Do AI rollups work? The bear case, tested against 36 graded firms

An AI rollup buys established services businesses and runs them on a shared AI and engineering layer, with the aim of expanding margins and holding the companies rather than flipping them. The question owners keep asking us is whether the model works. The honest answer today is that the public record cannot settle it. On our AI Rollup Market Map, 36 firms are graded on the quality of their public evidence: A for an outcome confirmed by a credible third party, B for a client-named or company-reported result, C for a real thesis with no verifiable outcome yet. Two firms hold grade A, and both are conventional private equity benchmarks. All 28 firms in the new-category lanes, owner-operators, rollup platforms, and transformation partners including our own, sit at B or C. That is the frame for what follows. Each of the four standard bear-case arguments gets the same treatment: what the map shows, what the skeptics say, and what evidence would change the verdict.

Argument 1: Services businesses do not re-rate to software multiples

What the map shows. The highest-valued firm on the map is Thrive Holdings, which raised $2 billion at a $12 billion valuation in August 2026 on a portfolio of accounting, IT services, and regulatory businesses. Its most detailed AI outcome, roughly 7,000 tax returns drafted by Current's Tax AI with about a third of preparer time saved, comes from a case study co-published by OpenAI, which is also a shareholder. Long Lake's $6.3 billion Amex GBT take-private is financed in part by $1 billion of 7.625 percent senior secured notes. Neither firm has published a margin result that a customer, auditor, or counterparty has confirmed. The valuations price a re-rating that the evidence has not yet earned.

The bear case. Nathan Benaich of Air Street Capital and Nikola Mrkšić of PolyAI made the argument in a June 2025 Fortune op-ed, republished in full on Air Street's site: the thesis confuses operational improvement with business-model transformation. Their comparables are public. Concentrix, Genpact, and Infosys, all of which sell AI-transformed services, trade at 5x to 23x EV/EBITDA, while Salesforce, ServiceNow, and Workday trade at 22x to 92x. Concentrix has generative AI deployed at more than 1,000 customers and still runs at roughly a 10 percent EBITDA margin. On a revenue basis, The Deep Feed's summary of the services-as-software thesis puts services at 1x to 2x revenue against 8x to 20x for software. PolyAI tested the rollup route itself: a six-month study in 2019 of buying contact centers ended in a decision to walk away, on three grounds the authors call the illusion of control, the pricing trap, and zero switching costs. Their verdict is that AI rollups are "at best, tech-enabled private equity: operationally heavy, valuation-capped, unlikely to scale like software."

What would change the verdict. A single acquired business with audited margins before and after the AI layer, held long enough to show the gain survived a pricing cycle. No firm on the map has published one. Until then the multiple gap is a fact about public comparables, and the rollups are asking to be priced as the exception.

Argument 2: Venture math does not close, so the category moved to permanent capital

What the map shows. The map splits by capital structure. Thrive Holdings was formed as a separate company rather than a fund and states in its own essay that it holds the businesses it buys forever. Long Lake plans to hold permanently, on the Berkshire model. General Catalyst, which coined the term "AI-enabled roll-up," describes its version as deliberately adding engineering cost, committing $100 million or more per platform, and holding for the long term. Venture-backed platforms such as Crescendo, Eudia, and Dwelly sit on fund clocks. Both groups are graded B at best, because the structure changes the timeline without changing the evidence.

The bear case. PitchBook's Leah Hodgson quoted Foundamental's Fabio Bronzin in August 2025: "the maths doesn't make sense." His worked example: a VC pays $20 million for 20 percent at a $100 million valuation, and if all of that cash goes into acquisitions, the VC has supplied 100 percent of the capital for 20 percent of the company, an effective 5x book. A fivefold increase in the business returns 1x. Venture debt on top eats the EBITDA that was supposed to be the point. CNBC's June 2026 report on the category names the same risk: operating companies return 100 to 200 percent over a long hold, two to three times the money, while venture investors underwrite 10x. RollupEurope's January 2026 survey of holdco capital describes long-term-hold vehicles as tying founder outcomes to MOIC "as opposed to IRR, which arguably disincentivises long-term compounding," and quotes Tenet's Sahil Patwa arguing that neither VC nor search-fund and PE structures fit, so a new type of inception-stage capital is needed. Permanent capital is the category's answer to Bronzin, and it is a real one. It also removes the exit that would have forced a verified number into the open.

A second piece of arithmetic comes from the bull side. General Catalyst's Hemant Taneja argues that AI has broken the software buyout itself. In Sourcery's written account of GC's first quarterly review, the example is a business bought at 25x EBITDA whose EBITDA grows 50 percent and whose exit multiple falls to 12x, returning 0.68x. A transcript summary of the same conversation renders his spoken example as a 15x entry with two-thirds debt, EBITDA doubling, and a terminal multiple collapsing toward 3x. The numbers differ between the two accounts; the conclusion does not. In Taneja's words, "the assumption of terminal value is gone." If the investor funding a dozen rollups believes terminal multiples are unknowable, an owner should ask what multiple the rollup's own model assumes at the end.

What would change the verdict. Distributions. The Capital Founders family-office guide notes that most platforms are under five years old and that claimed returns often include unrealised gains marked at valuations the sponsor determines. Cash returned to investors, or a dividend paid out of a transformed business's free cash flow, would settle the math argument in a way that a Series C mark cannot.

Argument 3: Integration is the whole job, and the base rates are poor

What the map shows. Pace is the reservation that recurs across the map. Beacon Software has acquired more than 30 companies since 2024 and now closes roughly one deal a week, and its one published outcome is a single company-reported aggregate: 50 percent growth in portfolio operating earnings with no per-company breakdown. Long Lake bought more than 30 businesses in about three years before agreeing to take a $6.3 billion public company private. Dwelly has acquired 17 UK letting agencies on $260 million raised including debt facilities. Fura's best result, an acquired broker moving from a $150,000 loss to $1 million in profit, came with headcount falling from 26 to 8, which is integration work as much as AI work. None of this is disqualifying. It is the exact profile the base rates warn about.

The bear case. Paul Carroll and Chunka Mui's Harvard Business Review study of 750 major business failures found that "more than two-thirds of roll-ups have failed to create any value for investors," and that rollups often "cannot sustain their fast rate of acquisition" once cultural and operational issues surface. CapitalPad's June 2026 review of buy-and-build data concludes that "integration, not sourcing, is the binding constraint": in the BCG and HHL sample it cites, deals with one or two add-ons earned a 35.5 percent IRR while deals with more than two add-ons earned 19.9 percent, below standalone buyouts, and the page carries a practitioner estimate that about 60 percent of rollups miss projected synergies within two years. The AI era has its own cautionary cases. Thrasio raised more than $3 billion in equity and debt to roll up Amazon brands and filed for Chapter 11 in February 2024; Equal Ventures' Rick Zullo cites it as the template for platforms that "burn through capital and then sell for less than their paid in capital raised," and PitchBook uses the Amazon aggregators as the precedent for venture money in rollups. Renovo Home Partners, a home-improvement platform that Audax Private Equity formed in 2021 and that reported $653 million of revenue in 2023, ceased operations on October 29, 2025 and filed Chapter 7 five days later. Capital & Clarity's April 2026 analysis of General Catalyst's portfolio cites Renovo as the case for integration risk compounding with speed, and adds that GC's portfolio margin figures are reported by the firms themselves, with the largest portfolio company under three years old. Luke Sophinos, who has tracked the category for two years and remains bullish on it, wrote in July 2026 that "every single failure mode I've seen in this category traces back to acquisition pace outrunning integration capacity."

The investors say the same thing about themselves. In the AI Roll-up Nexus 2026 investor survey of 102 respondents, about 80 percent of them venture and growth investors and about half European, 79 percent named integration and change management the biggest risk and 68 percent named overhyped AI value creation. One disclosure is required: the survey is published by Tenet, an inception-stage fund that backs AI rollups and appears on our map at grade C, so this is the category's own investors describing their own worry.

What would change the verdict. A firm that has integrated a dozen acquisitions publishing staff and client retention at 18 months, per acquired company. Sophinos's line cuts both ways. If pace is the failure mode, the firms that slow down and publish per-company results will be the first to earn grade A.

Argument 4: The moat is rented, unless the gains show up on the revenue line

What the map shows. Every headline AI number on the map is a cost or capacity number. Long Lake reports 25 to 30 percent productivity gains in HOA management in its lead investor's essay, and 20 to 40 percent when its CEO describes it. Beacon reports the operating-earnings aggregate above. Current reports preparer time saved on 7,000 returns. Multiplier's flagship, Citrine International Tax, reports cash flow up roughly 2.5x within eight months, company-reported, at a firm that was 12 people at acquisition. The one revenue-side claim, Long Lake's 10x increase in new-customer pipeline, also comes from General Catalyst's essay. All of it is real work. None of it has been checked by anyone outside the cap table.

The bear case. Zullo's argument in Equal Ventures' November 2025 essay is that cost advantages from AI "are largely ephemeral," because any business planning for long-term success "needs to assume that competition will implement similar AI interventions." Revenue-side synergies, in his view, "are far more sustainable": aggregation that lifts revenue, such as higher commission rates on larger premium volumes in insurance, produces a durable margin, while a cheaper cost-to-serve gets competed into price. Footnote, an accounting-industry newsletter, applied the same logic to Thrive's Current in July 2026. The AI capability Thrive built with OpenAI is impressive, "but a close cousin of it is now sold to every independent firm in the country by Ramp, Basis, Digits, and a dozen others," so an efficiency edge available for a monthly subscription amounts to "a rented advantage that your un-consolidated competitor can rent too." Footnote adds that most accounting deals were underwritten on billable-hour cash flows that the same AI compresses. A RollupEurope panel with OpenOcean and Unbound asked the question in its shortest form: "if the groundbreaking tech is widely available, what's the moat?"

What would change the verdict. Revenue. A rollup that can show an acquired firm winning clients it could not have won before, at prices it could not have charged before, has something a subscription cannot rent. Long Lake's pipeline claim is the right kind of number; it needs a customer or an auditor behind it. Until a firm publishes that, the operating layer is a cost advantage with a shelf life, and the shelf life is set by the vendors selling the same tools to the firm next door.

What would move a firm to grade A

The bar on the map is deliberately narrow: a specific, named outcome verified by a credible third party, meaning a customer, an auditor, or a counterparty rather than the firm or its investors. The two firms that meet it show what that looks like. Vista Equity Partners is the named exemplar in Bain's Global Private Equity Report, which documents Avalara's sales reps responding 65 percent faster with generative AI and LogicMonitor's Edwin AI saving customers about $2 million a year on average. Apollo's portfolio AI program is the subject of an MIT Sloan Management Review case study recording a 40 percent cost reduction in select content production at Cengage and a procurement saving of more than 65 percent from cross-portfolio contract analysis. Neither is an AI rollup. Both are conventional PE firms that let a third party inside. Grades move in both directions: on August 26 we regraded Metropolis from A to B because its deployment metrics turned out to be company-blog-reported, and moved Circeus, BHub, CVC, and Serent from B to C because their public records show corporate activity rather than AI outcomes. Any owner-operator on the map can move up the day it publishes one acquired company's before-and-after with a name on it other than its own.

Five questions to ask any of them

Whether the firm across the table is a rollup, a venture platform, an outside operating layer, or our own, these are the questions the evidence on the map says an owner should ask before signing.

1. Show me audited, or at least client-confirmed, before-and-after numbers from two businesses you already own: EBITDA margin at close versus today, and how long the improvement took.
2. What does your headline AI number measure, exactly, and who outside your company checked it?
3. What happens to my team, my brand, and my client relationships in year one, and which of your prior acquisitions can I call to confirm it?
4. If I roll equity, what is the second bite worth today, at whose valuation mark, and on what timeline does it become cash?
5. If we part ways, who keeps the software, the data, and the people you trained?

Footnote's warning about earn-outs belongs with question four: an offer is often a multiple on cash flows that AI may compress, wrapped in an earn-out tied to margin targets that depend on an AI rollout you will not fully control.

The third path

If you own an established, profitable services business, the bear case above is a list of things you do not have to underwrite. You already have the distribution, the clients, and the people that every acquirer on the map is paying a premium to buy. The cheapest way to find out whether AI changes your margins is to run one outcome inside your own company, with success criteria written down before the work starts and a judge you choose, and to keep the software, the data, and the trained operator whether it works or not. That is how we run every engagement, and we grade ourselves B on our own map because our results are client-named but vendor-published. The ledger the hub renders live shows every initiative resolved as a success, the success rate against criteria agreed in advance, and the ones that missed recorded alongside them. The owner's section of the market map lays out the four paths side by side. If you want to underwrite one outcome before you decide about any of them, book a fit call.