The $16 Trillion Services Market Is Being Bought, Not Sold To

Global services generate about $16 trillion in annual revenue. Global software generates about $1 trillion. Those two figures, set side by side, explain why a Silicon Valley venture firm has started buying accounting practices, law firms, IT outsourcers, and call centers. Marc Bhargava, a managing director at General Catalyst, laid out the playbook on *The Cognitive Revolution* with host Nathan Labenz.

His background fits the subject. He co-founded Tagomi, the first prime broker in crypto; Coinbase acquired it and it became Coinbase Prime, which his co-founder Greg still runs. Bhargava is an operator by temperament, now in his third year at General Catalyst. He works on creation, the firm's incubation track, which previously helped build companies such as Kayak and Livongo.

He calls what he is building an AI-enabled rollup, or a compounder. The logic is straightforward: incubate a vertical applied-AI company, prove it with customers, then buy traditional service firms, run the same work through AI, free up cash flow, and buy the next one. Public markets already know this loop. Danaher, TransDigm, Constellation Software, and Berkshire Hathaway buy companies, improve them, generate cash, and buy again — hedge-fund favorites for the past decade. Bhargava's claim is that AI is a new lever inside that same compounding machine, aimed at industries that barely buy software and label it a sliver of IT spend.

Venture capital traditionally ignored services for a blunt reason. EBITDA margins sit at 5 to 10%, and after capital expenditure there is little left to support venture-scale returns. Bhargava's team argues that if 30 to 70% of the tasks in a category can already be handed to AI, those margins can move toward 30 to 40% — closer to software. Once that arithmetic holds, services stop being uninvestable and become something you eat through acquisition.

They did the unglamorous work first. The team broke 70 service categories down to the task level, asked which steps are automatable with today's models, and kept about ten candidate verticals. That cut happened before the first check. After a first round of funding, the incubated team has to go one layer deeper: shadow customers in the field, split each person's day into tasks, and come back with an hour-level map of where time actually goes. Only then does General Catalyst put up capital to buy a company, often one the team is already serving.

"We map out the hours at a company," Bhargava says, adding up people and workflows until the calendar is visible. The firm has now placed eight of these AI-enabled rollup bets, including names such as Crescendo in call centers and Long Lake in HOA management. In five of them, General Catalyst incubated the company and wrote the first check. Founding teams are deliberately hybrid: applied-AI operators from places such as Rippling, Ramp, and Scale AI, paired with industry veterans who have done rollups and private-equity-style M&A before.

Over the past two years, the work that has actually automated has clustered in four buckets. The first is customer support and frequently asked questions; at Crescendo, Bhargava says 50 to 70% of call-center workflows can now be handled by agents and large language models. The second is content production — decks, copy, marketing, email — any business that moves a lot of text. The third is repetitive data work: filling forms, checking boxes, reading spreadsheets, the insurance claim that asks whether the roof caved in and whether the paperwork is complete. A related portfolio company in bookkeeping, Kick, is already seeing about 80% automation on that kind of grind.

The fourth bucket matured only in the last year: logic and reasoning. In insurance that means moving from checkbox review to a suggested price for a policy; in legal, a lawyer can ask what hard questions a case will draw. Everything outside those buckets they mostly walk away from. Heavy physical work — plumbing, roofing, HVAC, firefighting — can get help with routing and payments, but not enough to clear the 30% line. High-churn industries such as digital ad agencies and PR shops are also out: clients already use several vendors, so one wobble during the transformation can lose the account, and there is no six-to-twelve-month buffer to land the change.

A third screen is model capability itself. Two years ago insurance underwriting was not viable; today it is. That line moves every three to six months, so the list of ten verticals is not frozen. Legal is the positioning puzzle Bhargava dwells on. Most legal spend sits with outside counsel, not software, and selling efficiency tools to hourly-billing firms cuts the firm's own invoice. Eudia sidesteps that by selling to Fortune 100 general counsel: a mix of software and services so in-house teams can take more work with the same headcount and send less out.

To deliver that mix, Eudia acquired Johnson Hana, a legal services company, and folded the people into its offering. Bhargava's analogy is Palantir, which staffs forward-deployed engineers and looks like a services shop while selling an AI platform and an outcome — "an end result or work." He calls the category selling work. On the models, he says OpenAI is stronger at summarizing case text and Anthropic is stronger at reasoning and hostile questions, so Eudia routes both through, task by task. Gains every three to six months flow to the application layer if the team knows how to evaluate and assemble them.

The hardest piece of the model is the acquisition itself. Most deals do not create the value promised at signing, a consensus Bhargava does not dispute. The workaround is to start as a vendor. Crescendo served ten to fifteen call centers, showed high automation rates, then went back to buy a larger one. By then the conversation has changed: the owner has seen the technology, knows there is no internal engineering team or CTO, and cannot land the transformation alone.

Deal structure does the rest. Sellers can roll roughly 30% of their equity into the new entity, and if they believe there is a 10x from here, that stub can be worth three times the entire cash consideration. Bhargava says many owners ask to raise the rollover, which is unusual in classic private-equity deals. Early transactions also use little or no debt, so there are no interest payments forcing a cost-cut. The extra line items on the books are software subscriptions and engineers, a bet that revenue growth will cover them.

The quantitative bar is a 20% cost increase and a doubling of revenue inside three years. If that ratio does not underwrite, they do not buy. Staff reactions are the most earthbound part of the conversation. Bhargava says there is always apprehension at first, but the pitch holds because there is no interest burden and the first move is to add cost, not cut people. The work that remains is usually the harder, more human part of the day, which makes "we are not replacing the workforce" a claim the numbers can support.

On implementation, one detail is worth copying: do not change the workflow. In HOA management, the property manager still emails a request for a board deck; the agent drafts it and emails it back. Nobody logs into a new system. Nathan offered a contrast from Ambience Healthcare and Cleveland Clinic, where every doctor had to try the AI scribe once; from that single use, 75% of 4,000 doctors across 60 specialties became regular users. After the first acquisition, leadership from the bought firm talks to the next target's staff, and the persuasion cost drops.

Labenz says he is bullish on the strategy, then raises three objections that are more useful than the answers. First, the transformable set may be small. He estimates that only about 20% of service companies have the leadership, operations, and culture to be worth buying and converting. The other 80%, he suspects, lose customers and good people and slowly shrink; the compounder grows by taking those clients rather than by acquiring the laggards. Bhargava disagrees: the tools are built to sit inside existing email and Slack workflows, and a retiring baby-boomer generation is being replaced by people who already treat ChatGPT as an operating system.

Second, the employment math may not close. Bhargava's frame is the same people doing two to three times the work, which assumes demand is not capped. Labenz uses accounting as a counterexample: most business owners want the cheapest path to compliance, not a hidden wish list of extra accounting. Bhargava's reply is a severe accountant shortage — top firms could take 50 to 100% more clients if they could hire — plus a product shift from three statements and a tax filing toward proactive recommendations, models, and capital advice. The reply is coherent, but it still assumes demand follows supply.

Third, if more than 90% of tasks automate in one or two years, where do the displaced people go. Labenz bets on human-heavy fields with elastic demand: nursing, childcare, eldercare, and education. Bhargava points the same direction. General Catalyst incubated Hippocratic, an AI-native nurse that works with health systems for proprietary data; he says the country could use ten times as many nurses, because six- and nine-month post-surgery follow-ups simply do not happen now. Labenz still expects the abundance story to arrive with a much bumpier transition than Bhargava projects.

Labenz also asked whether big tech will keep the best models for its own products. Bhargava's answer is about supply: seven or eight serious model providers, plus Chinese labs such as DeepSeek and Alibaba, make a coordinated close-the-API move extremely unlikely — he puts the odds below 3%. He notes that the MAG 7 spent about $150 billion on AI research and capex last fiscal year and will spend closer to $250 billion this year. He also admits that last summer the models looked as if they were flattening, until reasoning models at the end of the year pulled the curve back up.

Map the tasks before you buy the tools. General Catalyst will not fund an acquisition until the team has shadowed customers and split each person's day into work that can and cannot be automated. Anyone pushing AI inside a company can steal that sequence and skip the license shopping. The usual discovery is that the bottleneck is not the model; it is that nobody can say what employees actually do all day.

Do not change the workflow; change who sits behind it. The HOA email is the pattern to copy. If a new tool asks people to change daily motions, adoption collapses. Hide the model behind the existing inbox so the switching cost is close to zero. Bhargava is explicit that they are not trying to reimagine the job on day one; they overlay AI on the same seven or eight steps, automating some and leaving others human, especially in regulated, emotional work.

Use a 20% cost increase and a doubling of revenue in three years as the investment hurdle. That frame beats "how much labor did we save," because it forces a story about where incremental revenue comes from. If an AI project only cuts cost and has no path to more output, it probably cannot support the valuation or the team's confidence. Bhargava says they talk about the strategy in public for two reasons: to attract people leaving applied-AI teams who want to found companies, and to tell owners who were about to sell to private equity that there is a third path. The second is the real moat for now — deals come from trust and a live demo, not from being the highest bidder.