The bottleneck in most AI-heavy businesses is approving much more so than generating. Just delivered an assessment to a founder who already runs Claude Code, a second brain on a Mac Mini, and a one dedicated wiki per client. His constraint is that every call summary and framework has to be manually reviewed before it enters the knowledge base. Generation is instant but ratification piles up. The fix, in two sizes: Quick win (this week) -Add his assistant to TailScale so she can reach the Mac Mini -She runs the existing intake skill on each Fathom transcript -Output goes to him for a yes or no, then into the second brain Major project (later) -Fathom fires a webhook when a transcript is done -An agent listens and runs the intake skill automatically -A routing skill sends the result to the right wiki (main brain first, then the client's) -Assistant approves or rejects in one shared channel Net savings we estimated: 1.5 to 2 hours a week after 30 to 60 minutes of review time.