A new paper rebuilds the DIKW pyramid for medicine and shows where AI actually breaks. Strong on the arrows nobody worried about, weak on the one the structure depends on. The pyramid always assumed you could tell which layer you were standing on. That assumption is gone.
A Cybercab was filmed with no steering wheel. Removing it was the honest part. The rungs below it, hands on the wheel, eyes on the road, ready to take over, were never supervision. They were accountability. Most enterprise AI is still on rung three.
Engineering changed its unit of work to intent, policy, and boundaries. Product and design mostly still ship user stories and screens. So the policy gets written by the last person to touch the ticket, in a prompt, at midnight. The control plane has inputs, and nobody upstream owns them.
Anthropic mapped product work into five archetypes: Prototyper, Builder, Sweeper, Grower, Maintainer. All five stop the moment the thing ships. But an agent keeps deciding after you let go of it, and there is a sixth archetype no one drew: the one who watches it decide.
Six people sent me their skills this week. The feed is full of them. But everyone is treating a skill like a macro, a way to go faster. It isn't. A skill is a template, and templates were never about speed. They were about everyone doing the work the same way. That changes who should build them.
Two things are both called "building with AI." One makes small teams faster at familiar products. The other ships products that act on their own, and that one needs a bigger, more senior supervisory layer, not a leaner pod. The ICU knew this before software did.
Every PM team I know is busy automating the wrong things. Skills, MCP servers, vibe coding prototypes. The demos are impressive. But if AI is handling your roadmap research, your status updates, and your PRD drafts, what are you doing with the time you got back?
You cannot build judgment about a probabilistic system by reading about it. You build it by living with one, daily, until its behavior stops surprising you. Then you fix the habit that makes it dangerous: it agrees with you too fast. The browser agent and the roadmap agent are the same system.
The PRD was built for software that does what it's told. An agent exercises judgment, and the moment it crosses from suggestion to decision, the old brief breaks. It splits in two: one document to decide, one to build. The new AI tools can write the second. The first is still yours.
I had undiagnosed ADHD for most of my life. Lecture halls were the wrong delivery mechanism for me. So I went to the library, ran my own research, built my own curriculum. Six months ago, the constraint of scale disappeared. The library is still the right model. The library is just bigger now.
Your dashboards are green. Your agent approved 17 wrong purchase orders overnight. Traditional O&M answers "is it running?" Agentic O&M must answer "is it behaving correctly?" These are different questions. They require different instruments.
Your agent passed every security check. The tools your team used were built for a different system. The frameworks that cover agentic AI are months old, the enterprise adoption cycle is 12 to 18 months long, and the models getting better at finding your gaps ship faster than your procurement cycle.