Pick one visibility surface you haven’t been measuring consistently and establish a clean baseline. You’re not trying to win the metric this week — without a baseline, improvement is mostly storytelling.

From the field What I’m seeing inside the work.

I’ve been learning this lesson inside my own business in a very literal way. Over the past several weeks, I realized I was using AI for too many kinds of work without clearly separating what actually required reasoning from what should have been handled by systems, automation, templates, or deterministic workflows. So I started rebuilding the operating layer underneath it: what should live in a source of truth, what should be automated, what needs human approval, which AI gets which kind of work, and what should never need an AI decision at all. It turns out “using more AI” is not the same thing as having a better AI system. The more capable these tools become, the more important the structure underneath them becomes. That is showing up in the market too.

• Persistent agents are becoming less theoretical: Microsoft is moving Copilot toward longer-running agentic work, and Claude Code is moving more work into cloud sessions that keep going after you step away.
• The interesting part isn’t that agents can work longer — it’s that this turns AI into an operating-cost and governance question: what work deserves an agent, what permissions should it have, who reviews the output, what usage cost is acceptable, and where human accountability is still required.
• I’d watch this closely before automating everything that moves.

Where are you still using AI manually for work that probably belongs in a system, template, workflow, or automation instead?