Peter Kim Frank writes on DEV about what it takes to move a team from a few AI power users to everyday use across everyone. The original article is at https://dev.to/peter/what-team-wide-ai-adoption-actually-takes-1gbh. His argument is that getting one person productive with AI is easy, while getting a whole team to use it daily is a different problem, and the model is rarely what stands in the way.
The factory, not just the engine
Frank frames better models and more compute as the engine, and says they don't build a factory on their own. In software, the factory is orchestration, review, testing and CI/CD around the code. Knowledge work needs the same surrounding structure, and the goal is what he calls usable productivity, described as a function of visibility, collaboration, transparency, judgment and leverage. A tool that is brilliant for one person but invisible to everyone else scores low on most of those.
Mindset shifts
- Build in public: people learn AI by watching each other, so work in shared channels and threads lets teammates see the prompt, the result and the correction, then reuse them. Private chats and personal terminals teach nobody.
- Hand off the busywork, keep the judgment: AI takes repeatable chores while people decide at the moments that matter.
- Experiments are cheap: a proof of concept costs minutes, so ask for more than seems reasonable. He cites a teammate who asked an agent to regenerate a day's emails for comparison and then requested a diff view.
- Turn what works into rules: when an experiment proves itself, make it deterministic as a script, a check or a scheduled job, so generated output becomes something the team can rely on.
Tooling that fits existing habits
Frank compares adopted AI tools to a humanoid robot: not the best at any single job, but general-purpose because the world is built for humans. The working world is made of chat channels, threads and docs, so tools should meet people there rather than adding another window. He argues for giving work a shape through projects and threads, for native integrations that the whole team connects once instead of per-person MCP auth setups, and for security that isn't homework for less technical staff.
On security, he contrasts the common pattern of API keys pasted into .env files on every laptop, personal security habits and localhost MCP servers, with a shared setup where secrets are entered once through a secure form, agents use credentials at runtime without seeing the values, and integrations are connected centrally.
He also argues good team tooling avoids forced tradeoffs: scheduled automations and on-demand help, many models in one place, cloud machines and your own hardware, autonomy and a human in the loop, exact code and model judgment.
Disclosure and closing
Frank names Capy, which lives in Slack, as the closest fit for his own team, and notes it is not the only way to get there. Readers should weigh that recommendation accordingly. He closes by asking how other teams use AI day to day, especially teams that aren't mostly engineers.