Ideas earn a name here by showing up more than once — in the lab, inside the enterprise, or both. They're working theory: useful until proven wrong. When one breaks, it gets said here.
The AI implementation gap is an operations problem, not a technology problem.
The distance between what AI can do and what organizations actually put to work is enormous. Better models don't close it. Closing it takes someone who sits between the strategy and the system, someone who knows the workflows, the data, the people, and enough of the technology to make it real. That's the operator. Everything else on this page is downstream of this idea.
Purpose-built micro-apps with constrained inputs and outputs beat general LLM access as an enterprise pattern.
Handing everyone a chat window looks like democratization, but it ships the hard problem (knowing what to ask, and whether the answer is right) to the person least equipped to solve it. A micro-app makes those decisions up front: what goes in, what comes out, what good looks like. The constraint is the feature. Fewer ways to be wrong, and the ways that remain are visible.
Walled garden buys speed and compliance with a lower ceiling. Custom raises the ceiling — and the floor.
Walled-garden tools — the Gems, Copilots, and Projects of the world — deploy in days inside what the organization already permits. That's their entire value, and it's real. But their ceiling is fixed by the vendor. Custom builds go higher and integrate deeper, and they demand more to stand up and keep standing. The frame that falls out: walled garden for internal productivity, custom for anything deeply integrated with proprietary data or workflows. I keep watching organizations get this backwards in both directions at once.
Humans read duration as legitimacy. When the agent finishes instantly, "done" looks like "broken."
Meetings, spreadsheets, and deck revisions are how humans prove effort happened. Agents measure work in tokens, not hours — and near-instant completion breaks the oldest quality heuristic we have. The adoption problem isn't convincing users the AI is right. It's convincing them it actually worked.
READ THE POST →Dashboards stop being the destination. They become the checkpoint between what the AI found and what the workflows do about it.
For decades, reporting was the integration layer of the enterprise — the way state moved between teams that couldn't see each other's systems. When agents do the fetching and the acting, the dashboard's job inverts: it's where a human inspects agentic data work before it feeds a workflow, and where you verify the work happened right after.
READ THE POST →AI that generates confidence without accuracy is a core organizational risk — and fluent prose is what lets it travel.
Between the pre-AI enterprise and validated AI systems sits the middle: general-purpose AI pointed at real questions with no validation layer between the answer and the decision. It's where most organizations are, because it's the easiest place to reach. The danger is false authority: fluent language makes a wrong answer look like an answer.
READ THE POST →IDEAS GRADUATE FROM THIS LIST WHEN THEY SURVIVE CONTACT WITH REAL WORK. SOME WON'T.