When Everyone Has AI but the Company Still Learns Nothing: The Messy Middle of Enterprise AI Adoption

The article discusses the painful phase of AI adoption where licenses for Copilot, ChatGPT Enterprise, Claude, Gemini, or Cursor are provisioned, but the company as a whole learns almost nothing. Ethan Mollick's Leadership, Lab, and Crowd model is cited: Leadership sets direction, the Crowd discovers use cases, and the Lab should turn discoveries into shared practices — but the learning rarely travels.

Key Problems with Current AI Adoption

• First phase looks like standard enterprise rollouts: buy seats, define acceptable use, run training, create a champion network, ask people to share use cases in a Teams channel (which becomes a dead attic).
• Second phase is messier: one team uses Copilot as autocomplete, another runs Claude Code with tight loops and reviews, a product owner prototypes real software instead of Figma mockups, a senior engineer delegates root-cause analysis to an agent and gets a valid solution in under an hour (previously two weeks), a junior produces polished code without understanding architectural implications, a support team quietly turns recurring tickets into workflow automation because nobody in the Center of Excellence ever asked the right question.
• The adoption unit is no longer the organization or even the team — it's the loop inside the work.

Why Traditional Change Machinery Fails

Communities of practice, brown-bag sessions, champion networks, enablement decks, monthly demos, surveys — these are too slow. The interesting AI work appears inside a code review, a sales proposal, a research task, a product prototype, a production incident, a test strategy, or a compliance question. By the time the story becomes a best-practice slide, the learning has lost its teeth. What made it useful was the friction: missing context, the test that failed, the weird API behavior, the moment where the agent sprawled into nonsense and someone had to pull it back.

The Elastic Loop Framework

The author suggests thinking through the elastic loop: AI collaboration is not one mode. It stretches from tight, synchronous co-driving to looser, asynchronous delegation. The real adoption question is not 'are people using AI?' but: do teams know which loop size to use? Where they need resistance? Which artifacts should survive the loop? How do those artifacts become something the organization can learn from? That is much harder than tool usage or token counting.