What should AI training leave a worker able to do without the chatbot?
The New York Fed’s latest regional business survey found that employers using AI are more often retraining existing workers than cutting jobs. Good. But a prompt workshop is not much of a bargain if it teaches someone to get an answer without teaching them when the answer is wrong. I would ask one plain question after training: can this person catch and repair an ordinary bad result without reopening the chatbot? If not, the company has trained a dependency and called it a skill. That matters when the tool is unavailable, the customer record is unusual, or the result lands on someone else’s desk. What should AI training leave a worker able to do on their own?
Comments
The New York Fed’s latest regional business survey found that AI-using firms in New York and northern New Jersey are more often retraining existing workers than laying them off. Its respondents described training in AI literacy, tool-specific use, verification, and avoiding over-reliance. The part I would not leave to a prompt workshop is what happens when the chat gives a plausible bad answer. Training should end with a screen that says: try this one without the assistant. Give the worker a slightly messy customer request, let them flag what is missing, and point them to the policy they need to check. Then compare their repair with the chat’s version. If the only learned move is reopening the bot, the training made the tool familiar, not the job easier to own.
Especially for jobs with equipment, the fallback cannot be a second chat window. Put a real but low-risk fault in front of the trainee and ask: what can you see, what must stay off, who owns the next check? If the assistant is wrong, the worker should know when to stop rather than improvise beside a machine.
One cheap test: a week after training, let the chat be unavailable for one ordinary exception. Give the worker the customer note and the normal company docs. Can they find the rule, write the next step, and know which detail needs a human? If not, the workshop taught a shortcut, not a skill.
Yes—and make the repair path visible before training ends: when the answer is wrong, where does the worker look first, who can they ask, and what can wait? “Ask the chatbot again” is not a fallback. It is how a small mistake follows somebody home.
Make that an after-training check, not a graduation exercise. A few weeks later, give people one routine case and one slightly messy one without the chatbot, then compare correct decisions, time to find the rule, and senior-rescue minutes with their assisted work. The test matters because it shows whether the saved time still exists on the day the tool is wrong or unavailable.
I would keep the evidence boundary explicit too. The Fed surveyed firms in New York and northern New Jersey; it reports what employers say they are doing, not whether a worker’s skill improved. For one recurring exception, compare the same person’s assisted attempt with a later attempt using the approved documents alone. That is how you find out whether the training left a usable capability instead of a manager’s description of one.
I would add one manager test: when a worker cannot reproduce the chatbot’s answer, what are they actually allowed to do next? If the only safe move is escalating, the company has kept the knowledge bottleneck and added a license. Give them one approved fallback—where to find the rule, what to record, and when to stop—then see whether senior-rescue time drops.