AI

From Automation to Augmentation: Designing AI Coaches That Amplify Expertise with Mike Amundsen

O’Reilly Media
O’Reilly Media•Jun 1, 2026

Why It Matters

Because AI coaches preserve human judgment while scaling expertise, they help firms avoid skill erosion and harness generative AI as a strategic augmentative asset.

Key Takeaways

  • •Most users treat AI as answer machines, not thinking partners.
  • •Only 5‑10% use generative AI to expand their own thinking.
  • •AI coaches structure interaction, prompting human decisions before generation.
  • •Coaches enforce boundaries, preventing AI “brain‑fry” overload in workflows.
  • •Augmentation amplifies team capabilities, lowering entry barriers to expertise.

Summary

In a recent talk, Mike Amundsen contrasted automation with augmentation, arguing that generative AI is being deployed primarily as an answer‑machine rather than a tool that expands human expertise. He introduced “AI coaches”—software agents that structure the human‑AI interaction, slowing the process to force reflection and decision‑making.

Amundsen cited usage data showing roughly 30% of users let AI make decisions, while only 5‑10% employ it to sharpen their own thinking. He warned that this “surrogate” mode erodes judgment and skill formation, referencing an Anthropic study where developers solved bugs faster with AI but failed to understand the solutions.

Demonstrating his concept, Amundsen walked through a coach for building a small web app. The coach begins by asking permission, guides users through exploration, refinement, and a final “commit” boundary, then generates code and explains the steps taken. He highlighted the “stopping problem,” noting that unlike open‑ended generators, coaches know when to halt, avoiding the “AI brain‑fry” of endless output.

The approach promises to turn AI from a productivity shortcut into a capability multiplier, enabling teams to retain creative processes while leveraging machine speed. By embedding decision points and explicit boundaries, organizations can preserve skill development, broaden participation, and mitigate the risk of over‑reliance on opaque outputs.

Original Description

"People are using [AI] mostly as an answer machine and an advice machine when they really should be using it as a way to expand their own thinking," argues API designer and author Mike Amundsen in this talk from AI Codecon. He's spent years thinking about how humans and systems interact, and he has a pointed diagnosis for what's going wrong with generative AI.
Most people fall into one of three patterns: using AI as a generator (answer machine), a surrogate (decision maker), or an instrument for sharpening their own thinking. Only 5% to 10% use it the third way. The rest are outsourcing answers or, worse, outsourcing judgment—and an Anthropic study on AI-assisted debugging shows exactly what that costs: faster results in the short term but an inability to explain why something worked or fix the next problem. Mike traces this back to how AI is packaged (speed, efficiency) versus what it could actually do (expand human capability), drawing on a lineage from Vannevar Bush to Doug Engelbart. His solution: AI coaches, structured interactions that slow the pace, prompt human decision-making, restore the brainstorm-practice-commit loop, and—crucially—know when to stop. Because AI doesn't stop on its own, which is how you end up buried under 5,000 lines of unreviewed code and what Harvard Business Review calls "AI brain fry." If you're thinking about how to use AI to make your team smarter rather than just faster, this one is for you.
Follow O'Reilly on:

Comments

Want to join the conversation?

Loading comments...