Key takeaways

• Agency has three parts: an autonomous decision, an action across systems, and loop-closure. Remove any one and you have an assistant, however well-briefed.
• The Button Test settles it in one question: does a human still have to press a button for the core loop to close? If yes, it is an assistant.
• The Agency Ladder has six rungs. Roughly 90% of so-called enterprise agents sit at rung 3 — a configured assistant — and the work that produces ROI lives at rung 5 and above.
• Microsoft is not the villain for lacking agents; it has real ones. The damage is that the same word covers a SharePoint Q&A bot, so 'we have Copilot' gets heard as 'we have agents'.
• Assistants make a human faster (linear, saves minutes). Agents remove the human from a class of work (nonlinear, removes the task). You are paying transformation prices for efficiency outcomes and wondering why the P&L never moves.