Not quite the exec-with-AI-roadmap version, but a smaller cousin of it: I gave Claude access to our Jira board to help with my own admin, and the first week it happily updated ticket statuses off stale Slack context nobody had corrected yet. Same root issue really, an AI moving fast without the judgment call a PM would make about what's actually still true.
What fixed it wasn't a smarter model, it was fixed rules: treat ticket comments as the only source of truth, never infer from a status field that might be stale, and nothing gets written back without being shown to me as a draft first. Once that was in place the 'wait, why did it touch that ticket' moments mostly stopped. Your three questions feel like the same idea applied a level up, force the AI (or the exec wielding it) to show its work before anyone trusts the output. Wrote the fuller rule set up as a small kit if useful: https://aipmkit.com
Not quite the exec-with-AI-roadmap version, but a smaller cousin of it: I gave Claude access to our Jira board to help with my own admin, and the first week it happily updated ticket statuses off stale Slack context nobody had corrected yet. Same root issue really, an AI moving fast without the judgment call a PM would make about what's actually still true.
What fixed it wasn't a smarter model, it was fixed rules: treat ticket comments as the only source of truth, never infer from a status field that might be stale, and nothing gets written back without being shown to me as a draft first. Once that was in place the 'wait, why did it touch that ticket' moments mostly stopped. Your three questions feel like the same idea applied a level up, force the AI (or the exec wielding it) to show its work before anyone trusts the output. Wrote the fuller rule set up as a small kit if useful: https://aipmkit.com