Every dot on this screen is one real episode from our interviews: a task, a decision, or a line a founder drew. The thesis is submitted. This page is what came out of it, written for the people it came from.
The dots move into the parts of the business they belong to. Two zones are full: building & operating and reaching the market. The zones about testing demand and about customer relationships are much emptier. The red dots sit where founders keep AI out on purpose.
In short: AI does the work; only the market can say if it was the right work. The risk: a website, a deck and a prototype can look like progress before anyone has paid anything.
Now the 94 episodes where AI was involved regroup by what AI did. Two things dominate: it made something (42 episodes: code, decks, pages, drafts) or it did the job of a person the team did not have (34: the developer, the designer, the analyst). Saving time was the main point in only 12.
Two columns are almost empty. Saving money is the one you would expect: it is a side effect of everything else. The other one matters more. In not one episode was reducing uncertainty about the business AI’s main job (it played a small part in two). AI does the work. It almost never tells you if the idea is right.
32 episodes sit to the right of the line. In each of them, a founder decided to keep a piece of work human. The reasons repeat from founder to founder: core IP, sensitive data, security, legal responsibility — and, most of all, customer relationships (7 of 8 episodes kept human, while AI was involved in 78% of the episodes about reaching the market).
A few of these “no”s are about waiting for better tools. Most are not. They are about what a mistake would cost, and who would be responsible for it.
findingIn exactly one episode, a founder used AI to work out what an outside change — a regulator’s draft scientific opinion — would mean for the business. Almost everything else is about the next step: build this, write that, reach this customer. Founders’ own bets — a segment, a channel, a price — are a separate question. There the picture was similar: founders asked AI to prepare the step far more often than to think about where the step leads.
my readThinking through scenarios used to be slow work with many documents. Big companies paid whole teams for it. AI makes it cheap. In these interviews, almost nobody described it as a habit.
try thisOnce a month, give your AI the three changes you noticed around you: a new regulation, a competitor’s move, a platform’s new pricing. Ask what each one means for your business, and what you would do differently. Then show the answer to someone who knows the field. The decision is still yours. But this is the cheapest experiment in this report.
The algorithm or the science at the heart of the product does not go into general-purpose models.
User data stays inside the company (GDPR came up more than once). Nobody gives AI a security-critical system.
AI can draft. A human signs.
If a number has to pass an audit, a black box is not allowed.
The clearest line in the whole sample.
The same founders give AI a lot of work where checking is cheap and a mistake costs little. This is not caution in general. It is choosing where to be careful.
| Level | Rule | Typical tasks |
|---|---|---|
| Delegate | AI does most of the work; the founder takes a quick look. | Drafts, research summaries, standard code, routine content |
| Draft | AI writes the first version; the founder decides what stays. | Decks, emails, financial sketches, proposals |
| Verify | Used only after it has been checked. | Analysis, citations, anything with numbers |
| Keep human-controlled | Kept human as a rule. | Key relationships, core IP, legal responsibility, security |
The four levels are my reading of the interviews, not a scale the founders used themselves. They are also personal: the same task can sit at a different level in a different company.
How founders checked AI’s work. Three kinds of check — and they are not equally strong:
Cheap, and they catch small mistakes. But a model checking itself is not an independent check.
Someone or something outside the model says yes or no.
The only check that answers “was this the right step?” — and the rarest one in the episodes founders told us about.
In this sample, heavy AI use usually came with more systematic checking, not less.
Producing is cheap now.
Knowing what to trust is the hard part.
Three founders from the interviews, each in a short snapshot: what they give to AI, how they check the result, and where they stop. Details are made general; the words are theirs.
Seven kinds of work that came up again and again in the interviews. For each one, say how you really handle it. After each answer you will see how the founders in the interviews handled the same kind of work, and one question to ask yourself. You can change an answer at any time and see what changes. There is no score at the end. The point is to see your own lines written down — and where they differ from the sample. Your answers stay in your browser; nothing is stored or sent.
Seven questions a founder might ask about this study, answered from the 130 episodes. There is no AI model behind this — just the data and my answers. Pick one.
You built a landing page. Which assumption does it test? What result would make you change direction? In the episodes founders told us about, a website or a deck was rarely tied to a test stated that clearly. That is how speed starts to feel like progress. A website, a deck or a prototype counts when it gets a real response: a reply, a payment, someone actually using it.
Put the four levels — my frame, taken from the interviews — on your own task list: delegate, draft, verify, keep human-controlled. Move tasks between the levels when the stakes change. The heavy AI users in this study often knew exactly where their lines were.
The more AI produces, the more of your work becomes checking and judging. Plan time for it. Otherwise speed just turns into unchecked activity.
Twenty-one founders in Austria gave me their time for this research in spring 2026. This page is the version I promised to send back to them — on purpose not the academic one. The thesis is over two hundred pages; this is the part that is worth twenty minutes of a founder’s day.
If you are a founder and this does not match your experience — or matches it exactly — I would like to hear it. One line is enough.
| design | Qualitative interview study (Master’s thesis, WU Vienna) |
| sample | 21 founders, early-stage (idea → seed, a few early post-seed), Austria |
| sectors | software, hardware, biotech, food, fintech, clean tech, health, media, other |
| period | March – May 2026 |
| unit | ~130 coded episodes: AI used / used with limits / kept human on purpose |
| anonymisation | names and company names removed; details that could identify someone removed or made general |
| limits | qualitative sample; self-reported; a snapshot from 2026 |
| analytics | anonymous page views and reading-depth counts (GoatCounter, no cookies, no personal data); your self-test answers are never sent |