The Five Trillion Dollar Bet: Buying Small Service Businesses With AI Agents

The headline number is not a software story. It is a retirement story. According to McKinsey's data on the ownership transition, roughly six million small and midsize businesses in the United States will change hands over the next decade, representing about five trillion dollars in value. McKinsey also estimates that without a transition plan, six million businesses could shut down and twelve million jobs could be lost. The speaker reads this wave not as an empty threat to employment but as a chance to change how service work is actually done.

Three conditions arriving at once

Three things have to land together for a new business model to appear. The first is ownership turnover: accounting, insurance and property management firms founded in the eighties and nineties mostly have no succession plan, the children do not want the business, and the employees do not want to run it. The second is that AI has become genuinely good at the work these firms do all day: pulling numbers out of a PDF into a spreadsheet, chasing a client a third time for a missing document, writing a similar status update fifty times a week. The third is pricing. A service firm runs at five to ten percent profit, and buyers have valued them that way for decades. The speaker's point is that the third condition is what just stopped being true.

What makes the outcome work is that last condition. Revenue stays flat, because the clients are the same and the invoices are the same. Costs fall, because agents do a large share of the work. The table the speaker draws is margin moving from a five-to-ten percent band to a thirty-to-forty percent band, the same business earning three or four times more profit. The attractive part is that nothing has to change on the revenue side. The risk sits in the same place: margin expansion only counts if the cost of delivery genuinely comes down.

What the big funds are doing

General Catalyst was an early mover. The Wall Street Journal reported in late 2024 that the firm had set aside one and a half billion dollars for a strategy of buying service businesses and transforming them with AI, and the framing in that piece stresses that this is not a private equity roll-up. General Catalyst's own writing on Europe draws the same line, saying the firm deliberately adds cost for technology build and platform work rather than extracting it. The portfolio companies span accounting, customer support, legal work, IT management, property management and, in the UK, letting agencies. What emerges is a strategy spread across a fragmented, labor-heavy services chain rather than a bet on one sector. Long Lake, a property management platform, says it bought eighteen businesses and reached a hundred million dollars in EBITDA in under two years, which Capital & Clarity calls unusual by any standard. Crescendo says AI resolves about ninety percent of frontline tickets, charges per resolved ticket rather than per hour, and runs margins four times higher than a traditional contact centre. Titan, incubated by General Catalyst, announced a seventy-four million dollar round in 2025 and its first deal was an IT managed service provider serving financial services clients. Dwelly, consolidating letting agencies in the UK, closed a sixty-nine million pound round. Every one of those figures is self-reported. Capital & Clarity adds the warning that matters most: integration risk compounds with acquisition speed, the largest portfolio company is under three years old, and none of them have been through a recession. Menlo Ventures' survey of enterprise AI spending shows the ground is genuinely prepared, with spend rising from 1.7 billion dollars in 2023 to 37 billion in 2025, six percent of the software market and 3.2 times year-on-year growth.

Thrive moves faster and is built differently. Thrive Holdings, spun out of Josh Kushner's Thrive Capital, put up a one billion dollar vehicle of its own in 2025 and announced an OpenAI partnership in December of that year. OpenAI's announcement says it will embed research, product and engineering teams inside the portfolio companies, with accounting and IT services first. Thrive Holdings' own statement says the raise is over two billion dollars at a twelve billion dollar valuation, with SoftBank, D1 Capital and Altimeter participating, that more than seventy businesses sit on its platforms, and that part of the new money goes into a new platform for the regulatory work of physical infrastructure.

A result measured in tax season

The most concrete result so far came on the accounting side. Tax AI, built jointly by OpenAI and Thrive Holdings, processed seven thousand returns across the participating firms this tax season, drafted them at up to ninety-seven percent accuracy, saved practitioners about a third of their preparation time and raised throughput by close to half. OpenAI's engineering write-up also measures the system improving itself: at launch only a quarter of returns had seventy-five percent of fields correctly completed, and within six weeks that share reached eighty-six percent. The system started on simple W-2 and 1099 forms and, as the season progressed, moved into K-1 schedules and rental income cases that are far harder to extract.

Data entry alone on a medium-complexity return could take eight hours, which is the single most useful number in the story about where the efficiency comes from. The speaker describes the people who used to prepare returns moving into reviewing them, and the nature of the work left for a human changing rather than disappearing. There is a two-sided picture here. Thomson Reuters found in 2026 that seventy-four percent of professionals use AI tools every week while ninety-one percent believe their organisation delivers less than the technology could. The work is changing; the margin claims still have not been independently audited.

The one-person version

Here is the central claim: the model is not only open to billion-dollar funds. A solo founder works in the segment those funds ignore, because large funds are built around mid-market targets and seven-figure revenue expectations. In the one-person version the founder captures three advantages at once: access to the same models and tools the funds use, the ability to do the integration themselves, and the fact that an owner who spent forty years building something cares who takes it over. Owners who would never sell to a large venture firm may sell to a person instead, and the competition does not come from the fund at all, because it lives on a different layer. The cost is that one person carries operational responsibility across several businesses and has to keep new ones from breaking the system.

The founder then turns the idea into an actual structure. At the top sits a holding company that owns the businesses; below it, a handful of firms each with its own general manager; at the bottom, a shared layer every business uses. The point the speaker emphasises is that this shared layer already exists when the second business arrives and is further settled by the third: agents, rules, back office and dashboards are built once. The general manager is usually the senior bookkeeper who has been there fifteen years and knows every client, and the equity given to that person is the real price of keeping a two-decade client relationship in place.

Folders, rules and the corrections log

The folder structure is the part of the model the speaker treats as most useful. A thesis folder records which sector you are in and why, so a decision made in excitement cannot be defended later. The shared folder holds agent files, global rules, real accepted examples of finished work, and runbooks for anything that happens more than once. Each business has its own folder with its clients, its people and its local rules: a client who wants reports on the third of the month rather than the first because they asked for that in 2011 is the kind of detail that belongs there, because it is an accumulation of what this particular office has learned. The most important file is the corrections log. Every change a person makes to an agent's output is written down, repeated corrections become rules, and every approved rule becomes a test using that job's real input and the accepted output. The speaker ties this loop to a weekly rhythm and treats that hour as the most important of the week, because after a few hundred jobs the rule list is the only thing the model truly owns: anyone can use the same models, but nobody else knows every way this kind of business goes wrong. The approach also lines up with the evaluation infrastructure language engineers use, where production failures become targeted evals and the system corrects itself.

The work itself moves along a three-stage line. The first agent takes the job in, collects the documents and chases what is missing. The second does the work: categorising transactions, drafting a month-end close. The third checks the draft against the rules and sends it back if something does not add up. The rule the speaker returns to is that the reviewer can block but can never send, and the preparer cannot send either. Nothing reaches a client without human approval, and each agent file reads like a plain-English job description that states what it may not do. It is the single design decision that most reduces the risk in the whole model.

The weekly rhythm and the first deal

The system has a weekly rhythm. On Monday you check five numbers for every business: profit margin, human minutes per job, how often agent drafts need fixing, client retention, and whether the key people are happy and staying. On Tuesday you speak to each general manager separately. On Wednesday comes the hour that matters, when every correction a person made to agent work is reviewed, repeated ones are turned into rules, and each approved rule is written as a test with its real input and accepted output. Thursday and Friday go to finding the next business: coffee with owners, learning a sector. The first purchase is not found by shopping for an investment but by selling a service first.

The path to the first business starts by picking one sector, say bookkeeping, and doing its most painful job with agents: cleaning up month-end for the messiest clients, or chasing missing documents. Doing that work teaches you how the sector actually runs from the inside and gets you in front of its owners. Six, nine or twelve months later one of those owners will be ready to step back, and you will be the obvious person to take it over. The idea the section names is multipreneurship: not growing one company, but running several on a shared layer. The speaker has run his own holding company this way for years, and the quality of the general manager is what decides whether the system works.

The arguments against it

The speaker gives serious ground to the objections. The first is that roll-ups fail, and this is private equity with an AI sticker on it; the Wall Street Journal piece on General Catalyst makes the same point from the other side. Roll-ups do fail, and they fail for familiar reasons: paying too much, buying faster than you can integrate, letting the culture break. AI does not fix any of that on its own. His answer is that the small version buys slowly, does the first integration himself, and prices the business on what it makes today rather than paying the seller for AI upside that does not exist yet. The second objection is that once everyone uses AI the margins get competed away. Yes, eventually, but there is a window between your costs falling and prices in your industry catching up, and in several of these sectors that window is measured in years. The things that keep clients around, twenty-year relationships and a private list of rules for how the work gets done, are harder to copy than they look. The third objection is that these are regulated, high-stakes businesses where a human must check everything, so there is no saving. There is one: checking a draft takes far less time than producing it from scratch.

The objection the speaker treats as most serious is that staff dislike disruption and will walk out. Some do, and that is why clients see no visible difference during the opening month, why automation is switched on out of sight before it is trusted with live work, and why the staffer who holds every client relationship is installed as general manager and handed a proper equity stake. The last objection is that this is a dressed-up word for layoffs. His answer is that roles are moving from doing the work to checking it and each person handles more clients, and he is careful not to claim otherwise: some people will not make the transition, and he expects some of those roles to disappear. There are real risks, he concludes, but none of them changed the size of the opportunity he sees.