Can AI replace a product manager?
Last Updated: 21.06.2025 12:38
AI is as good or better than I am at some of the tasks I used to spend time on, but it still falls short on many tasks. This answer will likely be different in years or even months, but here’s my evaluation of how AI compares to me on some common tasks, as of October 2024:
Writing user-facing communications — mediocre. Similar to writing product specs, it’s good at making a template and some boilerplate text, but I haven’t figured out how to get it to use a tone that doesn’t sound super AI-y, and I end up rewriting 70+% of what it writes.
Writing internal communications — bad. Maybe if my workplace wasn’t full of straight-talkers, it would be more useful. At least where I work, it’s valued to incisively get to the point about what matters, and AI lacks the judgement to do that. It’s also usually verbose, and I find that when I ask it to be more concise it usually just sounds bratty.
More seniors are using cannabis than ever before despite health risks, research shows - Fox News
Brainstorming — good with oversight. I’ve used generative AI to “premortem” an idea or to write out foreseeable questions given a product spec. AI did very well at both of those, but I still needed to go back through and answer the concerns and questions using real-world human judgment.
Data pulls - bad but presumably will be good soon. AI helps very little at this point, mostly because the tools I use don’t have knowledge of the data sources I need to use, and most of the work is in making sure I wield the data sources correctly. I believe this problem will be solved really easily by some enterprise RAG product that takes in something like table descriptions and maybe example notebooks.
Research — good with oversight. Very similar to labeling, I’ve found that AI is good at finding what matters and pointing me there, but I usually want to read the sources myself. It’s also super helpful that audio generative AI can let me listen to longer text while I do other non-brain-intensive tasks.
What the ‘Revenge Tax’ Is in the Tax Bill—and How It Could Pummel the Dollar - Barron's
Labeling — good with oversight. It can save a ton of time to have a model look through user feedback to extract quotes about particular topics, but from my experience, I still usually want to read the extracted quotes myself rather than using AI to synthesize them, as the synthesis sometimes misses the mark on what I actually needed to know.
Scripting (i.e writing small bits of code to automate something simple) - very good. AI is much, much faster than me, and often right. At the current moment, it needs a little bit of oversight from me, but using Cursor (AI code editor), a script that used to take me half a day can now be done in maybe 10–15 minutes.
Writing product specs / PRDs — mediocre. AI can be pretty helpful at writing a template quickly and filling in some basic ideas, but it fails at focusing on expanding on what’s important and at making wise decisions, since it lacks real-world context.