When agents take the routine work, how do people learn judgement?
Summary: Lawrence Knowles raises concerns about how the integration of AI agents in handling routine cases may impact the development of judgment in junior colleagues. They discuss the importance of maintaining opportunities for learning and suggest methods such as retaining representative cases for human involvement, enabling a review of decisions made by both humans and AI. They emphasize the role of human approval in ensuring task quality and learning, and inquire about others' experiences in maintaining judgment skills alongside AI deployment.
If AI agents take on more routine cases, how do junior colleagues develop the judgement we will later expect them to use when supervising exceptions?
I’m exploring this while developing Agent Control, an experimental approach to organising and verifying work carried out by agents. It has made me think about two responsibilities that belong together: deciding whether a task is good enough, and helping people learn why.
A human approval button is only useful if the person has the knowledge, time and authority to challenge the result. Yet routine work is often where that knowledge begins to develop. We could remove the practice while retaining the expectation of expertise.
One approach worth testing would be to keep a small set of representative cases for colleagues to work through, compare human and agent decisions, and review what happened afterwards—including occasions when both were wrong. Acceptance criteria should be understandable to the people doing the work and revisited when requirements change.
For those introducing AI into people processes: are you deliberately preserving opportunities to practise judgement? What has helped people become better at challenging an answer, rather than simply approving it?
Wristy -
·4 days agoThis is the question I keep coming back to in HR automation too. The "approve" button only works if the person pressing it could have done the work themselves 😅
Two things we've tried:
📋 Leave exceptions: juniors decided past cases before seeing the agent's answer. The agent followed policy well but missed context a person would catch, and the "why did we differ?" chat taught more than any training deck.
💼 Offer approvals: reviewers wrote one line on why an offer sat within band instead of just clicking approve. Reviewing the misses together, including ones where both of us were wrong, built sharper challengers.
So, what helped most so far is:
🔹 Practising on real past cases
🔹 Making approval an active step, not a Friday 4:55pm click
🔹 Giving juniors early exposure where judgement actually grows
🔹 Revisiting the criteria whenever policy changes, so they stay readable to the people using them
Your point about acceptance criteria being understandable to the people doing the work is the one I'd underline. Otherwise the reviewer is a rubber stamp with a job title 😄
Looking forward to seeing what Agent Control surfaces!