AI may make software disposable.

It doesn’t make dependency disposable.

That’s one of my takeaways from Benedict Evans’ latest essay on AI, tools and transformation.

He describes enterprise work as moving between two states: improvised and institutionalized.

For years, the cost of building software created a natural control point.

You needed engineering time. Budget. Procurement. A project.

AI weakens that constraint.

A salesperson can create a workflow.

An analyst can build a small tool.

A team can assemble an agent around a problem that would previously have needed a backlog.

That sounds like a software-creation problem disappearing.

I think it actually moves the harder problem downstream.

When implementation becomes abundant, selection becomes scarce.

Which of those thousands of experiments should remain disposable?

Which should disappear?

And which have quietly become important enough that the business now depends on them?

That last transition changes everything.

Identity. Permissions. Auditability. Reliability. Security. Ownership. Support.

This also suggests that the answer cannot be to apply traditional application governance to every prompt, script or agent someone creates.

That would destroy much of the benefit.

The more interesting architecture problem is knowing when an improvised capability has crossed the threshold into a durable dependency — and then giving it the controls appropriate to that state.

Evans puts it well:

“AI doesn’t change the question: it creates new choices and moves the thresholds.”

Perhaps that is the enterprise AI shift worth watching.

AI makes creation cheap.

It doesn’t make responsibility cheap.