First, check the requests you doubt. Build what the evidence supports.
A short verdict with evidence labels, written for product and engineering alike.
Request #212 · from product. Add an AI summary to every report. Planned for sprint 14. Checked before anyone writes code.
Request #215 · from product. Export the dashboard to a spreadsheet. Planned for sprint 14. It's a small build.
Request #219 · from product. A chatbot on the pricing page. Planned for sprint 14, after a competitor launched one.
Measured · interviews. 3 of the last 40 user interviews mention a summary. Counted from the interview notes, by keyword search and by hand. Interviews read 40; Mention a summary 3; Asked for it 1
Measured · the category. Four tools in the category already offer one. Users can compare it with theirs. The question is whether yours would be better.
Not known · stated. No data says how often people would use it. A short test can close that gap before the build.
Reviewed · the research. The survey behind the request asked: "Would an AI summary save you time?" A question like that invites a yes. We count it as weak evidence.
Measured · support tickets. 41 tickets in three months ask for an export. Counted from the support system, duplicates removed.
Measured · the category. Most tools in the category have one. Users expect it. Its absence shows up in the tickets.
Measured · accounts. 9 paying accounts asked for it by name. Each one is in the account notes, with the date.
Reviewed · the research. Tickets, account notes and the category review agree. Three sources, collected separately.
Measured · interviews. None of the last 40 interviews mentions a chatbot. Questions about pricing do come up. They are about discounts for annual plans.
Estimated · the category. Common on pricing pages, with little public evidence that people use them. An estimate from a review of public pages, not usage data.
Not known · stated. No data on whether visitors would use it. A test would need a threshold set before it starts.
Reviewed · the research. The request cites one competitor's launch. A launch shows what they built. It doesn't show that their users wanted it.
Work that checks itself, and reports what it found.
Prototype · sample reports. Our five sample reports pass 62 automated quality checks, with 0 failures. There's also a blind review covering 23 types of defect. An earlier internal blind audit found problems. Since then, every fix ships with an automated test.
Published · Walmart Labs. Add-to-cart rose in three of four categories and fell slightly for laptops. The paper shows the drop next to the gains. arXiv 1811.07996
Prototype · sample reports. Measured, declared or estimated, or not applicable with a reason. When two sources disagree, the more certain one wins.