How AI Becomes Economic Value—and Who Captures It

I've been looking at AI-enabled roll-ups, and one question won't go away: when the work gets faster, who actually makes the money?

Efficiency creates more capacity, not profit. Freed-up time has three exits: hire fewer people, serve more clients, or sell more services. If none of the three happens, the accountant just has three extra hours. Plenty of AI investment theses jump straight from "time saved" to "margin expansion," as if new capacity naturally belongs to shareholders. It can just as easily become employee slack—or price cuts that competition hands to customers.

Whenever supply grows faster than demand, productivity gains get competed away to customers. China's solar industry ran this live: module prices fell roughly 90% over the past decade while shipments multiplied many times over, and the sector still went loss-making. China's EVs are the same story: EV penetration hit a record 61.4% in April 2026, while the auto industry's profit margin sank to about 3.4%, down from around 9% a decade ago, touching 1.8% in the worst recent month. Better machines, better products, fast demand growth—and no profit, because supply and competition grew faster.

I divide this essay into two notes: what looks investable now (call it the low-hanging fruit) and the bigger directions still developing. This first note focuses on the former.

What Looks Investable Now

The low-hanging fruit: buy an asset that compounds without AI, then use AI to amplify it.

I set up three screens to help assess it.

01Undermonetized Monopoly Distribution, with AI as the Unlock

Monopoly is control of supply: whoever controls most of the supply in a market sets the price with customers. Everything else—network effects, physical density economics, regulation, aggregating fragmented supply—is only how you get there. One analogy in sports: turn the teams into a league, and the league negotiates the media rights as one entity to get better deals.

The asset worth acquiring has the control but hasn't monetized it. Customers, data, and payments are already in the system; the distribution is hard to replicate; the current product solves a fraction of customer needs; the price is small relative to the client's budget. Undermonetization persists for structural reasons—founders aging out, orphaned divisions inside larger companies, owners who ran a cost playbook instead of a product one, or adjacencies that simply weren't economic to deliver until now.

AI lowers the cost of software development, content creation, localization, and customer interaction. This creates two paths for revenue expansion:

1. The easy tier. Modernize the UI, smooth the interaction, add natural-language queries, make everything multilingual, and create more content such as courses and marketing materials. A better product experience built on monopolistic distribution supports an uplift in pricing.
2. The harder tier. With more effort and time, use insights gleaned from unstructured data to improve recommendations and matching, then create new products or business models to monetize.

02Fault Tolerance, or Cheap Verification

The common agent pitch of 2026 is "agents do the work, humans review it." This screen asks how much the reviewing costs.

LLMs fail on edge cases, and they fail confidently. The economics therefore depend on verification: what counts as correct, where the ground truth comes from, and how much it costs to obtain.

First, the ground truth. Three properties matter: whether the ideal response is easily defined, how fast the verdict arrives, and what a query costs. Coding maxes all three—the test suite defines "correct" in advance, the verdict arrives in seconds, and it costs cents.

Second, the stakes: what an error costs if it gets through.

The sweet spot is why coding fell to AI first: expensive engineers, cheap verifier. AI takes the generation seat, machines keep the checking seat, and humans review only what fails a gate.

03Uncrowded Acquisition Markets

Entry multiple impacts the return. When capital floods into roll-ups faster than deal supply grows, the return is handed to the seller on closing day.

Themes / Sectors Worth a Look

Applying the first two screens, below are several themes and sectors worth a look. The third—entry multiple—has to be tested deal by deal.

01Offline Operations with Regional-Monopoly Economics

Digitize fragmented operational information, then use optimization software to schedule, route, and coordinate physical work. Greater local density improves asset utilization, service quality, and unit economics.

• Waste managementRoute optimization based on service, weight tickets, disposal, and compliance data.
• Senior livingWorkforce scheduling across facilities, regional staffing pools, care documentation, and acuity-based staffing and pricing.
• Home careCaregiver scheduling, geographic matching, utilization, and compliance reporting.

AI can structure information from calls and notes, improve the front-end interaction experience, and increase pricing and efficiency.

02Community-Based Experiences Software for Children

Parents tend to cut spending on their children later than other discretionary spending, making this category recession-resilient.

LEGO offers a historical illustration. LEGO was born out of the Great Depression in 1932, when founder Ole Kirk Kristiansen pivoted from carpentry into wooden toys because that was the category that still made money during the recession.

Community relationships add another layer of resilience: parents care about children's friendships and sense of belonging through community sports. This creates continued engagement, high switching costs, and low sensitivity to price increases.

• Community sportsRegistration, payments, scheduling, streaming, and AI highlights.
• Youth activitiesDance, gymnastics, martial arts, and music, with scheduling, progress tracking, and parent communication.
• CampsSeasonal registration, personal recommendations, and annual re-enrollment.
• ChildcareDaily relationships with caregivers and peers, parent communications, and subsidy reporting.

03Marketplaces: Turning Unstructured Demand into Transactions

Marketplace information is scattered, unstructured, and often stale. Static documents capture stated requirements, but often miss underlying priorities, acceptable trade-offs, and real-time constraints. Legacy marketplaces struggled to keep both sides current or distinguish hard constraints from preferences, leaving human intermediaries to clarify intent, verify qualifications, and resolve exceptions.

LLMs and AI agents will increasingly change the economics of that workflow. AI can collect information from scattered sources, interpret how buyers describe their needs and suppliers describe their capabilities, and translate both sides into comparable requirements. It can distinguish hard constraints from preferences, identify missing information, verify qualifications against authoritative systems, proactively contact relevant counterparties, and propose matches or substitutes when terms do not fully align.

The next step is transaction execution. AI can increasingly coordinate qualification, quoting, scheduling, negotiation, contracting, and payment rather than stopping at a recommendation. The opportunity is turning fragmented, bespoke, and historically brokered workflows into repeatable transactions—and capturing part of the intermediary fee pool rather than only the software budget.

Examples

• Amex GBT / Long LakeA scaled corporate travel platform with established customer relationships, supplier connectivity, and transaction data; AI can improve service delivery and booking conversion.
• XometryCAD file upload to get price, lead time, and manufacturability feedback, translating an engineering requirement into a priced, executable order.
• BuildingConnectedConstruction drawings to bid packages, qualified subcontractors, and comparable bids, with a potential path into contracting and payment.
• Healthcare shiftsNursa and ShiftKey turn licenses and real-time availability into completed shifts, monetizing capacity that otherwise expires unused.

New wedges

• Data center and electrification tradesQualification, availability, dispatch, and payroll.
• Aviation MROTechnical RFQs, certification review, logistics, and settlement.
• Locum tenensLicensing, credentialing, hospital privileging, scheduling, and payment.

The sectors above share one important feature: AI does more than make existing work faster. It either helps a hard-to-replicate asset monetize more deeply, makes verification cheap enough to support automation, or turns fragmented demand into more completed transactions.

Back to Accounting

Accounting firms can be attractive traditional roll-up investments. But AI adds a huge amount of low-cost capacity to the industry. This is likely to lead to a barbell market. Large firms benefit from reputation and institutional credibility, while small independent practices can protect local niches through personal relationships and high-touch service. Mid-sized firms face pressure from both sides.

Over the past five years, I've seen and experienced an economic downturn firsthand in China. During this deflationary recession, large accounting firms have been able to hire more well-educated employees at low cost and continue winning work, although their profitability has still declined sharply. Smaller practices have been more exposed, losing clients both to bankruptcy and to larger firms. I see AI as a similar supply shock: an oversupply of skilled, well-educated labor at the price of tokens.

Having seen the disastrous results brought by the combination of overcapacity and a cut-throat competitive mindset in China, I've been cautious about the accounting-firm roll-up playbook.

The point is not that accounting firms cannot be good roll-up investments. It is that AI labor savings alone are not sufficient evidence of value creation. The investor still has to answer the original question: when the work gets faster, who actually makes the money?