Microsoft cut Copilot from its own apps and usage went up. That sentence should stop you cold. Jacob Andreou, the EVP running Copilot at Microsoft, told The Rundown AI that the company stripped Copilot out of several Windows apps after discovering that high-traffic entry points delivered almost no real value. They cut buttons. Usage per person went up. The lesson isn't complicated: putting AI everywhere doesn't make it useful. It just makes it ambient noise.
Andreou also admitted something owners running agents should hear. He personally hit a wall with Copilot Tasks. Ordering food and calling Ubers felt like magic for a few weeks, then felt hollow. And for harder work, like financial planning, agents were, in his words, "an inch deep." That's not a knock on AI. It's an honest map of where it is right now. The magic is real in the shallow end. The deep end is still shallow. Know which water you're wading into before you build a workflow around it.
One practical thing worth noting from the same interview: Copilot's Auto mode now picks the right underlying model for each task automatically. You don't have to know whether GPT-4o or Claude or Gemini handles your request better. The system decides. That's genuinely useful for owners who don't want to become AI model comparison shoppers.
Shelly Palmer is asking a question most owners are skipping. His piece is short and worth your five minutes. The argument is simple: personal AI agents only get powerful when they hold persistent context about your life. Your relationships, your purchases, your plans, your conversations across months. That compound memory is what makes them feel like a real assistant instead of a fancy search bar. But it also means handing one service an unusually deep, ongoing view of your business and your clients.
This matters right now if you're building AI-assisted customer experiences or deploying agents on behalf of customers. What context is the agent holding? Where does it live? Under whose terms? Palmer doesn't panic about any of this. He just asks you to make the trade-off consciously, not by accident. Most owners are making it by accident.
Trending on GitHub
bilawalsidhu/gods-eye-view, A live satellite view of Earth in your browser, built on real open-source spatial data rendered on a photorealistic 3D globe. Think Google Earth but pulling live intelligence. (886 stars today)
lyogavin/airllm, Run a full large AI model locally on a single consumer GPU with only 4GB of memory, the kind of card many people already own. Normally that takes way more hardware. (548 stars today)
chrislgarry/Apollo-11, The actual source code that flew the Apollo 11 lunar module and command module to the Moon in 1969, preserved and readable. A piece of history more than a tool, but worth knowing it exists. (466 stars today)
Also Worth Knowing
Datadog analyzed LLM failures across 1,000+ enterprise customers and found that 30% of all errors in production are rate-limit hits, meaning the provider throttled the request because too much was being asked at once. If you're building on AI APIs, plan for that failure mode before you scale, not after.
AAA Accelerator added a live session on running AI workshops for your employees, scheduled for Thursday October 15, 2:00 to 3:00 PM CDT. Good option if your team is stuck at "AI curious" and you want to move them toward actual daily use.
The AI Pulse has a simple framework for deciding what you should still do yourself vs. hand off to AI, worth a look if you're drawing those lines inside your own business right now.
Q4 2026 State of AI and Autonomous Economy event at LA Tech Week is on Zoom October 13, free to register, good for owners who want a broader read on where AI infrastructure is heading.
-Scott
SmartOwner is published (almost) daily by the team at DigitalTreehouse. Want AI consulting or automations for your business? Reply to this email.