The PM artifact stack is collapsing into prompts
Most of the documents we write are about to get cheap. Here is what does not, and why.
A PM job, stripped to its surface, is a stack of artifacts.
The one-pager. The PRD. The market analysis. The competitive teardown. The launch brief. The post-mortem. The roadmap. The status update. The exec summary. The Teams message that took 40 minutes to write because it had to land right.
For 20 years, the people who could produce that stack on demand, and produce it well, were the people who got promoted. The artifact was the job. If the spec was sharp, the launch tended to be sharp. If the post-mortem was honest, the next launch tended to be less broken. The artifact was both the work and the receipt that the work happened.
That model is collapsing. Most of these documents are about to get cheap. A few are not.
Here is my read on which is which, and why the difference matters more than the trend.
What is actually happening
The cheap version of the story is that AI “writes the docs for you.” The more accurate version is that AI removes the cost of a respectable first draft. The price of a passable spec, a passable competitive analysis, a passable launch brief, has dropped to near zero for anyone who can write a coherent prompt.
That sounds like a productivity story. It is not. It is a redistribution story. The work that used to live inside the artifact (the thinking that happened while the PM was writing the spec) now has to happen somewhere else, or it does not happen at all. And the place it has to happen is the PM’s head, before and after the model runs.
So the right question is not “which artifacts survive.” It is “where does the thinking move when the artifact gets cheap.”
The artifacts that collapse
These are the ones a model now produces faster than I can, at a quality level that is good enough to ship internally with a half hour of editing.
The competitive teardown. A model reads ten competitor pages, three analyst reports, two community threads, and produces a respectable two-pager. The PM job becomes choosing what is wrong in the draft and what is missing, not generating the draft. The artifact stays, the labor inside it changes shape.
The status update. The weekly “here is where we are” doc, the all-hands slide, the partner update, the monthly stakeholder note. These are aggregation work. Models aggregate. The PM who used to spend Friday afternoon assembling status now spends 20 minutes editing what the model assembled. The artifact survives. The Friday afternoon does not.
The market analysis. The 30-page deck that took two weeks and an analyst is now a 90-minute exercise. The deeper version (the one that includes a real point of view about where the market is going) is harder, and the model does not do it well. So the deck collapses and the conviction is what is left, written shorter and signed by a named person who will own being wrong about it.
What these three have in common is that they were always synthesis work. The PM was the human glue between sources. The glue is now mostly machine. The work that remains is editorial: what is true, what is missing, what does not pass the smell test, what is worth defending.
The artifacts that stay human
These are the ones I have tried to delegate to a model and watched fail in ways that mattered.
The spec, but only the part you would defend in court. Models write very good first drafts of specs. They write a usable user flow, a credible set of edge cases, a reasonable success metric. What they do not write is the line that says “we are choosing this trade-off and rejecting this other one, and here is why we will defend it in three months when it is unpopular.” That line is where the spec stops being a document and becomes a commitment. A model does not have skin. It does not get held to the decision. The part of the spec that gets shipped to a real customer and stays your call is exactly the part that stays human.
You can let the model draft the rest. You cannot let the model draft that line. If you do, you will not be able to defend it, because you will not actually know why you chose it.
The post-mortem about the launch that flopped. Models write fine post-mortems for launches that worked. The retrospective on a success is a recitation, and recitations are easy. The post-mortem about the launch that broke is hard because the honest version requires saying out loud what you did wrong, what you saw and did not say, who you avoided pushing back on, what assumption you were holding that turned out to be a fantasy. A model cannot do that work. It does not have the context, and it does not have the standing. The post-mortem on the flop is one of the most valuable documents a PM produces in a year, and it is exactly the one that does not compress.
The pattern: artifacts that require a named person to take a position on something they could be wrong about do not collapse. They are the ones that get more valuable as everything else gets cheaper, because a clear position with a name on it becomes the scarce thing.
The one I am still not sure about
The roadmap.
I have watched models produce roadmaps that are 80 percent right and 20 percent wishful in a way that is hard to detect. A roadmap is a sequence of bets, and a model that does not face consequences for the bets cannot tell you which bet is real and which one is the comfortable one. So the artifact looks plausible and is quietly wrong.
On the other hand, the roadmap is the artifact that most senior PMs spend the most time rewriting, because it changes every two weeks anyway. So the cheap-first-draft model is genuinely useful here, even if you cannot trust the draft.
I think the roadmap is going to split into two artifacts. The plausible-sequence document, which a model produces and a PM edits in 30 minutes. And the “here is the bet we are making and what we are willing to be wrong about” document, which is one page, signed by a person, and updated when the bet changes. The first one is the schedule. The second one is the spine.
If you are still writing both into one document, the cheap part is going to drag the expensive part down with it.
What this means for the week
If you are a senior PM right now, the play is not “use AI for everything you write.” The play is to figure out, for each artifact on your stack, which of these three buckets it lives in. Then ship the cheap ones cheaply, and protect the expensive ones from being mistaken for cheap.
The mistake I keep seeing senior PMs make is using the same workflow on both. They either AI-draft the spec and ship it without sharpening the load-bearing line, which leaves them exposed in three months. Or they hand-write the status update like it is 2022, which is fine for them and disrespectful of the time of everyone who reads it.
The artifact stack is not collapsing uniformly. It is collapsing where it was always thin, and getting heavier where it was always load-bearing.
The PM job in 2027 is going to be the same job it was in 2007, with most of the in-between work taken away. You will write less. You will defend more.
See you next Tuesday.
Luiz
Views are my own and do not represent Microsoft.