What an AI cofounder can't do

An AI cofounder owns a function end to end — it decides, acts and is accountable for a measurable outcome. That sentence is doing a lot of work, and most of what it excludes is more useful to know than what it covers. There are two kinds of limit below: things it should never do, and things it cannot do yet. The first list is the important one.

By design

Things it should never do

These are not missing features. They are decisions whose consequences land on a person, and handing them to a system that cannot carry the consequence is how you end up with a confident answer nobody owns.

Decide what you sell, or what it costs

Positioning and pricing are bets about who you want to be, and they are answered with taste and nerve rather than with last week's click-through. An agent optimising inside your positioning will make it sharper. An agent choosing your positioning is guessing with your company.

Pick which market to enter

It can tell you which of your current audiences responds best. It cannot tell you which market is worth ten years of your life, because that judgement depends on things it has no access to — what you want, what you are good at, what you would still care about when it stops being fun.

Negotiate, or sign anything

Partnerships, contracts and hiring all involve consequences that land on a person. It drafts outreach and tracks replies; the ones that carry commitment come to you first, on purpose.

Fix a product nobody wants

Distribution multiplies whatever the product already is. If the thing does not work, better ads make more people find that out faster. It will tell you when the numbers look like that rather than quietly spending through it.

For now

Things it can't do yet

Judge a creative before the data exists

It refuses to rank a creative until it has roughly a thousand impressions across at least two days. Before that the numbers are noise, and killing on noise throws away winners. This is a real constraint on speed, and it is deliberate.

Know something that isn't written down

It reads your site, your ad accounts and what customers say in public. A reason your last launch failed that lives only in your head is invisible to it until you tell it.

Guarantee a channel will work

It can find what is working in your niche and test into it faster than you would alone. Whether your particular offer wins on a particular platform is not knowable in advance by anyone.

Honesty

And one more thing

Cofounder is pre-launch. Everything above describes how the system is built to behave rather than an average of customer experience, because there are no customers yet to average. When there are, this page gets the parts we got wrong added to it.

Questions

Common questions

What can an AI cofounder not do?

Set positioning or pricing, choose which market to enter, negotiate or sign anything, or rescue a product people don't want. Those are judgement calls whose consequences land on a person, so a person makes them.

Will it publish something off-brand?

Brand rules — words it will never use, claims it cannot make — are checked on every asset before it reaches your queue, not after it is live. In approve-first mode, which is the default, nothing publishes without you regardless.

Can it spend more than I told it to?

No. The daily ceiling is an authority limit rather than a target: it is not a rule the system follows, it is a number it has no path to exceed.

Does it replace a marketing hire?

It replaces the volume work a hire spends most of the week on — briefs, variants, scheduling, reporting. It does not replace the judgement a good marketer brings, and a company with real budget should have both.

What happens when it is wrong?

Every action is reversible and logged with the reasoning it had at the time. That is the point of the log — not a feed of activity, a record you can audit after something goes badly.

Nothing to lose

Founding cohort seats include the first month and a written teardown of your current funnel.

Claim a seat

100 seats. No card. We tell you on day 14 whether it worked.