Teach delegation and verification together
Beginners should not be taught to treat AI agents as dependable colleagues by default. Independence is earned by demonstrated reliability on the particular task.
A lot of training now starts from the premise that the tools are reliable and the learner's job is to hand work over. In my experience the opposite order is safer and, oddly, faster. Teach people to specify bounded work, to recognise uncertainty, to verify outputs against a standard, and to stop or escalate. Then let the tool earn autonomy, task by task.
What this looks like in a session
Every exercise ends with a verification note: what was checked, what was found, what was corrected. Participants keep a record of common errors. By week three the record is a better training document than anything I could have written in advance, because it is about their work.
The internal champions are assessed by teaching a colleague and by handling a deliberate error case. If they can spot the failure and explain the recovery, the practice has a chance of surviving a staff change.
Why it matters for the people in the room
Staff are right to be wary of a programme framed as getting more output from them. Agreeing how released capacity will be used, involving practitioners in the design, and refusing to promise employment outcomes nobody can control are not soft extras. They are what makes the learning stick.
By Shadrack Munene
