Jonathon's AI Wiki

AI Roll-Ups — The Solo Holdco Blueprint (Corrections Log → Rules → Tests) (Greg Isenberg)

9 min read

Source: raw/5T_opportunity_-_AI_Roll_Ups.md — Greg Isenberg, Startup Ideas podcast, solo episode, youtube.com/watch?v=ZT4mpjx0JnE, YouTube. Auto-caption transcript fetched 2026-09-29.

Greg argues that three things together create a buying opportunity: owners are retiring, agents can now do back-office work, and service firms are priced as if their 5–10% margins can never change. Thrive Holdings and General Catalyst are already rolling up firms, and he says deals too small for them are open to solo founders. The part worth keeping even if you never buy a business is his operating layer:

  • shared agents and rules;
  • a corrections log for each business;
  • an intake → preparer → reviewer → human chain in which no agent can send anything;
  • a weekly prompt that turns human edits into rules and regression tests.

The fund figures are reported, not verified. Greg himself says they are “self-reported by young companies that are raising money.”

Key Takeaways

  • The thesis: three things at once.
    • Owners are retiring. Greg puts it at “$5 trillion” worth of businesses, a figure he attributes to McKinsey, with “about a million of those businesses” expected to sell by 2035.
    • “AI can do the work now.” He means typing numbers from PDFs, chasing the same missing document, writing the same status update. Set up correctly, “you get a draft that somebody checks.”
    • Margins look fixed. Service firms run at “5 to 10% profit” and are valued as if that can’t change. If the thesis holds, agents take margins to “30 or 40%” — “the same business making three to four times profit.”
  • What the funds are doing (Greg’s reporting, not checked here):
    • Thrive Holdings, spun out of Josh Kushner’s Thrive Capital, has bought “close to 50” accounting firms in 24 months and “committed another billion dollars.”
      • One of them is Larsson Gross (caption spelling) in Bellingham, Washington: founded 1949, five offices, 200 people.
      • It uses AI built on OpenAI’s Codex. This tax season it “processed 7,000 returns,” accountants saved “31% of their time on average,” and one task went from 180 hours a year to 15.
      • Interns who used to prepare returns now train to review them. “The people who used to do the work are becoming the people who check it.”
    • General Catalyst set aside $1.5 billion and has put “more than 750 million into at least 10 companies.”
      • Long Lake (property management) says it “bought 18 businesses and gotten to a 100 million in EBITDA in under two years with margins doubling.”
      • Crescendo (customer support centers) says “AI now handles about 90% of frontline tickets.”
      • Deal structure: “about 60 to 70% in cash and the founder rolls about 30% equity,” so the seller has a reason to make the handoff work.
    • Greg’s caveat: the numbers are “self-reported by young companies that are raising money and none of them have been through a recession yet.” He reads them as “a strong signal of where this is going, not as proof.”
  • The fund playbook, as Greg reads it:
    1. build the agents and software before buying anything;
    2. buy a firm that clients already trust;
    3. run agents in the background so clients see no change;
    4. once the agents are reliable, move over the back office (data entry, document collection, first drafts);
    5. plug the next acquisition into the same platform.
  • Why a solo founder has room:
    • A bookkeeping firm doing $2 million a year is too small for a billion-dollar fund.
    • Solo buyers have the same models: Thrive’s agents run on Codex, and “you can use Codex, Claude Code, Gemini.”
    • The founder personally does the integration, the hardest part of any roll-up.
    • Some owners would rather sell to a person than to a fund.
  • The holdco structure.
    • A holding company sits at the top.
    • Each business is run by a GM, usually an existing senior employee given “a real piece of the upside.”
    • Underneath is a shared layer — the same agents, rule system, back office and dashboards — so each acquisition is easier than the last.
    • Greg says he has run a holdco for six years, and that choosing the GM matters more than anything else.

The Operating Layer

Folders

  • thesis/ — which industries you’re in and why, and what a good business looks like to you. It keeps you from buying something because you got excited about it.
  • shared/ — what every business uses:
    • agents;
    • global rules;
    • “a folder of real accepted work that you use to test the agents every time you change something”;
    • runbooks for things that happen more than once, such as day one after a takeover, or what to do when an agent sends something wrong.
  • One folder per business:
    • clients: who has been with the firm 20 years, who is sensitive, who is loyal to a specific person;
    • people: who knows what;
    • rules: rules for this business only (his example: reports go out on the 3rd because one big client asked for that in 2011);
    • corrections log: “anytime a person changes something an agent did, it gets logged.”
  • If you set up only three files: “global rules, each business rules, and the corrections log.”

The Agent Chain

  1. Intake agent: collects documents and chases anything missing.
  2. Preparer agent: writes the first draft. In a bookkeeping firm that means categorizing transactions or drafting the month-end close.
  3. Reviewer agent: checks the draft against the rules and sends it back if something is off.
  4. A person approves: “usually the same person who used to do this work by hand.” Only then does it go to the client.
  • The rule Greg calls the most important: “The reviewer can block, but it can never send anything. And the preparer can’t send anything either.” The person is “the accountability layer.”
  • Every agent gets “a plain English job description.” It covers the agent’s job, what it may do, what it can never do, and when it must stop, ask and get a human. “Most of the agent problems I’ve seen come from nobody telling the agent where its job ends.”

Weekly Rhythm

  • Monday: five numbers per business.
    1. profit margin;
    2. minutes of human time per job;
    3. how often agent drafts need fixing;
    4. client retention;
    5. whether key people are happy and staying.
    • “If profit is going up while clients are leaving… something’s wrong.”
  • Tuesday: one call with each GM.
  • Wednesday: the corrections review, “the most important hour of the week” (prompt below).
  • Thursday and Friday: finding the next business. Listings on business-for-sale marketplaces rarely produce good deals unless the marketplace is curated; reach out to owners directly.

The Wednesday Corrections Prompt

As shown on screen (the caption’s “version of person” is corrected to “version a person”):

Compare each agent draft with the version a person approved this week. Sort every change into factual error, client preference, missing information or style. For any correction that happened more than once, propose a rule. Add each approved rule as a test using the original input and the accepted output.

Why it matters, in Greg’s words: “After a few hundred jobs, that list of rules becomes the most valuable thing you own because anyone can use the same AI models, but nobody else has your list of every way they go wrong in your kind of business.”

Getting the First Business

  1. Pick one industry; his example is bookkeeping firms.
  2. Sell those firms a service first: “take one annoying job and do it with agents,” such as month-end cleanup for their messiest clients or chasing missing documents.
  3. After 6 to 12 months of good work, one owner will be ready to step back, “and you’ll be the obvious person to take it over.”

Objections Greg Answers

ObjectionGreg’s answer
”Roll-ups always fail”; “private equity with an AI sticker”Roll-ups fail from overpaying, buying faster than they can integrate, or culture falling apart. Buy slowly, do the integration yourself, and price the business on what it makes today, not on the AI upside you plan to create
Margins will be competed away once everyone uses AIEventually, but there is a window between your costs dropping and prices catching up, and in many industries it is “years long.” Relationships and your rule list are hard to copy
Regulated, high-stakes work needs humans checking everythingChecking a draft takes much less time than making it from scratch
Staff and clients will leave (the one he takes “most seriously”)Change nothing clients can see in the first 30 days, run agents in the background before they touch real work, and give the GM real upside
It’s a fancy word for layoffsEarly examples show people moving from doing the work to checking it, each handling more clients. He expects some layoffs and says so

Try It

  • Start a corrections log on one agent output you already send out, such as client reports, blog drafts or review responses. Log every human edit.
  • Run the Wednesday prompt each week over that log. Add each approved rule as a test (the original input plus the accepted output) that the agent must pass before you change its prompt or model. This is the verifier-first discipline, with the test set built from your own corrections.
  • Take the send permission away from every drafting agent, so that approval is always a person’s action. It is the same gate as in the high-trust paperwork skeleton.
  • Track “how often agent drafts need fixing” weekly, next to margin and time per job.
  • Agencies can copy Greg’s entry path: sell one agent-run service to firms in a single industry. It has the same shape as a productized service such as the AI Tools Assessment business.

Open Questions

  • Unverified numbers. All fund and portfolio figures are Greg’s reporting of self-reported numbers; none were checked against primary sources. Firm names follow the captions: Larsson Gross, Long Lake, Crescendo.
  • The 31% figure. Greg first says accountants “saved 31% of their time,” then later calls it “Larsson Gross’s 31% margin.” ^[ambiguous] Read it as time saved.
  • Missing materials. Greg promised to post the folder structure, agent files and prompts in a pinned comment if the video reached 5,000 likes. They are not in this source.
  • No solo-operator numbers. The episode gives no purchase prices, financing terms or time to a first acquisition for a solo buyer.

Graph View

Backlinks