Every founder I talk to is building the same thing.
Not the same product, exactly. Same shape. A model wrapper, a vertical agent, a polite layer of orchestration on top of someone else’s intelligence. Slide one has a different industry on it. Slide seven is always the same architecture. Y Combinator’s Spring 2025 batch featured 144 companies, and 67 of them — nearly half — were AI agent startups [4]. The wave looks like software because the wave is loud, and because the people building in it are the ones writing the essays about it.
I think the wave is wrong. Or at least incomplete.
The durable money in a platform shift has rarely been concentrated in the layer everyone names after the platform. When the internet made information abundant, the winners were not “information companies.” They were live events, premium journalism, trusted curators, and the businesses that absorbed the demand abundance redirected. AI is going to do this again. The interesting question is not what AI can automate. It is what AI makes people need more of.
How Can 94% of Companies Get So Little From AI?
By the end of 2025, almost nine out of ten companies had deployed AI in at least one business function. Ninety-four percent reported not seeing significant value from those investments [2]. Deployment is nearly universal. Captured value is close to zero.
Wharton’s 2025 AI Adoption Report tells the same story from another angle. Weekly GenAI use inside the enterprise jumped from 37% to 72% in a single year. Spending was up 130%. Among teams using GenAI across business functions, only 58% rated the performance as “great” — a word doing a lot of work in a vendor survey [1]. Spending and use are rising faster than the reported outcomes. In older platform shifts, the unfinished work between buying a technology and benefiting from it was implementation debt. That work remains a durable business opportunity in every technology cycle.
Advisors, implementation partners, workflow designers, and trainers work in that gap. They help companies change the processes that must change before the technology can deliver its promise. None of those are AI companies. All of them get more valuable every time a CEO signs a six-figure AI budget and discovers nothing changed except the line item.
One stat is sharper than the rest. Non-AI-native companies use AI in direct sales at a 75% rate, versus 44% for the firms branded “AI-native” [6]. The label and the leverage have separated. The operating companies that quietly bolted AI onto their existing workflow are extracting more practical value than the firms that put AI on the homepage. The AI brand and the AI-extracted value are walking in opposite directions, and only the second one pays the bills.
What Does AI Make People Want More Of?
Every platform shift creates two opportunity zones, and most founders see only one.
The first is obvious: build in the new substrate. AI agents. AI software. AI infrastructure. This is the loud zone. It is also where competition is heaviest and where, as Bill Gates has noted, the agent market will be “exceptionally” competitive — investor-speak for cheap, plentiful, and unprofitable [7]. When you compete in the substrate, you fight everyone at once, including the people who own the substrate.
The second zone is the one most founders skip. When a technology makes one thing abundant, it raises the price of whatever that thing was suppressing. Information went free; attention got expensive. Streaming made recorded music cheap; live music became the most valuable product in the industry. Industrial automation made manufactured goods plentiful; handcrafted and artisan goods earned a premium they had never previously commanded [11]. The asset becomes scarce in inverse proportion to what the technology floods.
Take Live Nation. In 2010, Spotify was two years old and Napster’s ghost still haunted record-label boardrooms. The consensus was that live music would follow recorded music into the basement — why pay for a concert when the album is free? The opposite happened. Live Nation’s revenue grew from roughly $5.5 billion in 2010 to more than $22 billion in 2023, a 4x increase during the exact years streaming was supposed to be killing music as a business. Ticket prices in real terms doubled. The lesson is not that Spotify failed. It is that Live Nation built the business Spotify’s existence made more valuable.
Whatnot is the same story in a different decade. Founded in 2019, the live-stream commerce platform raised more than $385 million through 2023 at a valuation north of $3.7 billion [8]. It grew during the exact moment Amazon, Shopify, and every e-commerce player was rolling AI into product search, listings, pricing, and recommendation.
Whatnot did the opposite. It bet on real-time human sellers, live auctions, social chat, and the unpredictable energy of a person opening a graded card pack with two thousand strangers watching. The more algorithms automated standard commerce, the more valuable the un-automated layer became. A long-tail collectibles business is a strange answer to AI, until you notice that the answer to abundance is always what abundance cannot supply.
Now apply the lens to AI in 2026. If AI makes information abundant, human trust becomes more valuable — the advisors and curators you turn to precisely because you need someone to verify what the model just told you. If AI makes healthcare questions easier to ask, more people seek actual care, not fewer; primary care, screening, and triage absorb the released demand.
If AI makes everyone marginally more productive while marginally more overloaded, the categories that solve for time and attention rise: childcare, coaching, wellness, the concierge service that quietly removes friction from a life. And when AI floods every feed with content, the experiences that cannot be downloaded — live events, in-person community, the room — get more expensive, not less. These are not AI companies. They are businesses AI changes the price of.
So Did The Internet Help Or Hurt Local Businesses?
Both. And that is the point.
The same era that made Live Nation a giant flattened travel agents, record stores, local newspapers, and DVD rental. Abundance is not a blanket grant of wealth to everything human-centered. It is selective, and the line between “lifted” and “wiped out” is sharper than most thesis pieces admit. The distinction matters when choosing a business: which categories benefit from AI’s abundance, and which lose their reason to exist?
The test is simple. Is the human-centered business a complement to AI — AI lowers the friction to the first step, but resolution requires a human — or a substitute, where AI can do the job itself? Travel agents died because booking a flight was information work; the luxury and corporate niches survived because they sold judgment and access, not booking. Gyms survived because exercise is embodied — you cannot do it through a screen, no matter how excellent the screen. A category’s destiny depended on whether its core value was informational (vulnerable) or physical, relational, or regulated (protected).
MIT Sloan calls the surviving asset residual heterogeneity — what makes an organization distinct beyond the commoditized AI layer. As AI commoditizes, durable advantage migrates to whatever a competitor cannot easily replicate: proprietary data, a licensed clinician, an accumulated brand of trust, physical presence in a specific zip code [3]. The job-market lists confirm it from the other direction. Upwork’s catalog of 120+ jobs AI won’t replace and Microsoft’s research summarized by Inc. converge on one pattern [9]: the least exposed roles are not the most credentialed but the most embodied, regulated, and trust-dependent — phlebotomists, nursing assistants, primary care clinicians, skilled trades, instructors. The businesses that employ these workers are positioned to absorb AI-driven demand.
Amazon’s $3.9 billion acquisition of One Medical in 2023 is the most expensive bet on this logic any company has placed. The theory was not that AI health assistants would replace primary care, but that they would generate demand for it. AI lowers the friction of asking the question; the clinic resolves the need the question revealed.
Members come in earlier, more prepared, with more specific concerns, and convert into more care, not less [12]. The information layer became free; the trust layer got more valuable. If that mechanism plays out — and Amazon spent $3.9 billion betting it would — the same dynamic sits unbuilt in financial advice, mental health, legal guidance, any category where AI moves the asking and a human still has to do the answering.
But Won’t AI Companies Capture That Too?
This is the strongest version of the pushback, and it deserves a real answer. McKinsey is clear that the biggest expansion of profit pools in the AI era will come from business model reinvention, not incremental productivity [5]. The Harvey case in legal services is the cautionary example: if an AI-native player rewrites the unit economics of legal research deeply enough, it may capture the value that used to accrue to the law firm. The implication is uncomfortable — if AI software gets good enough at an adjacent layer, the AI-native player wins, and the operating business that owned the customer becomes the wholesale supplier of bodies and licenses.
That is true in some categories, which is exactly why this framework requires a defensibility test rather than blanket optimism. The thesis is not non-AI businesses always win. Second-order businesses win when they sit on a constraint AI cannot remove. Where the constraint is informational — the value was always just access to expertise AI can now match — the AI-native player likely captures it. Where the constraint is physical, regulated, relational, or embodied — a body in a chair, a licensed signature, trust accumulated over years, a hand on a wound — AI can lower the friction to the constraint but cannot dissolve it.
So the rule is not “build any non-AI business.” It is “build the businesses whose moat is precisely what AI cannot replicate, and use AI to lower the friction of everything else.” A primary care clinic with AI triage and human clinicians. A live commerce platform with AI-optimized logistics and irreducibly human sellers. A wellness studio with AI scheduling and a real coach in the room. That is the difference between Whatnot and a generic AI shopping app: one owns the human layer the algorithm cannot generate; the other competes with every model on earth, and every model on earth gets cheaper every quarter.
A Five-Step Map for the Adjacent Opportunity
If you are choosing where to build or invest in 2026, the gravitational pull will be toward the loud zone. This framework is mostly a way to de-select — to recognize when the second-order opportunity in your space is more defensible than the first-order one.
Call it the Second-Order Opportunity Map. Five steps.
-
Map the first-order AI disruption. Define the sentence: AI is making [X] abundant, cheap, or automated in [category]. If you cannot say it precisely, you cannot find the second-order opportunity.
-
Apply the complement-vs-substitute test. Does AI’s automation of the first-order task increase demand for the human-centered service, or destroy it? Information-only categories tend to be substituted; categories requiring presence, regulation, embodiment, or trust tend to be complemented.
-
Assess defensibility. Score the category on four dimensions: physical presence required, regulatory barrier, trust-and-relationship moat, embodied or specialized skill. Two or more high scores is the threshold.
-
Define the constraint AI cannot remove. Write it as a sentence: In [category], AI cannot remove the need for [constraint], and that constraint creates a scalable business because [mechanism]. A specific constraint lets you explain why a customer will still need the business after using AI.
-
Size the demand shift and find the entry point. Twenty customer interviews. One question: How has your use of AI tools changed how you think about [the human-centered service]? The answer pattern reveals where the complement dynamic is already live.
Practitioner investors already speak this language. The distinction circulating in venture commentary is between AI-native (the customer experience is AI; the brand is AI) and AI-enabled (the experience is unchanged; AI shows up in the operating model) [10]. The AI-enabled bucket is where the underwriting is interesting: hardware companies using AI for design, publishers using AI for distribution, healthcare access businesses using AI for triage. None are AI companies — all are AI companies if you read their gross margins instead of their marketing.
Most founders already know their category. A week of customer interviews helps them use this lens to choose between the AI layer and the adjacent human one, rather than trying to generate a business idea from the framework alone.
Build the Shore, Not the Wave
The wave is the AI software layer. The shore is where the wave changes what people want. Most founders are surfing. The bigger businesses may belong to the people who recognize which beaches are about to become beachfront property — provided the constraint underneath is real. Physical presence. Regulation. Trust. Embodiment. Anything less and the shore washes out with the next model release.
So stop only asking what AI can automate. The complementary question — what AI makes people need more of — is where the less obvious, more defensible businesses live. Live events. Primary care. Childcare. Coaching. Trusted advisory. The room where someone you trust looks you in the eye and tells you what they actually think, with no algorithm pretending to be them.
AI will get cheaper every quarter for the next decade. What gets more expensive is everything AI makes scarce by comparison. That, not the next agent, is the trade I would make.
References
[1] Wharton School, University of Pennsylvania. 2025 AI Adoption Report: Gen AI Fast-Tracks Into the Enterprise. 2025. https://knowledge.wharton.upenn.edu/special-report/2025-ai-adoption-report/
[2] McKinsey & Company. Where AI Will Create Value—and Where It Won’t. 2025. https://www.mckinsey.com/capabilities/strategy-and-corporate-finance/our-insights/where-ai-will-create-value-and-where-it-wont
[3] Davenport, Thomas H. and Mittal, Nitin. Why AI Will Not Provide Sustainable Competitive Advantage. MIT Sloan Management Review, 2024. https://sloanreview.mit.edu/article/why-ai-will-not-provide-sustainable-competitive-advantage/
[4] Epirus Ventures. Why the Next Big Startup Wave Must Go Beyond AI — The Case for Non-AI Innovation. 2025. https://www.epirus.vc/blog/why-the-next-big-startup-wave-must-go-beyond-ai-the-case-for-non-ai-innovation
[5] McKinsey & Company. Where AI Will Create Value—and Where It Won’t. 2025 (business model reinvention).
[6] Redlinger, Brandon. Why non-AI-native companies are ahead in AI adoption. LinkedIn, 2025. https://www.linkedin.com/posts/brandonredlinger_nonai-native-companies-are-adopting-ai-in-activity-7386784926278410240-7pqB
[7] Gates, Bill. The risks of artificial intelligence. Gates Notes, 2025. https://www.gatesnotes.com/meet-bill/tech-thinking/reader/the-risks-of-ai-are-real-but-manageable
[8] Whatnot funding and valuation data via Epirus Ventures (2025) and TechCrunch coverage. https://techcrunch.com/2023/07/31/live-shopping-startup-whatnot-raises-260-million-series-d/
[9] Upwork. 120+ Jobs That AI Won’t Replace. 2025. https://www.upwork.com/resources/jobs-ai-wont-replace; Crumley, Bruce. These 40 Jobs May Be Replaced by AI. These 40 Probably Won’t. Inc., 2025. https://www.inc.com/bruce-crumley/microsoft-study-identifies-jobs-most-and-least-affected-by-ai/91221483
[10] Preston McAndrew, Kate. Non-AI Companies: The Untapped Opportunity in Tech. LinkedIn, 2025. https://www.linkedin.com/posts/kateprestonmcandrew_the-biggest-untapped-opportunity-in-tech-activity-7404972790510018560-KoFZ
[11] Synthesis of historical platform-shift evidence; see Epirus Ventures (2025) and MIT Sloan Review (Davenport & Mittal, 2024). Internet → live events / premium journalism pattern.
[12] One Medical / Amazon acquisition context. CNBC, Amazon closes $3.9 billion One Medical acquisition. 2023. https://www.cnbc.com/2023/02/22/amazon-closes-3point9-billion-one-medical-acquisition.html