Writing
What makes AI products ship versus die.
Notes from the work: product judgment, production economics, and the gap between demo and production. Written by the two of us, one byline at a time.
Articles
-
Everything is buildable now. That's the problem.
Implementation cost used to be the idea filter - most ideas died of expense before anyone had to judge them. AI removed that filter, and most teams replaced it with nothing. Four tests that move the filter into the judgment, and how to say no without killing the idea that matters.
-
The work didn't disappear. It moved.
The one-click myth says AI deletes work. Watch a real workflow and the work moves instead - upstream into specification and context, downstream into verification and judgment. Why "AI didn't save me time" and "AI transformed my work" are both true, and what that means for anyone designing an AI product around a magic button.
-
Surviving the organisation.
A signature is not adoption. Enterprise software is adopted because it survived everyone who didn't - and AI has to keep clearing the organisation, not clear it once.
-
You can't procure the winner. So build what outlasts it.
SpaceX is renting Anthropic compute while buying Cursor to compete with it - the AI coding tools churn faster than any procurement cycle, so rent the tool and own the tool-agnostic context layer that survives the switch.
-
Europe's AI sovereignty has three layers. The legislation covers two.
The June 2026 sovereignty package funds data and compute and leaves the LLM layer - where capability lives - without a pillar. Three slices, why the missing one matters most, and what it means for anyone scoping an AI product for European buyers.
-
AI isn't too expensive. Your AI is unpriced.
The "AI is too expensive" wave averages two different problems into one headline: ungoverned per-seat usage, and product inference economics nobody modeled past the pilot. Five questions to run against any AI line item before approving, renewing, or cutting it.
-
Four questions before you commit months to a build.
Building got cheap, so teams skip the four questions good product judgment always asked. None of them are technical, all are answerable in an afternoon - and AI is the worst place to skip them. Each question, with the real answer and the comforting one that hides it.
-
The build got cheap. Being wrong didn't.
Vibe-coding collapsed prototyping from months to days, and the discourse decided validation got cheap too. It didn't - in B2B, the expensive part was never the build. Why the kill questions turned from gatekeeping into triage.