The Vertical AI Playbook

Authors: Kyle Harrison, Etienne Segal (Contrary Research) URL: research.contrary.com/report/the-vertical-ai-playbook Published: September 11, 2025 · ~42 min

One-line: AI’s margin upside is real but undistributed, so the interesting question is no longer whether to build the model but which ownership structure — sell to the operator, buy the operator, or be the operator — actually converts the technology into cash flow.

This is Kyle’s own deep dive, not a read of someone else’s. It is filed here because it is the canonical statement of the thesis the rest of the corpus argues with, and because the companion case study is QXO — A Vertical AI Case Study.

Key claims

• The adoption gap is the whole premise. Per S&P Market Intelligence, 42% of enterprise AI initiatives were discontinued in 2024, up from 17% in 2023 — at the same moment Dario Amodei was warning AI could erase half of entry-level white-collar jobs. The constraint is not model quality; it is that most non-technology firms cannot deploy.
• Three entry models, not two. (1) License software and let incumbents operate; (2) buy proven assets, install technology, recycle the cash; (3) build a full-stack company so code, capital and daily control sit under one roof. Most operators blend or pivot, but treating them as archetypes forces a clear first choice.
• AI expands the vertical TAM from software spend to labor spend. Traditional SaaS digitized the record of work; LLMs perform the work. US wages totalled over $11T in 2023, of which an estimated $4T+ is exposed to augmentation or automation — which is why markets that were too small for Vertical SaaS venture math are no longer small.
• Deploying AI is not shipping software. The Palantir answer — forward-deployed engineers embedded in the client, abstracting operations into reusable logic — is the reference implementation. AI is best understood as a new labor class: contextual, capable, and still requiring management.
• History rhymes, and it splits two ways. Strategic acquirers (Waste Management, United Rentals, XPO) integrate for synergy; financial acquirers (Berkshire Hathaway, Constellation Software) deliberately do not. Constellation grew from $165M revenue in 2005 to over $10B in 2024, ~30% CAGR since its 2006 IPO, by integrating companies financially only.
• Venture already ran this experiment and mostly lost. Thrasio raised $3B+, peaked near a $10B valuation in October 2021 doing ~1.5 acquisitions per week, and filed for bankruptcy in February 2024 when rates rose. The survivors — Metropolis (Company), Teamshares — were selective rather than volume-driven.
• The distribution wedge is often the real reason to buy. Metropolis had product-market fit and unit-economic fit and still could not get go-to-market fit against long-term parking management agreements — so it bought Premier Parking (600 garages, 2022) and then SP Plus for $1.5B in October 2023. Acquiring the incumbent turns customer inertia from an obstacle into a moat.
• Five-step playbook. I. Map the ontology (objects, states, transitions — find where time and headcount are absorbed). II. Define the terrain (mid-sized owner-operated niches with 150–200 viable targets reward ownership; micro-operator markets reward a SaaS wedge). III. Prove, then buy. IV. Test the distribution wedge. V. Match capital and talent to the path.
• The three hurdles that kill technologists in operator businesses: operational lift is real work, purchase price still decides returns, and deal/integration skill is scarce. “AI does not exempt acquirers from valuation discipline.”

Notable quotes

“The future is already here – it’s just not very evenly distributed.”

— William Gibson, the epigraph

“Suppose you believe that LLMs are now able to automate a lot of legal work. There are two things you might do with that idea. You could build an AI agent and sell it to law firms. That’s what most people do. Or, you could start your own law firm, staff it with AI agents, and compete with the existing law firms.”

— Y Combinator, 2025 Request for Startups

“only buy if the product absolutely rips (generates tons of value). You wouldn’t hire a sales team before building a product, and M&A is just a go-to-market motion here, so build then buy.”

— Slow Ventures, on sequencing

“Both models start with the same principles: (1) the average company in a given vertical isn’t run as effectively as it could be, and (2) you have the insights and playbooks to run the average company more effectively.”

— Matt Brown, on why PE and VC were never actually different in intent

“CSI’s strategy is to be a good owner of hundreds (and perhaps someday thousands) of growing autonomous small businesses that generate high returns on capital. Our strategy is unusual… We recognise that economies of scale, centralised management and world class talent competing in large and growing markets can be a great business-building formula. But, it isn’t what we do.”

— Mark Leonard, Constellation Software 2016 shareholder letter

“The easiest way to create tremendous shareholder value is to buy businesses at profit multiples lower than the multiple his own stock trades at, and then significantly improve those businesses.”

— Brad Jacobs

“Most AI rollup coverage highlights the venture firms pitching the idea, which makes it easy to cast them as the modern Berkshire Hathaways. However, in most cases, these funds are writing ordinary equity checks and moving on; more marketing than execution. The actual heavy lifting… sits with founders.”

“Today, 60% of all small businesses in America use vertical SaaS platforms to help run their business. Arborists use SingleOps, towing companies use Traxero… If you want to open a med spa, Moxie can get you going in 30 days.”

— Stripe, 2024 annual letter

“I am a better investor because I am a businessman and a better businessman because I am an investor.”

— Warren Buffett, the closing epigraph

The case studies, in one table

Vertical	Sell software	Buy the operator	Build AI-native
Real estate	EliseAI (350+ institutional owners, ~85% of conversations automated)	Metropolis (Company) (Premier Parking, SP Plus), Long Lake (18 HOA managers, $600M+ raised)	Wander
Accounting	Basis (top-100 firms, up to 30% time savings)	Crete ($300M+ revenue, 900 employees, 20+ firms), Multiplier (Citrine doubled margins with 12 people)	—
Legal	Harvey (Company) (300+ clients, $100M+ ARR by July 2025)	Eudia (acquired Johnson Hana, 300 people)	Atrium — the cautionary tale
Investment advisory	Inven (500+ firms)	—	OffDeal (two-person pods)
Contact centers	Replicant	Crescendo (bought PartnerHero; ~$90M revenue by May 2025)	—

The legal row carries the sharpest warning in the piece: “The best lawyers are the core of the business, and their trust, relationships, and judgment can’t be replaced by automation.” Atrium worked for high-volume repeatable work and broke on low-volume, high-stakes practice areas where attracting and keeping top lawyers decided the outcome.

How it connects

• QXO — A Vertical AI Case Study — the companion piece, and the incumbent-platform path the Playbook did not cover.
• AI Rollups — the concept page, which now carries the AI-native-firm discussion guide and the strongest published bear case (Benaich and Mrkšić on the multiple gap).
• AI-Native Firms — the org-design half of the argument; the Playbook argues the ownership structure changes, that page asks whether the org chart does.
• At Your Service — the services-as-software draft. Julien Bek’s copilot/autopilot split is the same taxonomy arrived at from the revenue-model side.
• Fifteen Accentures — the counter-argument worth holding: if services are structurally fragmented, the answer is a pack of firms, not one.
• Checking The AI Box — why the 42% discontinuation number may understate the problem. A pilot that was bought to satisfy a board mandate does not fail; it just never mattered.
• Constellation Software, Teamshares, Thrasio, Roll-up Strategy, Vertical SaaS, Vertical Market Software — the entity pages this piece leans on.
• Every Moat Becomes Moot — Kyle’s own published treatment of why the defensibility question here is about decay rate, not about whether a wall exists.
• The Mysteries of an Economic Engine — “tech-enabled” as a multiple-arbitrage story is exactly what that essay took apart, and it is the discipline this playbook’s step V demands.