NorthStar had a memory problem, not a meeting problem. Across about 120 meetings a week, decisions were hard to retrieve, actions were captured inconsistently and often without an owner or a due date, incomplete actions never made it into the next meeting, and managers had no view of overdue work across teams. The same discussions got repeated because nothing reliable was written down. The tools were all there (Teams, Otter.ai, SharePoint, Planner, Lists, Outlook, Power BI, Claude Team). What was missing was a consistent method and a shared, accountable record, with a human firmly in control of what gets published.
Modelled and pilot-tested under stated assumptions. The brief's five success measures are all met.
Across the meetings that warrant a formal record, the admin saving models to roughly 2,700 staff hours a year. That aggregate is deliberately conservative and sensitive to how many of the 120 weekly meetings get a formal record, so it sizes the prize; the per-meeting measures (the 80% cut, the 5-minute draft, the 90% owned-and-dated) are the robust, directly measurable results.
A Claude Team assistant that turns a transcript into a structured record: summary, decisions, risks, and actions each with an owner and a due date. It never invents an owner (it flags the gap for a human), never publishes on its own, and flags confidential content for careful handling.
Every action, decision and risk from every meeting in one place. Managers can finally see open, overdue and carried-forward work by team, with repeatedly-slipped items surfaced as stuck. Built as a working, clickable prototype loaded with the 20-meeting pilot data.
NorthStar's worst failure was that incomplete actions never made it into the next meeting, so they quietly died. The system closes that loop explicitly: every action gets a named owner and a due date, open actions are carried forward into the next meeting, and the dashboard counts the carry-forwards so an item slipped twice or more is flagged as stuck. The twenty-meeting pilot alone surfaced six overdue actions and four stuck ones that were previously invisible. Owned · dated · carried forward · visible
In a firm handling client-confidential material, an AI that could publish an unreviewed record would never be trusted. So the design keeps a human in control: nothing is auto-published (a reviewer approves every record), no owner is ever invented (unassigned actions are flagged, not guessed), and client-confidential content follows a clear protocol inside approved tools. The AI removes the two-hour write-up, not the accountability for what was decided.
The workflow was tested across twenty representative meetings spanning the five meeting types: client delivery, internal team, leadership, governance and sales. All five success measures were met, the human-review and confidentiality lines held, and, most tellingly, the pilot made overdue and stuck work visible for the first time. The twenty meetings are the same data loaded in the live dashboard prototype.