How Much Do We Need to Understand Before We Act?

There's a moment that comes up in almost every difficult problem. You've done the research, talked to people, gathered data, and mapped things out. You know more than you did when you started. And yet, somehow, your path ahead is still nebulous.

The more you learn, the more edges appear. The exceptions compound. What seemed like a clear problem may start looking like three different problems wearing the same coat.

So at some point, you have to decide: do we keep trying to understand, or is it time to do something?

We tend to think of this as a sequence:

Understand → Decide → Act

But complex problems rarely unfold so neatly.

Some things only become clear once we change something. An assumption that seemed reasonable doesn't hold up. A constraint we worried about turns out not to matter much. Something we didn't even know to ask about suddenly becomes important.

There is a long history of research around this, from learning by doing to how organizations learn. One recurring idea is that some knowledge doesn't come before action. It comes from it.

That suggests a slightly different question:

What don't we know that would actually matter for the next step?

Once we can name that, sometimes a small experiment can teach us more than another round of analysis. We try something, see what happens, and adjust what we thought we knew. We may not have solved the problem yet. But perhaps we've made the next decision a little less uncertain. That is a useful kind of progress too.

AI adds an interesting wrinkle to all of this.

It can now gather, connect, and synthesize information in hours that might once have taken us days. That's an extraordinary advantage. But having an explanation readily available isn't quite the same as having built the understanding ourselves.

Some of the effort AI removes was simply friction. But some of it—the searching, getting confused, following a trail, reconciling contradictions—is also how a mental model takes shape.

And as AI makes it easier to produce plausible answers, perhaps the value shifts a little. Coming up with possibilities becomes easier. Knowing which questions matter, which assumptions we should be careful about, and what still needs to be tested in the real world becomes more important.

So AI may help us get to better hypotheses much faster. But it doesn't necessarily remove the old loop.

We still try something, observe what happens, learn a little more, and try again.

And maybe that brings us back to a useful question:

What's the smallest responsible thing we can do that would teach us what we need to know next?

Further reading