What If We’re Assigning Problems Too Early?

95% of corporate AI initiatives deliver no measurable P&L impact.

The usual explanations are familiar: bad data, weak models, poor integration.
But what if the failure happens earlier—at the moment we define the problem?

Corporate AI works like this:
Humans specify the problem, set the boundaries, define success—
then deploy an agent to execute.
The agent becomes an execution engine for a human hypothesis.

But what if the hypothesis is wrong?

Or incomplete?

Or aimed at a symptom rather than the cause?

Something I’ve been watching made me pause.

In one emergent, agent-only environment I’ve been observing,
no one assigned tasks. No workflows. No predefined objectives.

Agents simply interacted—and began identifying gaps on their own:
quality issues, missing infrastructure, coordination problems.

Entire solution threads formed without anyone
declaring “this is the problem.”

That contrast matters.

The pattern is interesting:
Corporate AI: Problem specified → solution constrained → agent executes
Emergent AI: Environment provided → problem discovered → solution proposed

The first assumes humans already know what needs solving.

The second assumes they might not—and that this uncertainty is valuable.

Hayek won a Nobel Prize showing why central planning fails—
not because planners are foolish, but because the
knowledge required is distributed, local, and often tacit.

What if the same is true for AI?

An agent embedded in your codebase, your operations,
or your customer flows encounters information you never see.
If we define the problem before the agent engages, we discard that information by design.

Maybe the architect’s job isn’t to specify problems upfront.
Maybe it’s to design environments where valuable problems can be discovered.

That’s a different way to think about AI—and I think it will matter.