Start the business you have been thinking about. Keep all of it in one place.

No idea yet? Start with AI business ideas or, if you want something that has to grow, AI startup ideas. Have one? "Good" is not a feeling. An idea is worth building when it passes five tests, and you can score it in an afternoon. Score each from 1 to 5. Under 17 out of 25, change the idea or the customer before you build anything.

1. Painful. Does the customer lose money, time or sleep over this every week, or is it a mild annoyance? Invoice tool: an agency owner spends Friday afternoons writing awkward reminder emails and gets paid 20 days late. Pain: 4.
2. Already paid for. Do they spend money or staff time on this today, in any form? People who pay nothing now rarely start. They pay a bookkeeper or do it themselves; some pay for accounting software that does it badly. Score: 4.
3. Reachable. Can you name three places where these people gather, or ten of them personally? Agency owners are in Slack groups, LinkedIn, and the local design association. Score: 4.
4. Buildable by you. Can a first version exist in a few weeks with AI, without a team? Reminder emails plus a status page: yes. A full accounting suite: no. Score: 5 for the small version.
5. Repeatable. Is it the same problem for many people, or a custom job every time? Every agency has unpaid invoices. Score: 5.

Scoring is rung one of a five-rung ladder; the rest, from ten interviews to a pre-sale, is on the validation page. The scoring is where several AIs earn their place. Ask each one to score the idea independently and explain the score. When one gives "Reachable" a 4 and another a 2, the disagreement is the real answer: you have not actually found where these people are yet. That is the thing to do next, before writing code.

The sourced version of this, with Perplexity Sonar Pro doing the sweep and a reasoning model checking it, is on market research with AI. Most people do this backwards: they search for their idea, find three companies doing it, and either give up or decide those companies are bad. Do it in this order instead, and put every result into the Project so the AIs can work from real material rather than from memory.

1. List what the customer does today instead

Your competitor is rarely another startup. It is a spreadsheet, a bookkeeper, an accounting add-on, or "we just wait". List all of them: direct competitors, indirect ones, and the do-nothing option. Ask an AI with web research to find ten, then check each one yourself.

2. Read their worst reviews, not their homepage

For each competitor, collect the two- and three-star reviews from app stores, G2, Capterra or Trustpilot. Those are people who paid and were disappointed; they tell you exactly what the product must do that the incumbent does not. Upload screenshots or paste the text into the Project and ask for the recurring complaints. For invoicing tools the pattern is usually "too complex for a small agency" and "reminders sound robotic". That is your positioning, handed to you.

3. Build the comparison table

One row per competitor: who they target, their price, their most-praised feature, their most-hated gap, and the one sentence they use to describe themselves. Have it produced as a spreadsheet you can keep updating. The empty cell in that table, the thing nobody does well for your customer, is where you go.

4. Size the market from the bottom up

Ignore "the invoicing market is worth $X billion". Count instead: how many small agencies exist in the countries you can sell to, what share have the problem badly, what you would charge, and what share you could realistically win in year one. 50,000 agencies × 30% with late-payment pain × $29 a month × 1% reached = about 150 customers and $4,300 a month. Small, honest, and enough to know whether it is worth your time. Have two AIs do the arithmetic separately; if their assumptions differ by a factor of ten, you have found the number you need to check with real data.

Not "more features than the competitor". First products win on a short list of things that most first products get wrong. Treat this as the bar your version one must clear before you spend a day on anything else.

Requirement	What it means in practice	Invoice tool example
One job, done completely	It solves one painful thing end to end, not five things halfway.	Reminder goes out, client pays, status updates. Nothing else in version one.
A first win in ten minutes	A new user gets something valuable before they have to learn anything.	Paste one overdue invoice, approve the reminder, it is sent. Done in the first session.
Obvious to buy	One price, one page, a clear "what happens when I pay". Confusion is the biggest reason people leave a checkout.	$29 a month, cancel any time, first two reminders free.
Nothing embarrassing can happen	The failure that would make a customer look bad to their customer cannot occur.	Every reminder needs approval. No email ever goes out on its own.
It does not lose their data	Backups, a way to export, and nothing that disappears on a bad day.	Invoices export to CSV at any time.
A reason to come back	Something that is only true next week: new data, a report, a reminder, a streak.	Friday summary: what got paid, what is still late.
Trust signals	Who you are, how to reach you, a privacy policy, what happens to their data.	A real name, a support address, a one-page privacy policy.
A way to be found	You know where the first fifty users come from before launch, not after.	Two agency Slack groups and a LinkedIn post a week.

Just as important is what does not matter for version one: a logo, a mobile app when a web app works, integrations, a dashboard, a team, and any feature that only one person asked for. Every hour on those is an hour not spent finding the tenth customer.

Have several AIs score your plan against this table before you build, and again before you launch. They will disagree most on "obvious to buy" and "a reason to come back", which are exactly the two most founders neglect.

Less than you fear, and it is mostly time. The largest cost of starting a software business used to be building the software. When AI writes the code and you own the result, the first three months look like this. Figures are typical ranges, vary by country and provider, and are meant for planning, not as quotes.

Item	Typical cost	Notes
Domain name	$10 to $20 a year	Buy it the day you name the business.
Hosting for a small web app	$0 to $20 a month	Free tiers cover the first hundreds of users.
Database and sign-in service	$0 to $25 a month	Free until you have real traffic.
AI workspace	$0 to $90 a month	Free to start; the paid tier matters once the Project holds a lot of code and files.
Payment processing	About 3% plus a small fee per transaction	No monthly cost. You pay only when you get paid.
Email sending	$0 to $20 a month	For onboarding and support mail.
App store accounts (mobile only)	Apple about $99 a year, Google a one-time fee	Skip entirely if version one is a web app.
Registering the business	$0 to a few hundred, or more for a limited company	Many people start as a sole trader and incorporate once there is revenue. Check your country's rules.
Small marketing tests	$100 to $300	Only after free channels show people stay.
Legal and accounting	$0 to $500	AI drafts your terms and privacy policy; a professional reviews them once.

Total for the first three months: typically a few hundred to about $1,500. The same first version from an agency is commonly quoted in the tens of thousands. The capital you are actually spending is time: plan for five to ten hours a week for eight to twelve weeks to reach a launched product with first users.

The number that matters more than the budget

Months to first revenue. Write down how many months you can sustain the time and the small costs before the business must pay for something. Then build the smallest thing that can charge money inside that window. Ask two AIs to build the three-month budget from your plan, including the costs you forgot, and to tell you the month the business breaks even at three different prices.

A business is more than its product. Around it sit a dozen assets that used to need a designer, a copywriter, a lawyer and a bookkeeper. With AI, each one is a conversation in the Project, and because the Project already holds your customer, offer and price, every asset comes out consistent with the others. The table shows what the AI produces, what only you can add, and which ones you must have before launch.

Asset	AI produces	You add	Before launch?
The product	The code, screen by screen, checked by a second AI	Decisions, testing, publishing	Yes
Name and one-line pitch	Twenty options with reasoning, domain availability to check	The choice, and a trademark search	Yes
Landing page	Copy and a working page: headline, benefits, screenshot slots, one button	Real screenshots, your name and face	Yes
Pricing page	Price presentation, what is included, the FAQ under it	The price itself	Yes
Privacy policy and terms	A first draft in plain language, specific to what your app collects	One review by a professional	Yes
Onboarding emails	Three short emails: first win, the feature people miss, a question	Cutting every sentence that sounds like marketing	Yes
Help page and support replies	The ten real questions with answers; drafted replies to each email	Sending them yourself, from your address	Soon after
Financial model	A spreadsheet: costs, prices, customers per month, break-even	Real numbers as they arrive	Yes
Brand basics	Colour and type suggestions, a logo brief, image directions	Taste, and a decision to keep it simple	No
One-pager or deck	A short document for partners, a first hire, or your own clarity	Honesty about what is not working yet	No
Sales conversation	The questions to ask, the objections and answers, the follow-up email	Having the conversation	Yes
Contract and invoice templates	Drafts specific to your service	Professional review, your bank details	If you sell a service

Two habits make this work. First, make the assets in the order the customer meets them: name, landing page, pricing, product, onboarding, support. Second, when an asset matters, have a second AI review the first one's draft with a single question: "What in this would make a small-agency owner not trust us?" The answer is usually one sentence you would never have noticed.