You don't need to speak
to users anymore.

Our first memo made a claim we still stand behind: thirty seconds of conversation with one user told us more than a month of dashboards. So this headline deserves an explanation rather than a shrug.

Here it is. Talking to users was never the point. The answer was. Conversation was simply the only instrument that produced it — and it is an expensive, slow, badly calibrated instrument that we all agreed to pretend was fine, because there was nothing else.

What the interview was actually for

Nobody emails forty customers because they enjoy it. You do it because your funnel says users die at step three and cannot say why. The interview is a workaround for an instrument that reports that and never why.

As a research method it has known faults, and we lived every one of them:

• Most people never reply. The ones who do are disproportionately the ones who stayed — you are surveying survivors about a problem that killed everyone else.
• People explain themselves badly. Not dishonestly. They reconstruct a reason after the fact, because the real one — a moment of confusion, a switched tab — left no memory worth keeping.
• You can only ask people who already exist. Which means you can only ask after you have built the thing, launched the thing, and paid for the leak in reputation and weeks.

That last one is the real cost, and it is structural. Every answer you get by asking arrives after the mistake has already been bought and paid for.

The thing that changed

Until recently there was no way to ask a question about your product except to ask a human who had used it. That is no longer true.

A model can read your repository, walk your actual flows, hold the behaviour of a specific audience in mind, and tell you where that audience will stall — in the build, before anyone real has seen it. The interview stops being something you queue behind a launch and becomes something you run at the brief, then again at every pull request.

This is what Growthmind is being built to do: simulate your audience against the code you are about to ship, and hand back the finding while it is still cheap to act on. Not a panel of imaginary customers with opinions — a model of one specific product and one specific audience, checked against what really happened, and corrected every time it was wrong.

The part that would be dishonest to leave out

You still need users. You need them more than ever, because they are the only scorekeeper that counts.

A simulation that is never marked against reality is astrology with better grammar. The loop only earns its keep because the second half exists: after launch, the prediction is compared to real behaviour, the misses are named, and the model is rewritten by what it got wrong. A simulated finding that real data contradicts is a kill, not a debate.

And there is one question no simulation should ever be allowed to close: what a user wants that you have not built. Surprise is the one thing a model of your product cannot generate, because it is a model of your product. Keep talking to users about the future. Stop interrogating them about a past that your own code already records.

The one-sentence version

You don't need to speak to users to find out where your product loses them — that answer is available before the first user arrives. You still need them to find out what to build next, and to tell you, in behaviour rather than email, whether you were right.

One of those is research. The other was always just waiting.

Growthmind is open source at github.com/growthmind-ai/growthmind — the reasoning ships with the verdict, so you can argue with it. If you disagree with this memo, that is the more interesting conversation: tell the founders why.