The automation band
The automation band is the range, 30 to 70%, of a vertical's work that AI can automate in which buying companies beats building one. Below it, AI changes too little to pay for the acquisition premium; above it, you would pay for people you are about to replace, so you build instead.
The rule
| Share of the work AI can do | What to do | Why | Example |
|---|---|---|---|
| Below 30% | Grow organically, or stay away | The margin gain is too thin to underwrite or compound | Tree trimming sits at about 30%: a fragmented market, right at the edge |
| 30 to 70% | Roll up | You buy real operations and customer relationships that AI makes far more efficient but cannot replace | CPAs, IT service providers (MSPs), property managers, call centers |
| Above 70% | Build | Acquired staff become cost you will automate away, and customers switch easily | Harper Insure, an AI-native broker at over 80%, grows on inbound alone |
The data exception. Above 70%, buying can still make sense when the companies' client relationships produce data you cannot get any other way. Then you are buying the panel, not the work.
Worked example
The deals that have closed sit inside the band. Crescendo bought the outsourcer PartnerHero and is taking gross margins on its 200 clients to 60–65%. Dwelly bought property-management agencies and doubled EBITDA margins where its software is fully deployed. Titan bought the IT provider RFA and cut user onboarding from weeks to minutes.
Now the opposite case. A founder of an AI-native consulting firm, at 50% automation and heading to 90%, was offered $5M to buy a traditional consulting firm. The band says no: by the time the deal integrates, he would be paying for consultants his own product replaces.
How it differs from
- General Catalyst's 30–70% finding. General Catalyst mapped 70 service categories and found 10 where AI can already automate 30 to 70% of the work (source). That is a measurement of where AI fits. The automation band turns it into a decision: roll up inside the range, build above it, stay away below it, with the data exception on top.
- Multiple arbitrage. A classic rollup buys at a low multiple and sells the bundle at a higher one. The band is about operational change: the margin has to come from AI doing part of the work.
Questions
How do you measure the share? Count the hours or tasks in a typical firm and mark the ones AI can do with a human checking the result. Bookkeeping entries, document intake and ticket sorting count; fixing a router or signing a tax return does not.
Does the band move? Yes. Every model release pushes the share up, so a vertical inside the band today can leave it within a few years. The bet is that the firms stay needed for longer than the rollup takes to build.
Why not buy above 70% at a low price? Because the customers come cheap for a reason. When most of the work is software, customers can switch to another provider in days, and the staff you bought become cost.
What holds customers after the deal? In order: faster, better service; the combined team's experience; more options; and only then the old relationship.
Origin
Set out by Viktoria Izdebska in Roll Up or Build: Why Emergence Capital Is Wrong About AI Rollups (6 July 2026) and in The Prediction Economy (2026). The data exception was added in October 2026.