13 businesses that actually make sense in the age of AI agents

A window-washing robot that costs less than a human crew and never files a workers' comp claim. A sauna club that charges membership fees for something you could technically do at home with a space heater and a bucket of ice. A grinder company that started by selling to coffee nerds and ended up with a whole product line.

None of these businesses are "AI startups" in the traditional sense. And that's exactly the point.

As large language models get better at writing code, drafting contracts, and answering customer emails, the easy software wins are drying up. If an agent can spin up a SaaS clone in an afternoon, the moat was never the software — it was something else. This guide walks through the categories of businesses that hold up when AI can do most of the knowledge work, and why owning real assets, trust, and physical infrastructure might be the smartest move you make in the next decade.

Why does software stop being a moat in the agentic era?

For the last twenty years, the default startup playbook was: find a painful workflow, build software around it, charge a subscription. That playbook worked because writing good software was hard and slow. AI agents change the calculus. When a competitor can describe a feature in plain English and have a working prototype by lunch, the barrier to entry for pure software products collapses.

That doesn't mean software is dead. It means the kind of software company worth building has changed, and it means a whole category of non-software businesses just became more attractive by comparison, because they can't be cloned with a prompt.

What makes AI-native service firms different from regular agencies?

Traditional service businesses — law firms, accounting shops, marketing agencies — sell hours. AI-native service firms sell outcomes, and they use agents to compress the labor that used to require a dozen junior employees down to a handful of people managing AI systems.

The firm still needs humans for judgment calls, client relationships, and the parts of the work that require taste. But the ratio of revenue to headcount changes dramatically. This is the same thesis behind Y Combinator's push toward what partners call the "10-person, $100B company" — the idea that small, high-agency teams running AI aggressively can outcompete bloated organizations weighed down by meetings and middle management.

If you're building one of these, the edge isn't the AI itself (everyone has access to the same models). It's the workflow you wrap around it: the client relationships, the quality checks, the institutional knowledge about what "good" looks like in your niche.

Why do physical, offline businesses hold up so well?

Here's a question worth asking before you build anything: what's the food-truck version of your business? In other words, what's the smallest, cheapest, most physical way to test a demand before committing to a big buildout?

Physical businesses are interesting right now precisely because they're hard for AI to touch. An agent can write your marketing copy, but it can't pour concrete, cut hair, or hand someone a warm towel. The experience economy — gyms, wellness clubs, specialty retail, hospitality — is becoming more valuable as digital experiences become infinite and commoditized, not less.

Othership is a useful example here. It started as backyard sauna parties during lockdown and turned into a chain of bathhouses combining a 90-person sauna and eight cold plunges, blending performance, sober clubbing, treatments, and group therapy. The Flatiron location alone houses a 75-person cedar-lined sauna, and the Williamsburg location features a 100-person sauna with two sunken communal cold plunge pools. People pay for this not because the sauna is technically unique, but because of the shared, screen-free, in-person experience — something no chatbot can replicate.

Can robots really replace dangerous, dirty human jobs?

Some jobs are dull, dirty, or dangerous enough that automating them isn't just profitable, it's overdue. Window cleaning on skyscrapers is a textbook example — you're sending a person off the roof of a 45-story building on a rope, over and over, for their entire career.

Skyline Robotics built Ozmo to solve exactly this. The robot combines advanced robotics, artificial intelligence, and sophisticated sensors to clean windows three times faster than humans, and it's designed to automate all work at heights above 16 feet. It's already deployed commercially — Skyline announced a multi-year partnership with Platinum, Inc. to deliver Ozmo, a high-rise window-cleaning robot, to New York City skyscrapers, with Platinum owning contracts for 65% of Class A buildings in the city. The company is going after what it calls the $40 billion window-washing industry, and it's not alone — consumer versions like the Osmo Pro robot cleaner are bringing similar AI path-planning and edge-detection tech to home windows.

The lesson for founders: look for tasks that are physically repetitive, dangerous, or unpleasant enough that humans actively avoid them. If you can build (or distribute) hardware that handles it safely and cheaply, you're not competing with ChatGPT — you're competing with gravity and liability insurance.

How do you build a product brand that AI can't fake?

Branded physical products are another category that holds up well, especially when the brand starts narrow and expands. Specialty coffee gear is a good case study — companies that build cult followings among enthusiasts first, then broaden into a wider product suite once they've earned trust with the core audience. The AI can help you design the product faster and market it more efficiently, but it can't manufacture the taste, craftsmanship, and community loyalty that make people choose one kettle or grinder over a dozen cheaper alternatives.

Why are marketplaces and social networks still wide open?

Marketplaces and social networks are often described as more art than science, and that's not just a cliché — getting both sides of a marketplace to show up at the same time is genuinely one of the hardest problems in business. But that difficulty is also the moat. AI can help you build the matching algorithm or the recommendation engine, but it can't conjure liquidity out of nothing. Indie founders have shown repeatedly that scrappy, narrowly-targeted marketplaces can work even against giants, because large incumbents move slowly and niche communities reward founders who actually understand them.

If you're considering this path, the hard part isn't the tech — it's the chicken-and-egg problem of supply and demand. Solve that with a wedge (a specific city, a specific niche hobby, a specific professional vertical) before trying to go broad.

What does this mean for vertical AI agents specifically?

Not every opportunity in this list is anti-software. Y Combinator's own research pushes hard on "vertical AI agents" — AI that handles a complete, specialized workflow in one industry (tax prep, legal compliance, medical billing) rather than a general-purpose tool. The thinking here is that startups aim to actually solve a complete workflow in a target domain, rather than just provide a toolkit, because customers in 2025 demand tangible outcomes from AI. The defensibility doesn't come from the model — it comes from owning proprietary data and deep domain expertise that general-purpose AI companies won't bother building.

How do you decide which of these to actually build?

Start by asking three questions about any idea on this list:

1. Does it require something AI can't do alone? Physical presence, trust, taste, or regulatory relationships are all good answers.
2. Can you start small and cheap? The food-truck version of a sauna club is a pop-up event. The food-truck version of a marketplace is a spreadsheet and a group chat. Test before you build the full version.
3. Does the asset appreciate, or does it depreciate the moment AI improves? A real estate holding or a loyal customer base gets more valuable over time. A thin software wrapper around GPT-5 does not.

None of this means AI is irrelevant to these businesses — quite the opposite. The AI-native service firm, the vertical agent company, and even the sauna club all use AI heavily behind the scenes for scheduling, marketing, and operations. The difference is where the actual value sits: in the asset, the relationship, or the physical experience, not in the prompt.

Where should you start building?

Pick the category that overlaps with something you already understand — a hobby, an industry you've worked in, a frustration you've personally felt. Then build the smallest possible version of it this month, not next year. The founders who win in the next decade won't be the ones with the cleverest prompt. They'll be the ones who figured out which assets still matter when intelligence itself becomes cheap, and who moved fast enough to own them before everyone else noticed.