Why corporate AI training fails, and what builds capability instead

The programme ran. Attendance was good, the feedback scores were flattering, and the certificates went out. Six weeks later the workflows are exactly as they were, and the only durable output is a line in a compliance record.

Only about a quarter of companies say their training measurably improves performance. In AI, where the gap between knowing about a tool and changing a process is unusually wide, the number is unlikely to be better.

Four structural reasons

• Content divorced from the actual workflow. The examples are from a generic marketing team, not from your close process or your claims queue, and the translation is left to the learner.
• No real problem to solve. Exercises with known answers teach the tool's interface, not the judgement required to apply it where the answer is contested.
• No vendor in the room. The questions that matter — what happens with our data, what breaks at our volume — cannot be answered by a trainer.
• No artefact that outlives the session. When the room empties, the shared understanding evaporates because nothing was written down that the team owns.

Generic AI literacy is not useless. It is simply not the bottleneck, and it is being sold as though it were.

What builds capability

Four conditions, and they are largely the inverse of the list above: the department's own problem on the table; the partner who actually deploys the capability in the room; the people who will operate it present rather than represented; and a playbook that leaves the room with the team's name on it.

How to specify it

Hand a supplier a brief that says: we will bring one named workflow with a measured baseline and its P&L owner; you will bring the working software and the engineer who deploys it; the operators will be in the room for the full session; and the deliverable is a runbook our team executes the following week without you. Most training suppliers cannot accept that brief. That is the point of writing it.

How to tell within two weeks whether it worked

Do not look at completion rates or satisfaction scores. Look at whether the workflow changed, whether the change survived the first exception, and whether anyone has asked for the second workflow. If none of the three happened, you bought content, not capability.