In October 2024, a company called Crescendo completed its most significant acquisition. At the time, Crescendo had only about 20 employees. It acquired PartnerHero, which had 3,000 employees and over 200 enterprise clients. Following the acquisition, Crescendo took over PartnerHero’s operations with its own AI platform, replacing most repetitive customer service tasks with a 20-person AI team (Globe Newswire).
Behind this move is General Catalyst. GC allocated $1.5B for such deals in late 2023, Thrive Capital deployed $1B+, and Lightspeed raised $9B. Total capital in this space has exceeded $3B. What they are buying is clear: the customer relationships and data of traditional service firms. They then use AI to replace 30-70% of repetitive manual labor, pushing EBITDA margins from 15-20% toward 30-40% (General Catalyst).
If you are an entrepreneur, this matters to you. Not because you need to start a rollup, but because your pool of potential buyers, the competitive landscape, and even your company’s valuation logic are shifting. However, the truly valuable insight into AI rollups isn’t in the transaction figures; it lies in a much more universal problem.
The traditional PE rollup strategy involves acquiring many small companies in a fragmented industry at low prices, cutting costs through centralized back-offices and standardized processes, and exiting at a higher valuation after integration. The assumption is that labor can be organized more efficiently.
AI rollups replace this assumption with the idea that labor can be systematically replaced. The process follows four steps: first, incubate an AI platform to prove its effectiveness; second, acquire traditional service firms with established customer relationships; third, deploy AI to replace repetitive tasks; and finally, use the improved cash flow to continue acquisitions, creating a compounding effect.
There are two key differences from traditional PE. The order is reversed: you build the AI first, then buy the companies (Sourcery). The holding period is also longer; both GC and Thrive are benchmarking against long-term operating companies like TransDigm and Constellation Software (LinkedIn).
Sequoia partner Julien Bek proposed a complementary thesis in March 2026: “Services: The New Software” (Sequoia Capital). The core data point: for every $1 a company spends on software, it spends $6 on services. AI now has the capability to capture that $6.
General Catalyst is the pioneer. In late 2023, they allocated $1.5B from an $8B fund for what they call their “Creation Strategy” (QuantFi). As of late 2025, they have publicly executed this in six vertical directions, incubating or investing in at least 10 companies (Capital Founders).
Several representative cases include: Crescendo in customer service outsourcing, with a Series C valuation of $500M; Long Lake in property management, with approximately $670M in funding, 18 acquisitions, and $100M in EBITDA; Eudia in legal services, with a $105M Series A and the acquisition of Irish law firm Johnson Hana, gaining over 300 lawyers; and Titan in IT services, with $74M in funding, automating 38% of tasks.
Thrive Capital’s commitment is even larger. Joshua Kushner created Thrive Holdings, deploying over $1B (New York Times). Most notable is Thrive’s exclusive partnership with OpenAI, where OpenAI researchers are embedded directly into Thrive Holdings’ engineering teams to customize models for portfolio companies (Newcomer). Their portfolio includes Crete, an accounting integration platform with $300M+ in annual revenue and plans to spend another $500M on acquisitions, and Shield Technology Partners, an IT services MSP rollup.
The field is no longer limited to these two. Lightspeed raised $9B, 8VC incubated the stealth-mode IT services acquisition platform Sequence Holdings, and KKR and BlackRock are entering through infrastructure like data centers (Newcomer).
The transaction numbers for AI rollups are certainly eye-catching, but behind these deals lies an insight far more valuable to the average AI practitioner than the transactions themselves. This insight has nothing to do with billions of dollars and everything to do with the problems you encounter in your work every day.
First, look at the numbers. RAND Corporation found that over 80% of AI projects fail—double the failure rate of non-AI IT projects. A recent BCG survey showed that 60% of companies have not generated substantial value from AI. McKinsey reports that while 88% of organizations use AI in at least one function, only 39% have seen any EBIT impact (Talyx).
The cruelty of these numbers is that they are happening in 2025-2026, a time when AI tools are already good enough. With GPT-5 level models, mature RAG solutions, and reliable agent frameworks, the tech stack is no longer the bottleneck. So where is it?
The five root causes of failure identified by RAND are all organizational: unclear problem definition, insufficient data, a tech-first mindset, inadequate infrastructure, and problem difficulty. Not a single one is about the model not being smart enough or the API being too slow. One practitioner wrote on LinkedIn that the reason organizations fail is an “inability to reallocate authority.” Reports talk about “ambition” and “governance” but never mention the internal power struggles that block 80% of transformations (Duperrin).
In conversations with many early adopters over the past year, one observation has repeatedly surfaced: the primary factor determining whether AI can be successfully implemented in an organization is whether the CEO or team leader is a heavy AI user themselves.
The logic is straightforward. If you don’t use AI to write code, perform analysis, or handle emails, you cannot accurately judge what AI can and cannot replace in your business processes. You also lack the firsthand experience to identify which steps AI is already good at and where it still falls short. More importantly, your team will mirror your behavior. If the CEO doesn’t use AI, middle management has no incentive to drive process changes, and frontline employees will see AI as an extra burden rather than an efficiency tool.
Data supports this observation. Bain found that 88% of business transformations fail to achieve their original goals (Mavim). A September 2025 BCG report noted that only 5% of companies have achieved value from AI at scale (Talyx). What do that 5% have in common? Almost all of them have a leader who is deeply involved with AI. Not the kind of involvement where they say “we must embrace AI” at an all-hands meeting and then hand the task to the CTO, but the kind where they use AI for actual work every day and can specifically point out which steps can be automated.
Understanding this context explains why traditional consulting models perform poorly in the face of AI transformation.
Consultants are in an awkward position. You can write a beautiful AI transformation roadmap, build a PoC to prove technical feasibility, and train teams to use new tools. But when the transformation touches actual interests—when a department needs to be downsized, when a middle manager’s authority needs to be redefined, or when long-standing processes need to be scrapped—you don’t have the power to make those decisions. You can only write recommendations and wait for the next contract cycle.
This is the same problem traditional digital transformation faced, but AI has magnified it. Traditional digitalization (like implementing an ERP or building a data platform) primarily changes the tool layer; people’s ways of working change, but their roles remain the same. AI transformation is different; it directly changes the role layer—certain positions are replaced, skill requirements for others change completely, and some decision-making power shifts from people to systems. A transformation of this magnitude is nearly impossible to succeed without the continuous drive and final say of the person in charge.
Back to AI rollups. The fundamental reason GC chooses majority ownership over consulting contracts to drive AI transformation is right here. When you are the controlling shareholder, you can directly replace management teams, restructure processes, and deploy systems without having to convince anyone. Marc Bhargava of GC explicitly cited the fact that “Fortune 100 AI transformation often fails” as a counter-argument for the GC model (Sourcery).
The $1.5B bet is a wager on the value of this “enforcement gap.” AI technology is already good enough; the key to scaling implementation is organizational execution, and the prerequisite for organizational execution is sufficient control.
You likely don’t have $1.5B to buy your clients’ companies. But this insight is still useful because it helps you shift your focus away from the wrong things.
The instinct of many AI practitioners is to think: my tools aren’t good enough, my prompts aren’t precise enough, or my RAG pipeline isn’t fast enough. These are important, but if you are helping an organization implement AI, they are likely not the real bottleneck. The real bottleneck is whether the organization’s leader personally uses AI, understands what it can and cannot do, and has the will and power to drive process restructuring.
If you are the AI lead within a company, your most important task might not be optimizing models, but finding ways to make the CEO a heavy AI user. Let them experience the effects of AI in their daily work firsthand, and let them reach their own conclusions about which steps can be automated. Top-down consensus is far more effective than bottom-up pushing.
If you are an entrepreneur providing AI tools to enterprises, your product design and go-to-market strategy should account for this reality. Selling to a CTO is entirely different from selling to a CEO. A CTO cares about your RAG accuracy and latency; a CEO cares about whether their team can reduce labor costs by 20% without a drop in quality. If you can only talk to the CTO, your product will likely succeed in the PoC stage but fail at scale.
Klarna serves as a cautionary tale for enterprise AI strategy in 2026. After laying off about 700 customer service staff in 2023 to replace them with AI, they completely reversed course by mid-2025. Internal data showed a 27% increase in resolution time and a 35% rise in dissatisfied interactions. CEO Sebastian Siemiatkowski admitted they “focused too much on efficiency and cost, the result was lower quality” (LaSoft). Digital Applied called it “the canonical enterprise cautionary tale for 2026” (Digital Applied). A similar failure occurred with the Commonwealth Bank of Australia’s AI voice bot, leading to a public apology and the rehiring of laid-off staff a month later (SG Solutions).
Thrasio is a warning for the rollup model itself. The Amazon brand aggregator went from a 2020 unicorn to bankruptcy in February 2024. Acquisition valuations spiraled out of control, debt piled up to $855 million, and over 200 brands proved impossible to manage individually (eGrowth Partners). The entire industry invested about $16B, and it is estimated that 90% of these companies are now struggling or dead.
Linas offered a fundamental rebuttal to Sequoia’s thesis on Substack: “For every $1 companies stop spending on humans, they spend $0.03 on AI” (Linas). When machines replace human labor, the pricing of that work recalibrates to the machine’s rate, which is 97% cheaper than human labor. This means the revenue of an AI-transformed service firm could shrink dramatically. Even if profit margins jump, the total market capitalization might not meet expectations.
Results from an anonymous investor survey are also telling. The AI Rollup Investor Sentiment Report 2026 quoted one investor saying, “out of 100 operators I spoke with, maybe 3 are really capable of executing” (AI Rollup Nexus). Fortune questioned the fundamental assumption even earlier: “Services businesses aren’t inefficient by accident. They’re inefficient by design. The inefficiency is the product. Clients pay for flexibility, customization, and someone to blame when things go wrong.” (Fortune)
AI rollups have validated an acquisition-led growth path: build an AI platform first, then acquire distribution channels. Crescendo’s 20-person team gained 200+ clients overnight through acquisition.
But the more practical impact is that your competitive landscape is changing. If you are building AI products for enterprise service scenarios (customer service, legal documents, accounting automation, IT operations), your opponents are no longer just other AI startups. They also include VC-backed platforms armed with acquisition capital, an existing customer base, and the control to mandate AI transformation. Their customer acquisition cost might be an order of magnitude lower than yours.
Conversely, if you have built competitive AI capabilities in a vertical field, you might become an acquisition target for an AI rollup platform—they need your technology to arm the traditional companies they acquire.
Whether you are in BPO, legal, accounting, or customer service, your pool of potential buyers is expanding. The details of Crescendo’s 20 people acquiring 3,000 are worth a close look. They didn’t buy the team; they bought the customer relationships and the data. If your company’s primary assets are customer relationships and industry know-how, and you haven’t started your AI transformation, your position at the negotiating table may weaken.
One possible response is to proactively implement AI. If you can show these buyers that you are already using AI to improve efficiency, your valuation logic shifts from a traditional service firm to an AI-enabled service firm, with the difference in EBITDA margin directly reflected in your valuation.
The space is real and capital is plentiful, but the execution bar is extremely high. Teams that possess AI technical capability, M&A execution experience, and industry know-how simultaneously are incredibly rare. All margin data is self-reported by companies, and none have weathered an economic recession. The actual implementation of AI customer service is far more difficult than vendors claim. Klarna’s reversal and Qualtrics’ research (nearly 1/5 of consumers using AI customer service saw no benefit, CNBC) point to the same conclusion. Furthermore, Linas’s “$0.03 vs $1” data suggests that service budgets might simply evaporate rather than transfer to AI providers, posing a fundamental challenge to valuation models.
This section is for readers sensitive to numbers. Conclusion: the direction is supported by industry data, but the specific multiples should be discounted.
Regarding efficiency: The most specific data provided by GC is a 25-30% productivity increase at Long Lake (General Catalyst). Marc Bhargava mentioned “2 to 3 times more effective” on a podcast (Sourcery). Independent BPO industry reports expect 2-4x (HTC). I found no direct source for a 5x efficiency gain during my research.
Regarding gross margins: GC officially states they “aim to double profit margins, often targeting 30-40% margins”—this is a goal, not an achievement (General Catalyst). Industry benchmarks support the direction: basic call center EBITDA is only 15-20%, while specialized services can reach 30-45% (GoodCall). The average net margin for small businesses is only 7.1%, with labor costs accounting for 47% of revenue (VotedNumberOne).
Crescendo claims AI can “automate up to 90% of support tickets” with “99.8% accuracy in 50+ languages” (Crescendo), with no independent third-party verification. Client Rachio confirmed accuracy in the 95-99% range but admitted it is still in the “early days of production deployment” (PR Newswire).
厂商/官方来源：
媒体报道：
独立分析/行为证据：