We stopped describing the computer as a 'tool' the moment it became a teammate: we delegated to it, edited with it, and occasionally argued with it. Generative AI is heading the same way. But an organisation wanting to be 'AI fluent' cannot stop at teaching prompt formulas. It must enforce the same standards it applies between two competent colleagues: check the output, know when to escalate, and judge when a machine's suggestion is confidently wrong. Fluency is judgement relationship management, not vocabulary.
Define AI fluency as a ladder with four rungs. Operate, the basics of generating and refining output. Collaborate, handing work correctly back and forth and archiving the division of labour. Escalate, knowing when the model has reached its edge and a human must take a higher-risk decision. And audit: catching a plausible-sounding hallucination, introducing a test, and holding assumptions to account. Most enterprises are spending every hour on the first rung. The hidden risk is that they park there while the rest of the staircase quietly pulls the service down.
The practical build is rehearsal, not viewing. Role-based scenarios — a customer reply to an escalated ticket, a code review containing a fabricated API call, a contract summary with a silently dropped liability — cannot be watched; they have to be run. For each scenario the participant decides, argues, approves and logs their own escalation — not just clicks along. Empowerment, not attendance, is the currency of this training, and permission to be wrong is the safety feature that makes it experiment.
Finally, design for the moment the learning is most honest: the failure. Give staff leave to raise a bad AI result without fear of an 'AI scrape', surface these incidents to a review board, and let them refine the calibration. A team that treats the model as a peer — with limits to respect, outputs to challenge, and standards to hold — outperforms a team that treats it as a button to press twice. That shift, from novelty to trusted operator, is what separates desktops of tools from actual teammates.