We almost shipped a dashboard. Three months in, we threw it out.

We spent three months building a social media tool and then deleted most of the interface.

The product was going to be what everyone builds: a calendar, a composer, a queue, an AI button in the corner. We got it working. Then we watched a beta user open it, and he looked at the UI for about three seconds and went straight to the chat, skipping the onboarding and the walkthrough we had argued about for two weeks.

That was the whole lesson, and it was an expensive one.

What we had built assumed the user would sit down and create. Every social tool does. They hand you an empty calendar and excellent tools for filling it. But the reason we were not posting for our own company was never that scheduling was hard. It was that every Monday started from a blank page. We had automated the cheap half.

So we rebuilt it the other way round. It reads your website, learns your voice, researches your competitors, and turns up on Monday with the week already written and designed, each draft carrying the reason it exists: a trend in your category, an event near you, you have not posted in four days. You approve, edit, or skip.

Three things I would defend:

Nothing publishes without a human yes. Approval is on by default, and it is per platform, so the LinkedIn version can ship while the Instagram one goes back. Autonomy you cannot stop is not a feature, it is a liability, and it is the fastest way to lose trust in a product that speaks in someone else's name.

It does not do analytics. No likes, reach or impressions read back from any platform. We get asked for it constantly. It is a production and publishing system, and pretending otherwise would put us in a category we would lose.

The chat is not a bubble stapled to a dashboard. That is the part that took three months to learn. Most AI products right now are not AI-native, they are AI-flavored.

Where we are: open beta, deliberately small, three founders, no outside money yet. I am not going to quote numbers I cannot show you.

The question I would actually like an answer to, for anyone else building agent-shaped products: how do you decide what the agent does without asking? Every step you make autonomous saves the user time and costs you a little trust. We have drawn the line at "does all the work, then stops." I am not certain that is the right place for it.