Preserve the learning, not the paperwork.
AI drafts a customer response in seconds. A junior employee edits it and sends it. Handling time falls. But when the next customer has a different problem, has the employee learned what needs to change?
Part I asked whether removing junior tasks can also narrow the entrance to experience. Bringing back busywork is not the answer. What matters is whether a beginner can understand a problem and make a choice.
AI can support this process too. The NBER study by Brynjolfsson, Li and Raymond reports especially strong productivity benefits for less experienced and lower-skilled customer-support workers. The authors also suggest possible learning benefits. This is evidence from a particular support setting, not a guarantee of long-term expertise across occupations.
Offer a reason before receiving the answer.
Consider responding to a customer complaint. The following is an illustrative workflow proposed by the author, not a validated corporate training program.
Before reading AI’s draft, a beginner briefly identifies the customer’s problem, facts that need checking and possible responses. This is not a test of getting the answer right. It makes understanding and gaps visible.
They then compare their reasoning with AI’s suggestion. Rather than simply adopting polished language, they explain which facts were added, whether any promise lacks support and why they chose an option. A senior colleague focuses feedback on the points where judgment differed.
This takes time. Not every task should become a learning exercise. A team could start with a few real, low-impact tasks where feedback is available. Cost-saving automation and development assignments serve different purposes.
Give a small responsibility—and visibility into its outcome.
Practicing judgment must not mean transferring major risks to beginners. Define their decision authority, escalation points and final approver first. Legal, safety and privacy-sensitive decisions need separate controls.
Learning continues after a response is sent. Was the problem resolved? Did the complaint recur? Which assumption was wrong? Delivering a document and seeing its consequences are different experiences.
Evaluation cannot stop at counting drafts. Can the employee find relevant facts in a new situation, identify an AI error with evidence and recognize when to ask for help? These are starting questions, not validated assessment measures.
Record senior review time and rework costs as well. Without time allocated for guidance, a promise of learning transfers the burden to colleagues. If only immediate throughput matters, development will be deferred again.
A beginner in 2027 needs neither endless drafting nor a role copying AI answers. They need a small but real problem, a chance to explain a choice and visibility into the result. Participation in that process matters more than the volume of AI use.
Reduce the task burden without removing the first opportunity to exercise judgment.
SCOPE AND LIMITATIONS
The workflow and assessment questions are the author’s proposals, not a proven training program. The cited study concerns AI assistance in customer support. This essay does not establish increased recruitment in Korea or long-term skill gains. 2027 is a planning horizon.