Five businesses I'd look at building because of where AI agents are going. Each one gets more valuable as the models get smarter, not less. For every idea: who pays, what I'd build first, how I'd get customers, and what could kill it.
Everybody's trying to build the next great AI app. I'd rather build the things all those AI apps are going to need: memory, security, clean data, risk checks and data supply. Start with the test.
A new model drops next month. Does your business get better or worse?
Five ideas is the easy part. Picking one, testing it before you build, and getting the first person to pay is where most founders stall. That's the part I want to help you with.
I'm building a place where you take an idea like these from a list to a business on your own, with a CTO who's built two seven-figure software companies checking your thinking every week.
Test It Before You Build It.
Free to join the waitlist. Opens at $49/mo, and the price goes up as seats fill, so first in pays the least.
Greg Isenberg spent 48 hours in San Francisco with AI founders, data labs and execs, and posted what he heard. At one dinner, someone asked the table: "If a new model drops next month, does your company get better or worse?" Half the table went quiet.
A lot of AI startups are selling a model's current weakness. The product exists because today's model can't quite do the job alone, so the startup fills the gap with prompts, workflow and a nicer screen. Every model release closes part of that gap. Those companies are betting against AI getting better.
The other kind of business has customers whose needs grow as AI gets more capable. More agents doing more real work means more need for memory, security, clean data, risk checks and specialized data. Smarter models make those customers need you more.
None of these five is immune. Memory could become a built-in standard. Security testing could become a platform feature. The claim isn't that they can't be replaced. It's that their customers' need grows as the models improve, and that's where I'd want to be building.
Ordered the way I'd talk about them, not by how fast they make money. If you want revenue this quarter, start with 03. If you want the biggest swing, look at 01.
No prices or market sizes here on purpose. These markets are moving monthly. Talk to ten buyers before you trust any number, including one I'd give you.
Idea 01 · In The Video
Notion, But For Agents
A memory layer that you own and every AI can use. Your preferences, projects, people and voice live in one place, and you take them with you when a better model comes out.
Someone running a big consumer company told Greg the next fight is over memory: the AI that remembers the most about you wins. That's great for the provider. For you it's lock-in. The more your AI knows about you, the more painful it is to switch.
At the same time, the best teams treat models as swappable. They route one job to one model and another job to a cheaper one. Those two trends collide. People want to switch models, and memory is what keeps them from doing it. Someone should own the memory layer that sits outside any one provider.
Developers first. Open-source the MCP server, charge for hosted sync, and share it wherever people trade Claude Code and agent setups. Anyone hand-maintaining CLAUDE.md files and custom instructions across three tools is already doing this job manually.
Smarter agents do more of your work across more tools, so there's more worth remembering and more agents that need the same context. Better models also get better at deciding what to save and what to recall, which improves your product for free.
The providers want memory to be their moat, so they may make it hard to get out. Or the opposite happens: an open memory standard arrives, everyone supports it, and you're a feature. And you'd be holding the most sensitive data a person has. Encrypt it, offer self-hosting, and be boring about security from day one.
Idea 02 · In The Video
Every Agent Is A New Door
Test AI agents, and the software they connect to, for the ways they can be tricked or overreach. Start as a service. Turn what repeats into a scanner.
We're giving agents our email, our databases, our accounting and our customer records. Every company Greg talked to wants to pipe its internal data into one place and expose it to agents through MCP. As he put it, every agent with access to your systems is a new door into your company, and a lot of those doors were left open.
As screens disappear and agents become the main user of software, the attack surface moves from login forms to prompts, tool permissions and any text an agent reads. OWASP already lists prompt injection and excessive agency in its Top 10 for LLM applications. Most teams shipping agents have nobody who has tested for either.
Let enterprise security questionnaires do the selling. B2B companies adding AI features are starting to get asked how they prevent prompt injection, and most don't have an answer. Publish a free checklist, offer the audit at a fixed price, and partner with SOC 2 firms whose clients are asking the same question.
More capable agents get more permissions and more autonomy, which raises the cost of every mistake. Smarter models also make better attackers, and better testers. The thing you protect keeps getting bigger.
Security is crowded and runs on trust, and the big platforms will ship guardrails as features. Your edge is being independent and specific. Services don't scale like software, so plan the move from audits to a scanner early. And never sell "safe." Sell "tested," with the evidence.
Idea 03 · Bonus
The Fastest To Get Paid
Clean up and organize a company's data so its agents can actually reason over it. A company second brain is worthless if it's full of contradictions.
Here's the pattern Greg kept seeing. A company decides to pipe all its data into one place for agents. Then it looks at the data: five versions of the same customer, old Slack channels, PDFs nobody can search. The agent project stalls on the mess.
Agents are only as good as what they can read. If the data is organized so an AI can reason over it, every AI feature the company builds on top gets better, and keeps getting better with each model. That's compounding value, and most companies have no one to do the work.
Sell the outcome, not the cleanup: "your support agent gives the right answer." Run one department as a paid pilot with a before and after accuracy test. Partner with agencies building agents, who keep hitting bad data and don't want to fix it themselves.
Every new model reasons better over clean data, so the payoff from your work grows with each release. Smarter models also do more of the cleanup for you, which lowers your cost per job while your price holds.
It's consulting, with consulting margins. Smarter models may cope with messy data well enough that "clean" matters less. Data warehouses and enterprise search tools will bundle pieces of this. Your moat is the playbook and tooling you build from doing it fifty times.
Idea 04 · Bonus
Picks And Shovels For Insurers
Don't start an insurance company. Build the tests insurers will use to decide which agents they'll cover, and at what price.
Greg thinks agent insurance is coming: once an agent makes an expensive mistake, insurers will want audits before they cover anyone. Insurance companies are huge businesses, and nobody's starting one from their office. But insurers will need a way to measure agent risk, and right now there's no standard way to do it.
Whoever writes the evaluation everyone uses gets a seat in every policy. Insurers will build their own processes too. If you can define what "safe enough to cover" means for an agent, with tests that produce evidence, you're the expert they lean on.
Find one specialist broker or underwriter who wants to write AI policies and has no way to price them. Run the evaluations for their first clients at cost, in exchange for the data. Better yet, start with idea 02. Security audits are how you earn the expertise and the failure data, and this is where that business grows up.
As agents get more autonomy, their mistakes cost more and the demand for cover grows. Every new capability is a new kind of risk that needs a new test. Could models get so good you never worry about safety? I don't think that ever happens.
Timing. This market may be years away, and insurers move slowly. Big carriers may build evaluations in-house, or a standards body may publish a free framework. It's the most speculative of the five, which is why I'd get there through security audits, not start here.
Idea 05 · Bonus
Some Companies Are Data Companies
A marketplace where companies list licensed, permissioned industry data, and AI labs and vertical AI startups buy it.
Greg heard story after story of data brokers making around $2M a week selling company data to labs. He couldn't tell if it's the beginning or a bubble. He also heard that a shut-down startup with five years of Slack, email and code can get paid to license all of it, and that some SaaS companies are secretly data companies.
If software can be rebuilt quickly, the value moves to data. Vertical AI is beating generic AI, and vertical AI needs data about one industry that nobody can scrape. Right now those deals happen by hand, through people who know people. That's what marketplaces replace.
Supply first. Talk to founders winding down, investors holding dying SaaS companies, and trade associations. Build a catalog in one industry, then take it to the labs and vertical AI startups that need it.
Every lab is training on the same public internet. As models get smarter, the edge moves to specialized data they can't scrape. Smarter models also make it cheaper to clean, label and scrub what you sell.
Legal risk. Customer data often isn't yours to sell, and contracts, privacy law and terms of service get in the way. Labs could lean on synthetic data or deal directly with big data holders. It might be a bubble. And marketplaces are the hardest thing to start, with two cold sides. Broker before you build.
Four questions I'd ask about any AI startup, including the one you're already building.
If the honest answer is "less," you have a clock on your business. Know how long it is.
If your product exists because models can't quite do X yet, every release is a threat. If it exists because agents are doing more, every release is a tailwind.
Data, relationships, trust, distribution, a map of messy systems. Code isn't on the list anymore. People are cloning whole apps with agents.
In a good business, smarter models make your work cheaper to deliver. In a bad one, they make your product cheaper to replace.
Anybody can build software now. Not everybody can build a business.
Reading about tools is a great way to feel busy. The Lab is for Founders who'd rather get customers: weekly office hours with a CTO who's built two seven-figure software companies, and a room of other Founders who care whether anybody buys what you're making.
Less Theater. More Customers.
Want my 30-Day Launch Plan? It's for Lab members only: validate, plan, build and launch, one week at a time.
See The Solo Founder Lab ▶