Yesterday I spoke with a senior leader at a fast-growing investment firm. They’d done the “hard part,” at least on paper: enterprise AI tools were approved, rolled out, and available to everyone.

And yet, the reality looked familiar.

A few people were power users. They were automating real work, building repeatable workflows, even setting up lightweight “cron job” style routines to prep for meetings.

Others? Barely touching it. Not because they were resistant. Because day-to-day work is loud, incentives are local, and “try AI” is not a process.

The leader’s question wasn’t “Which model should we pick?” It was much more operational:

How do we move from ad hoc usage to systematic adoption without turning AI into another top-down initiative everyone tolerates and nobody owns?

The problem is most AI rollouts over-index on tools and under-index on shared context.

The visible part of AI adoption is prompts, skills, automations. It's what I think of as the tip of the iceberg. The part that actually makes the output consistent (and shareable) is below the waterline:

• A small set of context docs everyone agrees are “source of truth” (your thesis, your value prop, your operating principles, your communication style)
• A consistent folder structure so AI work products don’t become a junk drawer
• Clear boundaries on permissions (especially once you connect AI to shared drives)

This is also why “skill sharing” often disappoints. People don’t want to read each other’s AI output. They want results. A skill without context is usually just AI slop with a logo on it.

The unlock is surprisingly unsexy: small groups, real workflows, and time on the calendar. Not a big training. Not a mandate. A cadence where a team picks one workflow, builds the context around it, iterates, and compares “rework required” week over week.

AI adoption isn’t a license. It’s a management system.

AI looks so easy people think it great outcomes should be easy to create.

As with anything in life, real quality comes from effort.

Alex