How to start a company with AI

This is the method The Unnamed Roads uses to start and test every company in its portfolio. It is written so that you can use it too. The short version:

Pick one real problem for one specific group of people.
Build the smallest possible test with AI, in days rather than months.
Measure real demand against a goal you set before you start.
Keep, change or park the idea, and write down what you learned.

AI makes every step faster and cheaper. That changes the game: instead of betting a year on one idea, you can test many ideas and only keep growing the ones people actually want.

Why AI changes how companies get started

Starting a company used to need a team before you could even test the idea: someone to research the market, someone to design, someone to code, someone to write and someone to sell. That made each test expensive, so founders tested few ideas and held on to them for too long.

Today AI tools and agents can do most of that hands-on work:

Research: summarise markets, competitors, forums, reviews and search data.
Writing: draft the offer, website copy, emails and documentation.
Design and code: build websites, prototypes and working tools from a description.
Analysis: read the numbers from a test and point out what matters.
Operations: run routine jobs such as reports, data collection and follow-ups.

What AI does not do well on its own is decide which problem is worth solving, earn the trust of customers, or carry responsibility. Those stay with a human. In the studio, nothing is published, deployed or sent to someone outside without a person approving it.

Step 1: Pick one real problem

Good ideas start with a problem, not a product. Look for a problem that is:

Specific: one clear group of people, one clear situation. “Swedish parents planning how to split parental leave” is better than “families”.
Painful or frequent: people already spend time or money working around it.
Reachable: you know where these people are, online or offline.

How AI helps: ask an AI assistant to collect complaints and questions from forums, reviews and search suggestions, list existing solutions and their weaknesses, and estimate how many people have the problem. Treat the output as a starting point and check the sources.

Example from the studio: Parental Leave Planner started from the observation that Swedish parental-leave rules are hard to turn into a concrete plan.

Step 2: Build the smallest possible test

Before building a product, decide what you need to learn. Write down the riskiest assumption, for example “clubs will pay for weekly hockey insights” or “dealers want one portal for all brands”. Then build the smallest thing that can test exactly that:

a landing page that explains the offer and asks for an email or a pre-order,
a clickable prototype or a simple working tool,
or a manual service where you do the work by hand (with AI) before automating it.

How AI helps: AI coding assistants can build a complete website or prototype from a written description in hours. AI can also draft the copy, suggest names and create simple visuals. The studio builds most tests this way, with one person reviewing the result.

Rule of thumb: if the first test takes more than a week to build, it is too big.

Step 3: Measure real demand

A test is only useful if you decide in advance what counts as success. Set three things before the test goes live:

The signal: what people must do. Sign up, reply, book a call, pre-order or pay.
The goal: how many, for example 30 sign-ups or 3 paying customers.
The deadline: when you stop and judge, for example after four weeks.

Then get the test in front of the right people: search, communities, direct messages, partners or a small ad budget. Track visits and actions with a simple analytics tool.

How AI helps: AI can write variations of your message, help you find where your audience is, and summarise the results when the test ends. It is also good at pointing out when the numbers are too small to mean anything.

Step 4: Keep, change or park

When the deadline arrives, compare the result with the goal you set:

Clearly above the goal: keep going. Build the next, slightly bigger version.
Close, or interesting signals: change one thing (the audience, the offer or the price) and run one more test.
Clearly below the goal: park it. Write down what you learned and move on.

Parking is not failure. A parked idea that cost a week is a cheap lesson, and the parts you built (code, copy, research) are often reused in the next idea. In the studio, parked projects stay visible on the projects page with a short explanation.

How the studio runs many companies at once

Running several companies as one person only works with a few simple rules:

Only a few projects get full attention at a time. Everything else is either being tested quietly or parked.
Every project has an honest status: Live, Testing or Parked.
AI does the work, a human approves. Drafts, code and analysis are produced by AI; anything that goes public or costs money needs a human yes.
Shared tools. The same AI setup, hosting, analytics and templates are reused across projects, so each new test starts faster than the last.
Written lessons. What each test taught is written down, so the next idea does not start from zero. Some of these lessons are published as Field Notes.

Tools you can use

You do not need a special setup to start. A typical stack for testing an idea with AI:

an AI assistant for research and writing,
an AI coding tool to build websites and prototypes,
simple hosting for the website,
a lightweight analytics tool to measure visits and sign-ups,
an email or form tool to collect interest.

The studio’s own setup is listed on the stack and tools pages.

Common mistakes

Building too much before testing. AI makes building easy, which makes over-building tempting. Test the riskiest assumption first.
No goal set in advance. Without a goal, every result looks “promising” and nothing gets stopped.
Trusting AI research blindly. AI can be confidently wrong. Check important facts and numbers against real sources.
Too many projects at once. Testing many ideas is good; actively growing many at once is not.
Skipping real conversations. AI can draft the message, but you still need to talk to the people you want to help.

Common questions

Can you really start a company with AI?

Yes. AI can now do most of the hands-on work of an early-stage company: market research, writing, design, code, websites, data analysis and routine operations. What AI does not replace is choosing a real problem, talking to customers and taking responsibility for what goes out under your name.

How do I validate a startup idea with AI?

Write down the riskiest assumption behind the idea, decide what result would prove it (sign-ups, replies, pre-orders or payments), let AI build the smallest test that can produce that result, run it for a fixed time, and compare the outcome with what you decided in advance.

How long does it take to test an idea with AI?

A first test (a landing page, a prototype or a simple tool) can usually be built in a few days. The longer part is getting enough real people to see it. Plan for a few weeks per test and set the end date before you start.

How much does it cost to start a company with AI?

The tools are cheap: AI subscriptions, a domain and hosting often cost less than a few hundred euros a month in total. The real cost is your time and attention, which is why it matters to stop tests that are not working.

What is the difference between an AI venture studio and an accelerator?

An accelerator invests in outside founders and supports them through a programme. An AI venture studio starts its own companies and uses AI to do the work a team would normally do, so it can test many ideas in parallel.

Start your own test

Pick one problem this week, write down the riskiest assumption and the goal, and let AI build the smallest test that could prove you wrong. Then follow what happens with the studio’s own experiments on the projects page and in the Field Notes.

Questions or ideas? Get in touch.