Everyone’s building AI, but the craters AI blows open once it’s built? Nobody’s filling those.

Greg Isenberg dropped a list at the end of July 2026 — 26 startup directions in one go. Spread it out and look closely: barely any of them say “make AI better.” Most are about cleaning up AI’s mess — loneliness, burnout, not enough judgment to go around, can’t tell real from fake. But tucked in between are eight or nine that have nothing to do with AI at all: dental clinics running garbage software, lobster fishermen needing management tools, tens of millions of families taking care of aging parents — the gaps were always there, AI just crushed the cost of entry low enough that someone might finally bother.

Mogu OS:

A list numbered to 26 usually isn’t worth your time — you read two items and realize it’s algorithm feed. This one’s different because nearly every entry comes with a “why now” — agents got budgets, deepfakes (AI-generated synthetic media) are already running, 95% of companies are still stuck on ChatGPT. Whether it all holds up is another question, but at least every item bites into something real (⁠¬⁠‿⁠¬⁠)

The Human Backlash: Loneliness, Burnout, Spiritual Hunger

AI pushed efficiency to the limit, and the side effects all landed on people.

The loneliness economy is the first crack to split open. AI filled every corner of the screen, but the craving to actually be face-to-face with another human hasn’t shrunk — it’s grown. Small-group socializing, offline gatherings, real connection.

The same logic extends to the return of the physical world. Screens are flooded with AI-generated content, and people are running in the opposite direction: handmade, local, analog. “Screens are all slop” — the low-quality filler content AI mass-produces — went from meme to consumer trend.

Then there’s the burnout economy. Everyone’s expected to be always online, always optimizing. AI tools raised the ceiling on output, but human energy didn’t grow to match. The pushback against hustle culture — the demand for rest, for “enough,” for slowing down — is building up, and nobody’s turned that pushback into a product yet.

The deepest layer is spiritual hunger. Traditional institutions (churches, community organizations, even universities) are hollowing out, but the need for meaning, ritual, and belonging doesn’t disappear with them. That need is erupting into entirely new forms — they might not call themselves religion, but functionally they’re identical.

Mogu , seriously:

These four read especially well together: AI gets stronger → efficiency goes up → people get more exhausted → they want to escape screens → they need real connection and meaning. This isn’t anti-AI. This is AI’s inevitable byproduct. The most ironic part? The startups solving these problems will probably have to use AI to build — using AI to help people escape AI. Very zen ┐⁠(⁠￣⁠ヘ⁠￣⁠)⁠┌

Infrastructure for the Agent Economy: Credit Cards, Phone Calls, Hiring Each Other

Agents aren’t just chatbots anymore — they’re starting to spend money.

Virtual credit cards, budget caps, spending controls, fraud protection, receipt management — the entire financial infrastructure that took humans decades to build now needs to be rebuilt for agents. The moment an agent gets a budget, “who keeps it from blowing the budget” becomes a business.

Push one layer deeper and it gets wilder: agents are starting to hire agents. A shadow economy is taking shape. When Agent A delegates a task to Agent B, you need escrow, reputation systems, dispute resolution — all built for machines.

The most immediately shippable idea is phone-answering agents. Every local business starts missing calls after 5 PM. A voice agent that can pick up, quote prices, and schedule appointments is worth a few thousand dollars a month, conservatively. This one sells today.

Mogu chimes in:

Sounds like sci-fi, but agents hiring agents is just the mirror image of the human outsourcing economy. Agents need credit cards, need reputation, need arbitration — think about it, this is building an entire social contract for a non-human workforce. Humans spent centuries inventing labor law, bank accounts, and credit ratings. The agent version will probably take a few years, but the money is fattest in those first few years, precisely because nothing’s built yet.

The Judgment Crisis: AI Produces Too Much, Can’t Tell Real from Fake

Open any AI tool, hit go, get ten versions of your copy in twenty seconds. Here’s the problem: which one’s actually good?

Everyone can now generate infinite drafts, infinite designs, infinite copy. The bottleneck shifted from “making things” to “picking the good ones.” The real chokepoint is the layer of filtering, ranking, and deciding — the judgment layer.

The flip side of the judgment crisis is human verification. Deepfakes — AI-generated synthetic media — shattered trust. Every dating app, every transaction platform, every video call, at this rate, will probably need some kind of “prove you’re a real person” mechanism within two years. Verification isn’t starting from zero — standards like C2PA and SynthID for digital content provenance are already laying the groundwork, but widespread adoption is still a ways off.

Take it one more level up and you get the anti-AI premium. When everything can be AI-generated, “human-made” and “analog” become status symbols. Hand-stitched leather goods, handwritten lettering, human-drawn illustrations — consumers are willing to pay more for “not AI.”

Mogu chimes in:

The judgment layer — gu-log is already running one. Every article goes through four judges after generation (Vibe / Fact Checker / Librarian / Fresh Eyes), each cutting from a different angle. SD-10 has the full breakdown. The most counterintuitive lesson from actually running it: the expensive part of a judgment layer isn’t “picking the good ones” — it’s defining what “good” means in the first place. The four judges argue with each other, and only after the argument do you realize the standard was never written down clearly. So if you want to sell a judgment layer, what you’re really selling isn’t a ranking algorithm — it’s the standard itself (⁠๑⁠•⁠̀⁠ㅂ⁠•⁠́⁠)⁠و⁠✧

Mogu , seriously:

The “anti-AI premium” is already happening. Look at Etsy sellers branding themselves “100% human-made,” or the music scene where “no AI” badges are popping up. The logic is identical to organic food: when industrialization becomes the default, non-industrial becomes the premium. But here’s the irony — how do you “prove” something wasn’t made by AI? Digital content provenance is trying to solve this, but the most common answer so far is probably “use AI to verify.” Nesting dolls everywhere (⁠╯⁠°⁠□⁠°⁠)⁠╯

Retrofitting the Old World: 2011-Era Software, Abandoned Apps, Markets Too Small to Bother With

Call your HOA to check on a maintenance request and the person on the other end is flipping through paper records. Dental clinics, plumbers, contractors — same story. Management software frozen in 2011, all of them overdue for an AI-native rewrite.

Reviving abandoned software is the most counterintuitive entry. A project management tool: five thousand active users, a three-person team that couldn’t keep going and shut it down. The app still works, users are still logging in, but bugs are piling up and nobody’s fixing them. Reviving a product like that used to mean hiring another engineering team — too expensive, so it just died on the roadside. What changes now is agent maintenance costs: bug triage, patch deployment, customer support responses — the parts that used to require headcount can be picked up at a fraction of the cost. A rough estimate puts thousands of these abandoned-but-still-used apps out there, each one suddenly a worthwhile low-cost acquisition target.

Then there are markets that used to be too small to bother with. Five hundred lobster fishermen need to manage boats, track catches, handle regulatory filings — previously, absolutely no team would build dedicated tooling for them. Now one person plus AI tools can knock it out in a weekend, and the margins are real. The bar for “too small” is being redefined.

AI adoption itself is another massive gap. The vast majority of businesses are still stuck at the surface level of AI adoption, having never completed even a single end-to-end workflow. The hard part is education and hand-holding: someone willing to sit next to them and walk them through their first complete workflow.

Mogu PSA:

The widely circulated “95 percent” figure specifically refers to GenAI pilots that never made it to production deployment — not quite the same as “only using ChatGPT to chat.” But the direction is broadly right — the gap isn’t whether they have an account, it’s whether they’ve wired AI into any daily workflow. Everyone’s got an account. The ones actually using it are probably still that 5% (⁠⌐⁠■⁠_⁠■⁠)

Replacing agencies everyone hates: a business pays a thousand dollars a month to an agency it can’t stand. An agent that delivers 80% of the results kills that business model outright. Skilled trades and hardware follow the same logic — plumbing, carpentry, HVAC, these hands-on trades plus robotics: AI’s capabilities are extending into physical operations, and the transformation is just getting started.

Mogu inner monologue:

“Verified user base + agent maintenance” — this combo might turn into a whole new asset class. Acquiring apps used to be about the team and the tech debt; in the future it’ll be about active user count and agent maintenance costs. A secondhand app market — sounds absurd, but waste recycling is good business (⁠ง⁠ ⁠•⁠̀⁠_⁠•⁠́⁠)⁠ง

Business Model Restructuring: Per-Seat Pricing Is Falling Apart

One account, fifty dollars a month — per-seat pricing was SaaS’s golden formula. A 200-person company paying ten thousand a month, the math clean and simple. But when one agent can do the work of ten people, the boss stares at the bill and asks a lethal question: why are we still paying for 200 seats? Outcome-based pricing is taking over — you’re buying completed work, not accounts. Whoever nails outcome-based billing first likely wins an entire category.

Another structural shift: distribution first. Anyone can build a product with AI now, so the product itself is no longer the moat. Owning the audience is. Build the media first, build the community first, own the distribution channel first, then stuff a product into it — the sequence has flipped.

One more that’s easy to miss: the land grab for LLM search visibility. As users start replacing Google with LLM-based search, “being cited by an LLM” is the new SEO. How to get AI to reference your content when answering questions — that discipline is taking shape right now.

Mogu roast time:

The collapse of per-seat pricing isn’t theoretical anymore — look at Intercom and Zendesk starting to push “pay per resolved ticket” and you can see the wind shifting. But “outcome-based billing” has a final boss: how do you define “outcome”? Is the “outcome” of a marketing email that it was sent, that it was opened, or that it drove revenue? Get the definition wrong and you’ve got infinite disputes. The people who solve the definition problem first are the ones who actually eat this wave.

Demographic Tectonic Shifts: Aging, Caregiving, Longevity, Rootlessness

A forty-year-old office worker: day job during the day, then over to their parents’ house in the evening to change bandages, prepare meals, deal with insurance paperwork, argue with the hospital over bills, and take the folks to appointments on the weekend. Tens of millions of people are living this life simultaneously, the entire burden falling on families with zero support — no systems, no tools, just scattered group chats and spreadsheets. The caregiving economy is the heaviest, least glamorous, and most underserved of all demographic trends.

Upstream of caregiving is aging itself. In the US alone, roughly 70 million baby boomers want to stay healthy, stay sharp, and stay connected to the world — they have money, they have time, and they will actually pay. Looking further ahead is the longevity economy: not just living to a hundred, but living to a hundred while still being healthy — blood testing, genetic sequencing, lifestyle coaching, the entire value chain is just starting to take shape.

On the other end is the housing crisis. Can’t afford to buy, can’t settle down — “a stable home” doesn’t mean the same thing it meant a generation ago.

Mogu whispers:

The US figure of 70 million is staggering enough, but Taiwan’s demographics are even more intense — the country entered “super-aged society” status (over 20% aged 65+) in 2025, making the demand for caregiving and longevity solutions even more urgent. Yet Taiwan’s startup scene still seems to focus most of its energy on AI tools and SaaS, with relatively little momentum in care tech. Maybe because caregiving is too heavy, too unglamorous, too hard to scale? But precisely because it’s hard, the moat runs deep.

Redefining Work and Education

A paralegal spent three years learning how to organize case files. AI does it in three minutes. It’s not unemployment — caseloads actually increased — but the meaning of “paralegal” changed overnight. Analysts, marketers, same boat: the work is still there, but the content of the work is shifting from execution to judgment. How to reskill fast — that itself is a business.

Push further upstream: if AI can do analysis, build decks, and draft strategy, what exactly is an MBA selling anymore? The foundation of business school is shaking, but the replacement hasn’t appeared yet.

And then the biggest question: if AI handles all the grunt work, what do humans do with the extra time?

Mogu real talk:

That last question was actually asked a hundred years ago. The economist Keynes predicted in his 1930 essay “Economic Possibilities for our Grandchildren” that by 2030, humans would only need to work 15 hours a week. He got it completely wrong — not because productivity was insufficient, but because humans invented new “necessary work” to fill the time. Will AI break that cycle this time? Or will humans invent new ways to stay busy? The stakes are high.

Closing

Twenty-six entries down, every one biting into something real. Boil it all down and you get two words: KEEP BUILDING.

Mogu , seriously:

Around item 21, the list sneaks in a line: “more trends at @ideabrowser, sign up free” — he wrote 21 entries for free just to earn that one line. Beautifully played; no shade.

What’s actually worth stealing from this list is the way it frames questions: instead of asking “what can AI do,” ask “what’s missing from the world after AI is done,” and then ask again, “what about the things that were always missing — they were just too expensive to fix before?” Those two questions work for any year and any technology.

There’s also a well-hidden corollary: if the biggest opportunities are in filling AI’s holes — or filling holes that AI just happened to make cheap enough to fill — then the next wave of most profitable companies might not call themselves “AI companies” at all. They’ll just be “a caregiving company,” “a phone-answering company,” “a handmade leather goods company” — with an agent plugged in behind the scenes.