Part 11 of The Founder's Guide to AI-Enabled Roll-Ups
If a platform wants to buy your firm, or a sponsor wants your money, the useful question is how this particular one could fail. So far, the cases fall into three patterns. Each leaves signs you can check before you sign anything.
I'm bullish on AI roll-ups. But staying bullish without being naive means looking hard at how things could break.
Most failures in this space will be ordinary. The idea was fine, but the operator wasn't, or the market wasn't ready, or the money was too cheap to make anyone careful. A wave this big pulls in people who shouldn't be in it. Knowing how they fail is how you avoid selling to one of them, or backing one.
What's Inside
- Three ways these break. The valuation gap never closes, the buyer buys faster than it can integrate, or the AI doesn't do the work the plan assumed.
- The software multiple is still a hope. The cleanest accounting exit so far was reportedly priced at about 15x EBITDA, up from about 11x. Software companies trade at 22-92x.
- Integration is the oldest failure in roll-ups. A private equity home-improvement roll-up went into Chapter 7 in 2025, and the investors backing this wave name integration as their biggest risk.
- AI claims are being tested. Regulators have acted against companies whose "automation" was mostly people, and one careful owner called its own AI experiments "shockingly poor".
- No AI roll-up has exited yet. The big deals of 2026 are investors buying in. Nobody has sold one of these platforms at any price.
- Six signs of a platform built to last. AI that works on messy data, a deal pace set by integration, patient capital, a tested integration process, founders who stay, and returns that work without a software multiple.
Three ways AI roll-ups fail
Every failure in the cases below comes back to one of three things:
- The valuation gap between services firms and software companies never closes, so the exit the plan depends on doesn't happen.
- The platform buys firms faster than it can integrate them. Coordination costs then eat the gains.
- The AI turns out not to do the work the plan assumed.
And a platform can have all three at once.
Why the valuation gap between services and software may never close
The boldest claim in the pitch is that a services firm improved by AI should trade like software.
Writing in Fortune in June 2025, Nathan Benaich of Air Street Capital and Nikola Mrkšić, PolyAI's chief executive, compared two groups of companies. Outsourcing firms that have invested heavily in automation, such as Concentrix, Genpact and Infosys, traded at 5-23x EV/EBITDA. Software companies such as Salesforce, ServiceNow and Workday traded at 22-92x.
They called the roll-up pitch a "category error": "confusing operational improvement with business model transformation".
The cleanest exit so far supports the sceptics. When a Blackstone-led group bought New Mountain's stake in accounting group Citrin Cooperman in January 2025, the firm reportedly traded at about 15x EBITDA, up from roughly 11x when New Mountain invested in 2021. Neither side disclosed terms. That's real expansion and real money, and it's still a services multiple.
The biggest accounting deal of 2026 points the same way. In July, Grant Thornton's New Mountain-backed platform agreed to buy the listed accounting group CBIZ for about $5bn in cash, which William Blair's Andrew Nicholas put at roughly 10x EV/EBITDA.
General Catalyst, the model's most committed backer, makes the point itself. In its first-quarter letter, its chief executive, Hemant Taneja, set out a hypothetical buyout: pay 25x EBITDA, sell at 12x, and you get back about 0.68x your money even after profits grow. Pay 10x and sell at 14x, and the same work returns about 2.7x. The entry price does most of the deciding.
The UK showed the other side this year. Xeinadin, an accountancy group partially owned by Exponent, paused its sale in February 2026 when bids fell short of the more than £1bn it sought.
In May, Sumer's owner, Penta, paused its own £1bn sale and was reportedly weighing a continuation vehicle instead. In October, reports said TPG would lead that vehicle at a valuation above £1bn, which none of the firms has confirmed.
Both are traditional consolidators. Both show how hard a clean exit gets when it depends on buyers paying up.
A disciplined underwriter models a good services or private equity outcome and treats any software re-rating as a bonus. Chapter 6 makes the same point from the investor's seat.
Everything here is free. Subscribing just tells me the content is useful — and helps me decide what to write next.
Buying faster than you can integrate
The second failure is operational, and it's the oldest story in roll-ups. You can buy faster than you can integrate, and the gap between the two is where you lose money.
Renovo Home Partners is a clear recent case, and it wasn't an AI deal at all. Audax Private Equity built it from regional remodelling businesses, including Dreamstyle, Alure and Remodel USA.
Renovo shut its doors at the end of October 2025 and filed for Chapter 7 in November, with liabilities listed at between $100m and $500m. Its main lender, BlackRock, marked its loan to zero from 100 cents on the dollar.
Pro Remodeler's analysis of the collapse blamed the debt service, a slowdown in demand and the bureaucracy that came with centralising local businesses. Buying more firms is meant to create scale. Without a working integration machine, it mostly creates coordination, the same drag that slows large services firms in the first place.
Debt makes those mistakes expensive. The Private Equity Stakeholder Project, an advocacy group, counts private equity as having a role in 19 of the 35 US bankruptcies of 2025 with $1bn or more in liabilities. This year's list already includes Multi-Color, a label maker that made at least eight acquisitions after its 2021 buyout and filed in January with $5.9bn of debt.
These figures are for private equity in general. They show how little room a debt-funded roll-up has before anyone adds an unproven AI assumption on top.
The investors backing this wave know it. In Tenet's 2026 survey of 102 investors, most of them venture capital firms, 79% named integration and change management as a risk, ahead of overhyped AI value (68%) and poor M&A discipline (62%). Only 3% named regulatory scrutiny. Tenet backs these companies, so treat the survey as a record of what investors expect.
Integration problems can also surface as legal ones. Metropolis, the computer-vision parking company that grew by buying SP+, paid $8.75m to settle a Tennessee investigation into deceptive parking practices in January 2026. Its chief executive has linked the problems to legacy operations and past acquisitions.
This is one reason the patient-capital, build-first approach of the better platforms matters. Patience and a proven integration playbook prevent this failure.
When the AI doesn't do the work
The plan only works if the automation does the work. When it doesn't, the margins never arrive, and sometimes the claims were never real.
Regulators have acted. In January 2025, the SEC found that Presto Automation misled investors about its drive-through "AI": most orders needed human help, largely from order takers in the Philippines and India. Presto settled without admitting or denying the findings.
In April 2025, US prosecutors charged Nate's founder, a shopping app, with raising more than $40m on automation that, they allege, was done largely by contractors in the Philippines.
Even where the technology is real, deployment is hard. Klarna said in 2024 that its AI assistant was doing the work of 700 full-time agents. By May 2025, it was hiring people again for customer service after its chief executive said the cheaper AI service had produced lower-quality results.
Permanent Equity, which has 11 portfolio companies held through 30-year funds, hired a technologist with a PhD in AI and ran dozens of agentic-AI experiments across the firm and its portfolio in 2025. Its annual letter called the results "shockingly poor compared to expectations", and it is pulling back.
Thrive Holdings deserves credit for publishing its own benchmarks in October. On its tax test, the best set-up scored about 76%. On its IT test of 52 simulated support tickets, the best set-up solved 32. That's real progress, and a long way from work nobody has to check.
The wider surveys point the same way:
- MIT's NANDA project reported in July 2025 that 95% of organisations were getting no return from generative AI. The report calls its findings preliminary and based on interviews, so treat the figure as a warning.
- In a May 2026 survey of 2,527 senior decision makers published by Sinch, which sells customer-messaging software, 74% of enterprises said they had rolled back or shut down a live AI customer communications agent.
- In a survey of managed IT providers published in October 2026 by software vendor Nerdio, 32% had adopted AI in their own operations, and only 12% could point to measurable productivity gains.
Some of the automation is real. The gap between a working platform and a hopeful one is still wide, and you can test it. The live demo on your own data in Chapter 4 exists because so many claims don't survive one.
Will there be a shake-out in AI roll-ups?
No AI-led platform has failed yet, as far as I can find. My view is that some will. Some platforms overpaid in the land grab and will struggle to change what they bought. Others sold automation they couldn't deliver, or simply bought too fast.
When that happens, the coverage will turn sour for a while. That's how waves like this usually go: cheap money and hype pull in people who shouldn't be there, a correction clears them out, and the disciplined ones keep compounding through it.
If you're deciding who to trust with your firm or your capital, a shake-out is useful information. It shows which platforms were built properly.
Signs of an AI roll-up built to last
The failures point straight at what to look for, whether you're choosing a buyer or backing a platform:
- The AI is real, and shown working on messy, real-world data.
- Integration capacity sets the pace of acquisitions, and deals wait when the team is busy.
- The capital is patient, so nobody is forced to sell half-finished.
- Many firms have refined an integration process.
- Founders stay on and stay aligned because their operating knowledge is part of what was bought.
- Returns are underwritten to a services or private equity outcome, with any software re-rating treated as a bonus.
Take the list to your next conversation with a platform and ask for evidence on each line. It takes one meeting. Treat any line they can't evidence as a warning sign.
Why the exit question is still open
The largest deals of 2026 were all purchases. Long Lake, backed by General Catalyst, completed its $6.3bn take-private of American Express Global Business Travel in September. Thrive Holdings raised more than $2bn at a $12bn valuation in August. Both are investors paying up to buy in.
The nearest thing to an exit is a listing. Teamshares, which buys small firms from retiring owners and calls itself a tech-enabled acquirer, went public on Nasdaq through a SPAC in June 2026.
Its first results as a listed company showed organic revenue growth of 3.4% in the quarter, and its filing warned of substantial doubt about its ability to continue as a going concern. In October, it repaid a loan due in December with a new $130m four-year term loan.
As far as I can tell, no AI-led roll-up has been sold or recapitalised, so no one has sold one at a software multiple either. Even among ordinary accounting roll-ups, the private equity firms selling to each other this year, such as Broad Sky Partners when it sold Smith + Howard to TPG, disclosed no terms.
That worries me less than it might, because the services outcome stands on its own. It does mean the biggest prize the bulls talk about is still a hypothesis, and a platform that needs it to be true is taking a risk a better-run one isn't.
So the failures don't argue against the model. They argue for doing it properly and building each lesson into a design choice from the start.
Common questions
Why do AI roll-ups fail?
Mostly for three reasons: the valuation gap between services firms and software never closes, the platform buys firms faster than it can integrate them, or the AI doesn't do the work the plan assumed. Many failures are ordinary execution problems.
Has any AI-enabled roll-up been sold at a software multiple?
Not yet. As far as public sources show, no AI-led roll-up has been sold at all, and the largest deals of 2026, such as Long Lake's take-private of American Express Global Business Travel, were purchases. The cleanest accounting exit, Citrin Cooperman in January 2025, was reportedly priced at about 15x EBITDA, a services multiple.
What is AI-washing?
Claiming AI capabilities a company doesn't have. US regulators have acted against companies whose "automated" products relied heavily on people, and in AI roll-ups the risk is margins promised on automation that doesn't exist yet.
How can you tell an AI roll-up built to last?
Look for AI shown working on messy real data, an acquisition pace set by integration capacity, patient capital, an integration process used across many firms, founders who stay, and returns that work without a software re-rating.
Previous: When to Walk Away
Next: UK and Europe
Playbook Hub: The Founder's Guide to AI-Enabled Roll-Ups
Capital Founders OS is an educational platform for founders with $5M–$100M in assets. Frameworks for thinking about wealth — so you can make better decisions.
Explore more: Playbooks · Capital Signals · Wealth Architecture · Investment Strategy · Business Building · Life Design
Found this useful? Forward it to a founder who's thinking about this stuff. Got a question or disagree with something? Get in touch.
New here? Subscribe for one email a week.
