Leading AI-fluent teams

Every organisation is somewhere on the AI adoption curve, and most are stuck in one of four recognisable failure modes. Naming yours is the first step; this playbook covers the four, and the leadership work that gets a team past them. 1. The four failure modes Prohibition: AI is banned or so restricted it might as well be. The work does not stop — it goes underground, onto personal accounts and personal devices, where none of your controls apply. Prohibition does not reduce AI risk; it removes your visibility of it. Free-for-all: everyone uses everything, nobody reviews anything. Quality becomes uneven in ways that surface months later, and sensitive data quietly flows into tools nobody assessed. The energy is real; so is the liability. Tokenism: there is a pilot, a demo day, and a slide for the board — and no change to how real work is done. Tokenism is the most comfortable failure mode because it photographs like progress. Tool-worship: licences are bought, victory is declared, and the hard work — redesigning the work itself — is skipped. Adoption metrics look healthy. Output quality tells the truth. 2. Fluency is a property of work, not people Training courses produce familiarity; redesigned work produces fluency. Pick the workflows where assistance is expected, build the AI step into the way the work is actually done, and make review of AI-assisted output an explicit skill — taught, practised, and valued. The team that learns to review well learns faster than the team that learns to prompt well, because review is where judgement compounds. 3. Policy that enables Good policy is short and legible: what is allowed by default, what needs a gate, what is prohibited and why. Then the part most policies miss — a fast path to ask. If the answer to "can I use this for that?" takes three weeks, people stop asking, and you are back to the underground. Make the safe way the easy way: sanctioned tools that are genuinely good, defaults that protect data without ceremony. 4. Measure honestly Usage is not fluency. Licences active and prompts sent measure enthusiasm, not capability. Measure what the work shows: the quality of reviewed output, cycle time on the redesigned workflows, escalation and rework rates, and whether the review step is catching what it should. If assisted work is faster but rework is up, you have measured the truth early — which is the point. Publish the measures to the teams doing the work, not just upward: people improve what they can see, and hiding the scoreboard from the players helps nobody. 5. Leaders go first Teams calibrate on what leaders do, not what they announce. A leadership team that visibly uses the tools, visibly reviews the output, and visibly talks about what worked and what did not gives the organisation permission to do the same — including permission to report honestly when the tool made something worse. 6. The first 90 days Find out where AI is already touching your work, including the unsanctioned uses — that map is your real baseline. Choose two workflows that matter and redesign them with assistance and review built in. Publish the short policy with its fast path. Teach review before you teach prompting. Report the honest measures to your own leadership, especially the unflattering ones — the fastest way to license honesty below you is to practise it above. The goal is not an organisation that uses AI. It is an organisation whose judgement about AI compounds — workflow by workflow, review by review — until fluency is simply how the work is done.