Skip to content
article

Our PMs started drafting PRDs with AI. Six weeks later, this is what I would keep

Elin Sjöberg@elin_pm

July 16, 2026 · 2 min read

#9 in AI Hype Check

Contents(3)

In June our product team (four PMs including me) agreed to draft PRDs with an AI assistant for six weeks. Health tech, so no patient data went anywhere near it; we used it for structure and wording, with our own call notes as input. Some honest notes, since this topic asks for what was actually tried.

What got better

  • Blank page time went to almost zero. Getting from "notes from three customer calls" to a first draft took 20 minutes instead of a morning.
  • Edge cases. Asking "what happens when the clinic has no internet / the patient changes clinic / the user is a proxy for an elderly parent" surfaced two cases we had genuinely missed.
  • Consistency. Every PRD now has the same sections. Engineers said this was the biggest win for them, which was humbling.

What got worse

  • Length. The average PRD went from about 1,400 words to 3,100. Nobody reads 3,100 words. Our engineers told us, kindly, during fika.
  • Plausible requirements. The drafts were full of requirements that sounded right and that no customer had asked for. "Users should be able to export their data in multiple formats." Should they? Which users? Why?
  • False confidence in the problem statement. A well-written problem statement feels validated. It is not. We caught one PRD with a beautiful problem section based on a single call.

The rules we kept

  1. AI may draft. A human must cut at least a third before anyone else reads it.
  2. Every requirement links to evidence: a call note, a ticket, a metric. No link, no requirement.
  3. The problem statement is written by hand. Always.
  4. Nothing about patients, clinics or contracts goes into the tool. We have a list, and it is long.

Net result: lagom useful. Not the revolution our CEO forwarded a LinkedIn post about, and not useless either. I estimate we save 2 to 3 hours per PRD and spend about 1 hour of that on cutting. Happy to be proven wrong!

#ai#startups#ux#productivity

Written by

Elin Sjöberg

PM at a Stockholm health tech startup. Ex-UX researcher. Roadmaps are hypotheses, not promises. Fika is a ritual, not a break. she/her

View profile

Get the best of #ai_hype_check weekly

One email each Monday with the most useful posts. No spam; unsubscribe in one click.

Comments

8

Sign in to join the conversation.

Sign in to comment
Giulia Bellandi@giulia_designs1 rep in AI Hype Check
"Users should be able to export their data in multiple formats" 😂 I have seen this exact sentence in three different PRDs this summer, from three different companies!! The robot has a favourite requirement ✨
0 net votes
Elin Sjöberg@elin_pm#9 in AI Hype Check
It also loves 'intuitive and user-friendly interface'. We banned the word intuitive. Team morale improved.
2 net votes
Karl Lindqvist@kernel_karl4 rep in AI Hype Check
"Lagom useful" is the most honest AI review on this site. Rule 2 should apply to human-written PRDs too.
0 net votes
Elin Sjöberg@elin_pm#9 in AI Hype Check
It does now. You are welcome to quote that one back at me, Karl.
0 net votes
Wolfgang Engel@dsgvo_wolf2 rep in AI Hype Check
I would like to point out that rule 4 is not only good practice but, depending on your setup, a legal necessity under Art. 28 GDPR if the tool provider processes personal data on your behalf. I am glad to see it written as a Rule and not as a hope.
3 net votes
Elin Sjöberg@elin_pm#9 in AI Hype Check
Yes, our DPO wrote that list, not me. Happy to be proven wrong on most things, but not on that one.
0 net votes
Danny Cho@danny_cho_la3 rep in AI Hype Check
3,100-word PRDs are a gift to every engineer looking for an excuse not to read the PRD. Rule 1 is the whole post.
2 net votes
Rachel Kim@rachelkim1 rep in AI Hype Check
Late, but sharing this with my PM group. The 'beautiful problem statement based on a single call' failure mode is the one I'd tattoo on every new PM. Writing the problem by hand is such a simple rule. Stealing it.
3 net votes