Software Will Eat Services by Buying Them

Software spent three decades selling tools to businesses and capturing a thin slice of what those businesses earned. AI changes the arithmetic. When software can do the work rather than assist the people doing it, the largest prize is no longer the subscription. It is the service business itself. This post argues that the most direct way for software to capture that prize is to buy it, lays out the math, the evidence, and the failure modes, and states the conditions under which the thesis is wrong.

Committed 2026-10-31: the next generation of software companies will own the services they automate, financed conservatively and held for the long term. Distribution and data are the moat; the model is not.

Principles

Six lines that summarize the argument.

Sell the outcome, not the tool. The value is in the work done, not the software that helps do it.
AI is not the moat. Owned customers, proprietary workflow data, trust, and licenses are.
Pricing power decides who keeps the savings. Without it, efficiency gains are passed to customers.
Never over-lever. Operational excellence does not survive a capital structure that cannot absorb a bad year.
Trust is the whole asset. It compounds slowly and collapses fast.
Hold, do not flip. Where exit markets are thin, permanent ownership is the safer default.

Software Is Eating Services

The phrase “software is eating the world” described software replacing industries. In practice, most software did something narrower: it sold tools to the businesses that did the work. An accounting firm buys bookkeeping software; the software vendor earns a subscription; the firm earns the fees for the actual bookkeeping. The vendor captured a sliver. The service captured the rest.

The size gap is enormous. General Catalyst frames global services as a market of roughly $16 trillion a year against roughly $1 trillion for software. Software margins are what made the smaller market attractive. Services margins are what kept the larger market unattractive.

Citation: “The AI services transformation may be harder than VCs think,” TechCrunch, Sep 2025.

Consider a services firm with revenue $R$ . A software vendor selling it tools captures a small fraction $\varphi$ of that revenue:

V_{\text{vendor}} = \varphi \cdot R

An owner of the same firm captures its entire margin $\mu$ :

V_{\text{owner}} = \mu \cdot R

The ratio $V_{\text{owner}} / V_{\text{vendor}} = \mu / \varphi$ is the whole argument. While services margins stayed thin, $\mu$ was small and owning the firm was a mediocre business. Selling it software was better. AI breaks this by raising $\mu$ . If a firm spending 2% of revenue on software moves from an 8% margin to 25% through automation, owning it captures over twelve times what selling to it does.

Selling to the firm captures a sliver of its revenue; owning it captures the whole margin. Automation is what made that margin worth owning: at 25% it is over twelve times the vendor's 2%, inside a services market roughly sixteen times the size of software's.

Once software can do the work, the rational move for the software company is to stop selling to the firm and start owning it.

Why Ownership Beats Selling - A Personal Lesson

Engineering problems have a pleasant property: effort converts to progress. Write the pipeline, fix the bug, train the model, measure the lift. Distribution does not behave this way. Leads, partnerships, and trust do not respond to raw effort; they respond to relationships, timing, and reputation, which accumulate on a different clock.

A software startup asks its founders to solve both problems at once, usually with the second one on the critical path. A technically strong team can build a product faster and cheaper than almost anyone, and still stall because the number of businesses willing to buy from an unknown vendor grows linearly at best.

An acquisition inverts the order. The customers, the contracts, the revenue, and the operating history already exist. What remains is an optimization problem: make an existing system run better. That is the kind of problem engineers are built for.

Buying the business changes the order of the two problems, not their number. The distribution step does not vanish: it moves from finding customers to finding sellers and lenders, and it still comes first.

This is the operator’s version of the macro argument above. It is also, as later sections show, only half true.

The Ownership Models

“Private equity” covers several businesses with very different day-to-day lives. Choosing between them matters more than the label.

Each cell is ranked 1 (best fit on that row) to 4 (worst) for a technical operator with limited starting capital.

Decision metric	Buyout fund (GP)	Search fund	Vertical software holdco	AI-enabled roll-up
Capital needed to start	4 · requires LPs who back track records	2 · small search capital, then acquisition financing	3 · real capital, or cash flow from an existing business	3 · capital plus a working automation platform
Day-to-day operating load	1 · board-level; operators are hired	4 · the searcher becomes the CEO for 3 to 7 years	2 · decentralized; acquired teams keep running	3 · heavy during integration, lighter after
Leverage of technical skill	4 · mostly financial engineering and fundraising	3 · general management dominates	2 · understands the product and its economics	1 · the automation layer is the value creation
Exit dependence	4 · fund life forces a sale in 5 to 10 years	3 · investors expect an eventual sale	1 · built to hold forever	2 · can hold or sell
Network dependence	4 · LPs, bankers, sellers	2 · one seller at a time	2 · founders selling small companies	3 · many sellers, repeatedly

No column wins every row. The holdco and roll-up columns are where software capability is an unusual advantage; the buyout column is where a finance career is the entry ticket.

The Value Equation

Every model above creates value through the same identity. Enterprise value is a multiple of earnings:

EV = m \cdot \text{EBITDA}

and equity value is what remains after debt:

E = EV - D_{net}

The change in equity between purchase and today decomposes into three levers:

\Delta E = \underbrace{m_0 \cdot \Delta \text{EBITDA}}_{\text{operations}} + \underbrace{\text{EBITDA}_1 \cdot \Delta m}_{\text{rerating}} - \underbrace{\Delta D_{net}}_{\text{deleveraging}}

Classic leveraged buyouts leaned on the third term: borrow heavily, pay the debt down with the company’s own cash flow, and let the equity slice grow. Modern private equity leans on the first: improve the business itself. The EY Southeast Asia review describes the regional market moving toward exactly this, value-creation-led ownership with more control and operational improvement.

AI-enabled ownership adds a second engine to the rerating term. A services firm that begins to earn software-like margins may eventually be valued like one, so $\Delta m$ becomes positive alongside $\Delta \text{EBITDA}$ . The operations term is still the foundation: it is the only one that does not depend on cheap debt or a generous buyer.

Equity grows on three levers. Operations is the foundation, the only one that needs neither cheap debt nor a generous buyer; AI-enabled ownership adds a second engine by letting a services firm be rerated like software.

The AI-Enabled Roll-Up

Labor-heavy services businesses (IT support, bookkeeping, back-office processing, call centers) typically run thin margins because revenue scales with headcount. If a meaningful share of the work can be automated, the margin structure changes.

Let the labor share of cost be $\lambda$ and the automatable fraction of labor be $a$ . With an initial margin $\mu_0$ , the post-automation margin is approximately

\mu_1 \approx \mu_0 + a \cdot \lambda \cdot (1 - \mu_0) - c_{AI}

where $c_{AI}$ is the cost of running and supervising the automation, as a share of revenue. A services firm at an 8% margin, with labor at 60% of cost and 35% of that labor automatable, moves to roughly 27% before AI costs. On a fixed revenue base, EBITDA more than triples.

Automation moves a slice of labor cost straight into margin, but the owner keeps only what is left of that slice after the cost of running and reviewing the automation, which the example does not estimate.

This is no longer a fringe idea. General Catalyst dedicated $1.5 billion to incubating AI-native vertical companies that then acquire established services firms and their customers, across sectors from legal services to IT management. One of them, Titan, raised $74 million, bought an IT managed services provider, and reported automating 38% of typical tasks in pilots. The stated target is moving services firms from the 5 to 10% EBITDA typical of the sector toward the 30 to 40% associated with software.

The legal sector shows the shift from tool to outcome most clearly. Eudia, another General Catalyst company, sells fixed-fee legal services powered by AI to in-house legal departments instead of billing by the hour, then acquired an alternative legal services provider to extend its reach. It does not sell lawyers a tool. It sells the legal work, and it bought a firm to deliver more of it.

Citation: “The AI services transformation may be harder than VCs think,” TechCrunch, Sep 2025 · “The General Catalyst Behind $1.5 Billion of AI Roll-Ups,” Capital & Clarity.

In this model the automation layer is not a supporting function. It is the source of returns.

The Moat Is Not the Model

The tempting conclusion is that better AI is the competitive advantage. It is not. Foundation models are rented by everyone. Whatever automation is possible this year will be available to every competitor next year, and the model providers themselves can sell directly to services firms. Neither the model nor the vendor position is safe. Only ownership of customers and data is.

The savings are only kept if the business has pricing power. Let $\pi \in [0, 1]$ be the fraction of cost savings the owner retains rather than passing to customers through lower prices:

\mu_{\text{retained}} = \mu_0 + \pi \cdot (\mu_1 - \mu_0)

In a commoditized market, competitors automate too, prices fall, and $\pi \to 0$ : the entire efficiency gain flows to customers. The margin expansion in Eq. 06 is real but temporary. In a market with switching costs, regulatory licenses, or deep trust, $\pi$ stays high.

The same gain from automation, split two ways. Where competitors automate too, prices fall and customers keep nearly all of it; the owner keeps the savings as margin only where switching costs, licenses or trust give it pricing power.

What actually raises $\pi$ :

Moat source	Why it holds	Erodes when
Owned customer relationships	the acquisition bought them; switching is costly	service quality drops during integration
Proprietary workflow data	only the operator of the service sees it	data is not captured or structured
Licenses and trust	a licensed professional or trusted firm must sign off	the license-holder leaves, or trust is broken
Scale across many firms	one platform amortized across a portfolio	integrations stall and each firm stays bespoke

Every competitor can rent the same model, so the model cannot be the moat. Pricing power lives in what the owner holds around it, and each of those holdings has its own way of wearing away.

The model is the engine. The moat is everything around it.

From Financiers to Software Owners

The history of ownership in software runs in three steps, and the roll-up is the third.

Who	Background before buying	What they built	Lesson
Robert F. Smith	chemical engineer with patents, then six years of tech M&A at Goldman Sachs	Vista Equity Partners (2000), enterprise software only, now over $100B under management	financiers can own software and standardize how it is run
Mark Leonard	eleven years in venture capital	Constellation Software (1995), seeded with $25M from a pension fund; over 500 acquisitions	small vertical software firms are too small for VC and perfect for permanent owners
Andrew Wilkinson	founded a design agency at 20	Tiny, a holding company for internet businesses, funded by the agency’s cash flow; listed in 2023	own cash flow removes the need to convince LPs first
Sandiaga Uno & Edwin Soeryadjaya	Uno lost his job in the 1997 crisis; Soeryadjaya came from the Astra founding family	Saratoga (1998), buying distressed Indonesian assets after the crisis	a crisis turns desperate sellers into opportunity; network made the deals possible
General Catalyst roll-ups	AI-native vertical startups	Titan, Eudia, and others acquiring services firms	software now buys the services it automates
KKR, Bain, Vornado (Toys “R” Us)	the largest names in buyouts	$1.3B equity and $6.2B debt (2005); bankruptcy in 2017	no operational fix survives heavy leverage in a declining sector
Abraaj Group	a celebrated emerging-markets PE firm	built over sixteen years; collapsed in four months amid fraud allegations	in emerging markets, governance and trust are the entire asset

Read top to bottom, the progression is the title’s story. First, financiers learned to own software (Vista). Then software owners learned to hold software forever (Constellation, Tiny). Now software is learning to own the services it used to sell to (the AI roll-ups).

Software learned to own in three steps: financiers showed software could be owned, permanent holders showed it could be held forever, and the roll-ups now turn that ownership on the services software used to sell to. The two failures sit beside the path rather than on it, and either one ends the climb at any step.

Two patterns run through the successes.

The network came before the purchase. Every success had a source of deal flow and capital before buying anything: a bank, a VC career, an operating business, a family network. Ownership does not eliminate the distribution problem. It relocates it, from finding customers to finding sellers, lenders, and co-investors. Sellers of good businesses do not hand them to strangers.

Almost every success was a pair. Smith had a co-founder, Uno had Soeryadjaya, Wilkinson had Chris Sparling. One partner thinks in systems and operations; the other brings relationships and deals.

The failures are just as instructive. Toys “R” Us is the leverage failure: the reported interest bill ran around $400 million a year, and the company kept underinvesting in stores while competitors took its customers. The interest coverage ratio

\text{ICR} = \frac{\text{EBITDA}}{r \cdot D}

must stay comfortably above one even in a bad year. A business whose revenue is declining structurally will eventually push ICR below one, no matter how well it is run. Abraaj is the governance failure: the operating model did not matter once investors stopped trusting the reported numbers.

Interest is fixed when the debt is signed; the earnings of a declining business are not. Once the two lines cross, no operational fix survives an interest bill the business cannot cover. Conservative debt turns the same bad year, and the same decline, into something the business can absorb.

Where It Breaks

The thesis has five known failure points.

Each stage of the playbook has its own way to fail and its own condition that holds it. The rust numbers match the list below.

Savings get competed away. Eq. 07 with $\pi \to 0$ . Target sectors with pricing power, or the margin expansion is a one-time gift to customers.
AI output still needs review. A September 2025 survey of US desk workers by BetterUp and the Stanford Social Media Lab found 40% had received polished-but-hollow AI output (“workslop”) in the previous month, and each case took close to two hours to clean up (BetterUp). Cut too many people and errors go uncaught; keep them all and the margin gain never arrives. The automation has to be measured, not assumed.
Leverage. Eq. 08. Conservative debt or none at all, especially early.
Ownership rules. Many regulated professions restrict ownership to licensed practitioners. The most labor-heavy targets (accounting, legal, notarial work) may not be ownable outright in some jurisdictions, which forces partnership structures or rules them out. This must be checked sector by sector before any target list is built.
Integration. Roll-ups that acquire faster than they integrate end up as a collection of bespoke firms with a shared logo, not a platform. Scale economics only appear when integration is actually done.

The Indonesian Variant

Southeast Asian private equity became more selective in 2025: $9.1B deployed across 59 deals, down from $16B across 67 a year earlier, with Singapore accounting for over 74% of value. Within Indonesia, activity shifted toward consumer, healthcare, and financial services, with investors prioritizing cash generation and scalable operating models.

Citation: “Southeast Asia Private Equity Pulse 2025: year in review,” EY, Feb 2026.

Three local facts shape the strategy.

Exits are thin. As argued in Why Innovation Does Not Live in Indonesia, the domestic exit market is close to IPO-only, and PE-backed listings remain rare. A strategy that depends on selling in five years depends on the weakest part of the ecosystem. A hold-forever structure removes that dependency.

The AI roll-up is early here. Almost all visible activity is in the United States. Indonesian services markets are fragmented, labor-intensive, and largely untouched by this playbook.

Trust costs more and is worth more. High transaction costs make relationships expensive to build, which is exactly why they are a durable moat once built.

What This Means for Software Vendors

If owners capture $\mu R$ while vendors capture $\varphi R$ , the vendor’s position weakens from two directions at once. AI agents reduce the number of human users inside each customer, which erodes per-seat pricing. And a growing share of the customer base may itself be acquired by owners who build their own automation rather than renting tools.

Vendors then face a choice: price on outcomes instead of seats, become owners themselves, or shrink. That argument deserves its own treatment, in a forthcoming post, The Last Generation of Software Vendors.

What Would Falsify This

A thesis that cannot be wrong is a belief. These are the conditions that would break it.

Vendors capture outcomes without owning firms. If software companies successfully price on outcomes (per resolved ticket, per closed book, per contract reviewed) and capture most of the value without buying anyone, ownership is unnecessary. Software would eat services by pricing, not by buying.
No sector passes both filters. If no services or software sector has both meaningful pricing power and ownership rules that permit control, the roll-up variant fails in that market.
The automation share is small. If careful measurement on real workflows shows $a \cdot \lambda$ in Eq. 06 too small to move margins after review costs, the margin math does not work.
Integration does not scale. If acquired firms resist standardization so strongly that each one stays bespoke, the platform economics never arrive and the roll-up becomes an expensive collection of small businesses.

Open Questions

First capital. Own cash flow, a partner’s capital, or outside investors? Each changes the structure and the pace.
The partner. Who brings the seller network and the commercial muscle?
Sector shortlist. Which sectors pass pricing power, ownership rules, fragmentation, and automatable share at the same time?
Hold or roll. A single permanent acquisition, a vertical software holdco, or a multi-firm roll-up?

Appendix - For the Operator Who Wants to Do This

Expected skills

An operator-owner, whether an operating partner at a fund or an executive installed in a portfolio company, needs a broader stack than the operations term alone suggests. The weights below are estimates of relative importance and time, not survey data.

#	Skill area	Weight	What it covers
1	People and leadership	20%	assessing and replacing management, hiring fast, incentive design, leading change among people who did not choose the new owner
2	Value creation diagnosis	20%	finding the 3 to 5 levers that move enterprise value, the 100-day plan, KPI trees and operating cadence, choosing what not to fix
3	Finance and value linkage	15%	reading the three statements fluently, working capital, unit economics, EBITDA bridges, translating operations into IRR and MOIC, debt covenants
4	Commercial excellence	15%	pricing (often the fastest lever), sales effectiveness, segmentation, churn reduction
5	Operations, cost, and technology	15%	process redesign, procurement, org design, data infrastructure, automation and AI
6	Deal support and integration	10%	operational due diligence, integrating add-on acquisitions in buy-and-build strategies
7	Governance and communication	5%	board reporting, lender and investor confidence, exit readiness

The uncomfortable observation: technical skill lives mostly in row 5 and parts of rows 2 and 6, roughly a fifth to a quarter of the job. The two heaviest rows, people and commercial, are the same distribution and persuasion skills that make selling software hard. Ownership does not route around them. It demands them at higher stakes.

For software targets specifically, row 5 expands into its own stack: SaaS economics (ARR, gross and net revenue retention, CAC payback, Rule of 40), pricing and packaging, cost to serve, engineering productivity, and services economics (utilization, realization) where software is bundled with implementation.

A 24-week curriculum

Roughly 5 to 7 hours per week, two phases of twelve weeks, ending in one artifact that tests the whole thesis.

Weeks	Focus	Inputs	Output
1 to 4	Financial statements	Financial Statements (Ittelson); annual reports of two IDX-listed companies	a one-page breakdown of each: revenue model, margins, cash conversion, debt
5 to 8	Valuation and unit economics	Damodaran’s free NYU valuation course	a DCF and a comparables valuation of one company
9 to 12	Deal mechanics	an LBO modeling course; Stanford GSB search fund primer; Buy Then Build (Deibel)	an LBO model of a hypothetical small Indonesian services firm
13 to 15	SaaS economics	Skok, “SaaS Metrics 2.0”; Bessemer benchmarks	a metrics dashboard specification for a small software business
16 to 18	Pricing and retention	Monetizing Innovation (Simon-Kucher); The SaaS Playbook (Walling)	a pricing teardown of 3 to 5 Indonesian vertical SaaS products from public pages
19 to 21	Cost structure and services	Accelerate (Forsgren et al.); Managing the Professional Service Firm (Maister)	a cost-to-serve model with the automatable share quantified
22 to 24	The holdco playbook	Constellation Software shareholder letters; public material on Vista’s operating model	a 100-day plan for a hypothetical acquired vertical software firm

Capstone: merge the LBO model and the 100-day plan into one investment memo on a real Indonesian business. What to pay, how to finance it, what to change, what it is worth in five years. In parallel throughout: own one revenue number somewhere, because commercial skill is learned by reps, not reading.

For thirty years, software stood beside the work and charged rent for the tools. AI lets it step into the work itself, and the cleanest way to do that is to own the business where the work happens. The distribution problem does not disappear in the process; it moves from customers to sellers, from leads to lenders. But what software holds while solving it changes: a working company with cash flow, customers, and data, instead of a subscription that can be cancelled. Software stops being a tool sold to businesses. It becomes the business.