Measure AI literacy before you announce the rollout

The common sequence runs backwards. An organisation picks a tool, announces it, and then asks how to train people to use it. By then the announcement has set expectations, the licences are paid for, and the gaps in understanding show up as support tickets, workarounds and quiet avoidance.

A baseline taken first changes the sequence. It should not be a quiz on what a large language model is. It should test applied literacy: can this person judge whether an output is trustworthy, do they know which data must never be pasted into a tool, and do they know when to stop and escalate. Those three answers predict more about how a rollout will go than any measure of enthusiasm.

The baseline then feeds a fluency gap map, laid out by role and team, running from basic literacy to the point where someone can work comfortably with an agentic system that takes multi-step actions. The map shows where agentic systems will change roles first, and those roles should get support first. Training everyone at the same pace spreads effort evenly across people whose exposure is very uneven.

For a multi-market organisation the baseline matters even more. Literacy differs by market, by language, by local regulation and by how comfortable local managers are with the tools. HR teams that have run people operations across more than twenty markets see this quickly: one global curriculum lands well in two offices and badly in the rest, and the headline completion rate hides it.

There is a trust dimension. A workforce that does not understand a tool cannot calibrate its trust in it, so it either defers to the output or rejects it wholesale. Both are failures. Literacy is what allows people to trust a system to the right degree, which is why it should be settled before the choice of model is final, not after.

For boards, this is a governance question, not an HR one. Leadership needs safe environments for experimentation and clear lines of accountability, and neither works if the people involved cannot tell a reliable output from an unreliable one. A direct test of readiness: can you show your board a literacy baseline by function?