Writing
Practical notes on AI product management, product strategy, and building useful AI systems.
Why do good teams keep building things the room already suspects won’t work? Not out of stupidity. Everyone in the room is capable. Somebody had an idea, the idea sounded good, the meeting warmed to it, and a decision got made on the strength of how good it sounded. There was no test. There was momentum, and momentum is not evidence, though it wears the same clothes. Photo by Serge Taeymans on Uns…
Something interesting happens in B2B product meetings, and I wrote about it on LinkedIn this week, but the post could only hold the small version. Here is the larger one. In the room, there are always at least two clear interests. The side selling has a vision it needs you to buy. The side buying has a business goal, which is the reason everyone cleared their morning. Both are legitimate. Both are…
Why do good people defend a failing product for years? Not the cynics, and not the fools. The decent ones. The people who got into it for a reason they can still say out loud without embarrassment. They will sit in a room where the numbers have told the same story for six quarters, and they will explain the numbers away, and they will do it with real conviction, and they will not be lying to you. …
The pilot suggested. The agent acts. Everyone is making that jump right now, and almost nobody is asking the only question that matters: is it good enough to act without you? Start with why the pilot stalled, because that part gets remembered wrong. It was not the model. The model was usually fine. The pilot stalled at the last mile, at the handoff, at the moment a human had to take the output and…
Why does a RAG pilot that answers five demo questions perfectly fall apart three weeks into production? I have been reading Chip Huyen’s AI Engineering , and her chapter on RAG confirmed something I keep running into: when an enterprise AI assistant disappoints, the model is almost never the problem. The retrieval is. Quick orientation, in case RAG is new to you RAG stands for Retrieval-Augmented …
You can now build almost any AI feature in a weekend. So why do most of them still fail? Over the last month I pulled four of them apart in public. Why they run late, why nobody can tell if they work, why they cost so much more than the estimate, and what happens when they ship and quietly break. Four separate autopsies. Except they were not separate. Pull the cover off all four and the same body …
I. NAME THE PROBLEM A silent operational failure is when an AI feature acts confident and correct while something critical is quietly breaking. Not a crash. Not a timeout. Something that looks like success while failing operationally underneath. Photo by Ticka Kao on Unsplash Take an AI assistant that triages incoming support tickets and drafts the first reply. Week one, it routes tickets to the r…
Your AI feature has been live for three months. How would you know if it has been quietly wrong the whole time? That question makes people uncomfortable, and it should. With normal software, breakage announces itself. Something crashes, an error gets logged, a page goes white, someone files a ticket. You find out because the system stops. An AI system does not stop. It keeps answering. It just sta…
You approved the budget for the AI project. So why does it now cost three times that, and still climbing? The instinct is to blame the estimate. Someone lowballed it, the scope crept, the vendor under-quoted. Sometimes that is true. But the overruns I see again and again are not estimation errors. They are category errors. The estimate priced one kind of thing, and the project turned out to be a d…
Let’s picture the situation. The data scientist is proud, the number is real, and the room is uneasy in a way nobody can quite name. Someone finally says it out loud. The model is accurate, but is it actually doing anything for us? Here is the thing. “Working” is not a property you discover about a model after you build it. It is a definition you write down before you build one. And the teams that…
Newsletter
Occasional, considered writing. No noise.
