dozens of potential acquisition targets, I’m deeply skeptical of.
• (3) Well-Defined Capital Model: Finally, something we haven’t really touched on but is actually the lynchpin of the whole enchilada. How will you capitalize the engine? The business school bros are quick to say “venture funding,” then hesitate; doubting themselves. “Bootstrapped.” Their mental faculties depleted. In reality, there are dozens of models with which you can weave a capital web around your ambitions. Venture funding, sure. SBA loans, sometimes. Earnouts. Captive funds. Fundless sponsors. Asset-backed lending, depending on what you’re buying. Capital can come from lots of places, but it can also come from none of the places. Understanding the structure of where the capital comes from is the lifeblood of an acquisition engine. The unappreciated secret to Buffett’s capital engine was insurance float; cash sitting around letting you make acquisitions. What’s your capital lifeblood?

There are some models I find very compelling.

I first invested in [[Teamshares]] in 2019 when it was just three co-founders. The company has since grown 20x in valuation and is headed for the public markets this year. I would describe them as one of the more complicated models I’ve seen. They’ve built optimizations around the acquisition process, the transition process, and the post-transaction process. I believe the company has a long way to run, but that complexity is expensive and drags down their valuation. They’ll have a lot to prove.

Another one I find interesting, though I’ve never been close nor do I have any real details, is the Advanced Manufacturing Company of America (AMCA). They focus on aerospace and have acquired components companies that sell into Boeing, Airbus, Honeywell, etc. The company started with “a vision to renew the entrepreneurial spirit and engineering legacy of the aerospace and defense industry’s golden age.”

Other models I’ve seen strike me as compelling are existing tech platforms that can extend into becoming owner operators. Folks who have built technology for chemicals, clinical trials, or drug discovery, and then weaving themselves into chemical manufacturers, trial sites, and healthcare biobanks. Though, to date, I haven’t seen the software extension into operating play out successfully thus far.

In pursuit of AI rollups that don’t suck, there’s a fundamental equation to explore. If explored effectively, it can drive a significant amount of value. If explored poorly, it will likely lead to dramatically amplified levels of destruction because of the capital magnifiers.

God speed!