What using Claude as a co-founder actually looks like

I gave a talk in Belgrade called Claude as Co-Founder to a room of about a hundred people, mostly founders and people working inside programs. The framing got attention, which is fine, but the framing is also the part most likely to mislead. A co-founder carries risk, holds a share of the outcome, and disagrees with you at cost to themselves. A model does none of that. What it does is remove the excuse that you were waiting for someone else to do the work.

So here is the version I would give to one person over coffee instead of to a room.

There are four jobs where an AI partner genuinely changed my week. First, thinking out loud with structure. I describe a program design or a product idea in messy Serbian-inflected English and get back a version with the assumptions separated from the decisions. That is not insight, it is ordering, and ordering is most of what an early idea is missing. Second, first drafts of anything: an application form, a partner email, a landing page, a cohort curriculum, a data schema. A bad first draft in two minutes beats a good first draft in two days, because the argument starts earlier. Third, adversarial review. Asking it to attack my own plan as an investor, as a skeptical corporate partner, as the user who has no reason to care. Fourth, translation between worlds. Turning what a founder said into what a corporate partner needs to read, or turning a technical constraint into a sentence a non-technical stakeholder can act on.

Now the parts it cannot do, and this is where people get hurt. It cannot tell you whether anyone wants the thing. It has no access to the twelve conversations you have not had yet, and it will happily produce a confident market rationale for a product with no buyer. It cannot carry a relationship. Every real partnership I have closed came from a person who trusted a person, and there is no prompt for the second year of that relationship. And it cannot decide. Decisions are where you spend reputation, and a model has none to spend.

The failure mode nobody warns you about is fluency. The output is always articulate, and articulate reads as correct. I have caught myself accepting a paragraph because it was well built, not because it was true. My defence is boring and it works: for anything that will be read by someone who can say no to me, I make it state its sources or its assumptions explicitly, and I check the two numbers that matter. Almost every serious error I have made with AI in the last year was a number I did not check.

The second failure mode is that it makes you fast enough to skip the part where you were forced to think. When building took three weeks, you thought hard about what to build. When it takes two days, you can build four wrong things in a month and call it momentum. I now write the first paragraph before anything else: who the first user is, what they do today instead, what would make them stop. If I cannot write it, the model does not get the prompt.

Practically, my working setup is unglamorous. Long-running context per project rather than one-off questions, because most of the value comes from it knowing the constraints I already explained. Everything factual gets verified. Everything user-facing gets rewritten in my own voice, because the default register is smooth and forgettable and my whole professional identity depends on not sounding like that. And anything touching real people's data stays inside a boundary I can explain to the person whose data it is.

The part I did not expect: it changed what I am willing to start. Ideas that used to die as a note now get four hours and a working screen, and most of them still die, but they die with evidence instead of with a shrug. That is the real shift. Not a co-founder. A permanent removal of the excuse.

If you are trying to work out where this fits in your own week, the useful question is not which model. It is which of your bottlenecks is actually a writing, structuring or drafting bottleneck, and which one is a trust or distribution bottleneck. The first kind moved a lot in the last two years. The second kind has not moved at all.