It never prioritises, and it never scores.
Two PMs at a mid-size SaaS company, one at a marketplace, one who is the only PM at a 40-person startup and therefore also does support, analytics and release notes. They are the group most likely to have already tried three AI tools themselves, and most likely to have abandoned all three — because the tools automated the writing, which was never the bottleneck.
Two clusters break. The first is feedback: it arrives in Zendesk, in Slack, in sales calls, in app store reviews, in three different Notion pages, and reading all of it properly would take a day a week, so nobody does. The second is the status layer — the weekly update, the release notes, the stakeholder email, the ticket written up from a meeting that ended nine minutes ago.
- Pulling feedback from every source into one place, continuously
- Clustering it by theme and deduplicating the same complaint in eleven wordings
- Attaching volume, segment and revenue to each cluster
- Turning meeting recordings into written tickets with acceptance criteria
- Detecting duplicate backlog items filed four times over eighteen months
- Drafting release notes from merged PRs and closed tickets
- Drafting the weekly stakeholder update from what actually moved
- Monitoring named competitors’ changelogs, pricing pages and job ads
- What to build
- What to say no to, and to whom
- Whether the loudest cluster is the important one
- The strategy
- The feedback funnel
- Every source, one destination, clustered weekly with volume and segment attached. Crucially it keeps the raw quotes underneath each cluster — you read five real sentences per theme, not a summary. Summaries of user feedback are how PMs lose contact with users, and that is the failure mode of every tool in this space.
- The meeting-to-ticket pass
- Recording in, draft tickets out with problem statement, acceptance criteria and open questions marked as open. You edit and file. Saves roughly four hours a week for anyone running discovery calls.
- The duplicate detector
- Runs across the whole backlog. The first run is uncomfortable — one PM found 340 items collapsing to about 90 — and it is the single cheapest improvement available to most backlogs.
- The release notes draft
- From merged PRs, in the product’s voice, with the internal language stripped out.
- The competitor watch
- Weekly digest of what named competitors shipped, changed in pricing, or started hiring for. Hiring is the leading indicator and nobody reads it, because reading it is a chore — which is exactly the definition of machinery.
No RICE score, no impact estimate, no ranked roadmap.
Prioritisation is the job. A PM who accepts a machine-ranked backlog has outsourced the only part of the role that is not paperwork. Practically, the machine cannot see the commitment made to a customer in a room it was not in, or the strategic bet that makes a low-volume cluster the most important thing this quarter.
It also never sends the stakeholder update. Drafted, edited, sent by the PM. Status communication is political and the machine cannot read the politics.
Percentage of incoming feedback actually read by a human — from about 15% to near all of it. Hours per week on status and write-up. Backlog size after deduplication. Time from discovery call to a filed ticket, which should drop from days to the same afternoon.
When the real problem is that nobody at the company will make a decision, and the PM wants data to force one. More clustered feedback will not produce a decision. It will produce a better-evidenced stalemate.
Reading five quotes per cluster per week. Which was the job anyway.
I had already tried three AI tools and abandoned all three, because they automated the writing and writing was never my bottleneck. The feedback funnel keeps the raw quotes underneath every cluster, so I read five real sentences per theme instead of a summary, and that distinction is the whole thing. The duplicate pass took our backlog from three hundred and forty items to about ninety in one afternoon, which was humbling. It does not score or rank anything and I would not use it if it did.
I am the only PM at a forty-person company so I also do support and release notes. Meetings now become tickets with acceptance criteria the same afternoon instead of three days later when I have forgotten the nuance. The competitor watch reads their job ads, which I would never do myself because it is boring, and it is the earliest signal there is. Roughly four hours a week back.
Tell me what your week looks like.
I'll tell you honestly which parts you can hand over — and which you can't.