95% of PMs can't really code. Claude and Codex didn't change that.
What they changed is that those PMs can now create code instead of writing it. The 5% who were already decent engineers make "just build it" look easy, so everyone else tries to follow and ships output they can't fully reason about.
The bottleneck was never typing the code. It was knowing what to build. Making code cheap to produce doesn't help you if you're pointed at the wrong thing. You just get there faster.
Roadmaps have the same problem. Now that Claude generates one in seconds, people ask whether dedicated roadmap tools still earn their place. They're asking about the cheap part.
A roadmap is a visual summary of decisions you already made, arranged on a timeline. Claude produces a clean one. A roadmap tool produces a clean one. The artifact was never where the value sat.
Most of the value sits in the process around it: sync, ownership, accountability, links to work items, statuses, dependencies, the record of what changed and why. That's what you're paying a roadmap tool for, and generating a document doesn't replace it.
The rest sits in a layer almost no tool handles. A good roadmap shows why you decided to build something, not just what. What customers say on calls. What keeps coming back in support. What stalls deals in the pipeline. What's already half-built in your dev tracker.
This is where the AI conversation lost the plot. The energy goes into token maxxing: faster generation, more output, cheaper tokens, more documents an hour. But a roadmap built faster isn't a better roadmap. Impact maxxing is the thing that matters. Did what you built move the metric you cared about. That comes from the context feeding the decision, not the speed of the artifact.
And it forces a question:
how do you know you put effort where it actually matters? Right now the answer is a PM squinting at the output and deciding it looks about right. Same trap as the code they can create but can't reason about.
So the answer for the 95% isn't more code or more roadmaps. Go back to the signals. The feedback. What customers and support and the pipeline keep telling you. Make sense of it well enough to act on.
That's the layer we built at Bagel. We pull scattered signals into structured decision context, so when an AI tool generates the roadmap or the brief or the ticket, it works from what's real instead of what happened to be in the prompt. Speed was never the problem. Producing the wrong thing quickly was.
The win is reaching the right thing, the right way, at the right time, faster.
@ohadbiron