A startup idea is different from a business idea: it has to grow, and it has to hold on to what it wins. Most "AI startup" lists fail both tests; they are features that a platform will absorb, or markets so broad that no small team can serve them. The ideas below are built the other way round: a narrow wedge in a specific industry, a customer who pays now, data or workflow that accumulates, and a clear answer to "why did this not exist two years ago". Then a method to evaluate your own idea the same way.
Free to start. For this kind of thinking, the latest models are worth it.
All twelve share a pattern: a workflow in one industry where people currently pay in staff time, where the data a customer gives you makes the product better for them, and where the big AI platforms will not build something this specific. Each starts with one customer, in person.
Ask these of any idea, including the twelve above. Put the idea and your answers in a Project and have several models answer independently; where they disagree with you or each other is where the risk lives.
What changed that makes this possible or necessary today: a model capability, a regulation, a cost that fell, a behaviour that shifted? If the answer is only "AI exists", the idea is late. Field reports: on-device voice and vision got good enough to work in a basement with no signal.
The single narrow job you win first, before anyone notices you. Not "AI for insurance"; "first-notice-of-loss intake for water damage claims at regional brokers". Wedges get you a paying customer in a month and a reputation in a year.
Name the buyer, the budget line, and what it costs them now in staff hours or software. A product that replaces $40,000 of admin time a year can charge $6,000 with a straight face.
Why is the product better for a customer in month twelve than in month one? Their data, their rules, their history, integrations they will not want to redo. Without this, a bigger company copies you the moment you prove the market.
Because it needs industry-specific data, workflow, compliance or distribution that a general assistant cannot get to. If a general chatbot with a file upload does 80% of it, you have a feature, not a company.
Ask each model to answer the five questions for your idea without seeing the others' answers. One will call the wedge too narrow, another too broad; one will find a platform risk the others missed. You are not looking for consensus. You are looking for the objection you had not considered, before an investor or a customer raises it.
Talk to twenty people with the problem. Ask what they do today, what it costs, what they have tried. Get one to agree to be a design partner: they give you access and feedback; you give them the product free for a year.
Build only the wedge with AI, with your design partner's real data in the Project. Measure one number: minutes saved per claim, quotes turned around per day, tickets resolved without a human. How to build it.
Use the measured result to sign three more, at a real price. Write one page: the customer, the number, the price, what accumulates. That page is your pitch whether you raise money or never do. Launching it.
Evaluating a wedge, reasoning about why a platform would or would not absorb a product, sizing a market bottom-up from real assumptions: these are the questions where older models give confident generalities and the newest ones give specific, checkable answers. If you are going to spend a year on an idea, spend a month of the strongest models on it first.
Newer models hold your whole Project in view instead of the last few messages, reason through trade-offs instead of picking the first plausible answer, and are wrong less often and with less confidence. For the questions on this page, that is the difference between advice that sounds right and advice you can act on.
Pro from $20 a month, cancel any time. The free plan stays free.
Add what you know about the industry, any documents you have, and the idea in one paragraph.
A business idea needs a customer who pays. A startup idea also needs a reason to grow and a reason to hold on to what it wins: a wedge, something that accumulates, and a defence against platforms. Plenty of good businesses are not startups, and that is fine. See AI business ideas for the other kind.
Less than before. AI can build the wedge product with you, and several models reviewing each other catch most of what a second engineer would. You do need someone who knows the industry deeply. If that is you, start.
Not before you have a design partner and a measured result. A page with a customer, a number and a price is worth more than a deck, and it changes what you can raise on. This is general information, not financial advice.
Search, then look at what the customer actually uses today: often a spreadsheet or a person. An existing competitor is not a reason to stop; a competitor your customers love is. Read their two-star reviews.
Because startup questions are judgement calls, and one model gives you one judgement with no error bars. Several independent answers show you where the idea is solid and where it rests on a guess.
Your first Project, file uploads and the major model families. The latest models, sourced web research, more Projects and the multi-model verification modes are on paid plans. See plans.
Wedge, why now, who pays, what accumulates, platform risk. Several models, independently, before you spend a year on it. Free to start.
Also: AI business ideas · validate a business idea with AI · launch a product with AI
Continue learning
Run the same prompt through ChatGPT, Claude, Gemini and Grok before trusting one answer.
Core featureLet several models draft, challenge and verify each other instead of trusting one answer.
ProjectsKeep files, instructions, code and chats together so every model works from the same context.
FeaturesModels, Projects, collaboration, verification, Studios and images in one workspace.
HubIs the idea good, who else does it, what must it do, what will it cost: answered, then built in a Project.
ValidationA five-rung ladder with pass marks, from scoring to pre-selling.
Market researchSourced market size, segments, pricing and competitors with Perplexity Sonar Pro.
PlanTen sections, one page each, bottom-up financials, kept alive in a Project.
Online businessA real problem, a small product or service, a demand test, honest costs.
IdeasSixteen narrow, startable ideas with who pays and what you build.
FoundersOne Project for research, code, decisions and launch work.
LaunchThe four weeks before, the day, the week after.
