An SME owner approves an AI agent to prepare reconciliations, draft customer replies, and produce first-pass analysis.
The result looks good. Work moves faster. Senior staff spend less time clearing routine queues.
Then a harder question arrives: if the agent now does the work juniors used to learn from, how will those juniors become the people trusted to review the agent three years from now?
This is easy to miss because the short-term numbers can look positive. The queue gets shorter. Turnaround improves. The team produces more without adding headcount.
None of those numbers shows whether the business is still developing judgment.
Junior work was doing two jobs
Routine work produces an output. It also trains the person doing it.
A junior accountant preparing a reconciliation learns more than where to place the numbers. They see which source documents are late, which balances regularly fail, which explanations sound plausible but are wrong, and when a small mismatch points to a bigger process problem.
A junior operations executive drafting a customer reply learns which details calm a frustrated customer, which promise creates trouble for fulfilment, and when the standard answer does not fit the case.
A junior analyst building a first pass learns where the data is weak, which assumptions change the conclusion, and which questions a manager will ask before accepting the work.
Bluegain's 5 October 2026 paper on the human operating model for agentic enterprises makes this point directly. Visible production work also exposes people to ambiguity, challenge, feedback, customers, exceptions, error diagnosis, and progressively greater responsibility. The paper argues that leaders must redesign these development paths as AI absorbs more routine production.[1]
That distinction matters. A business can automate the output and accidentally remove the learning.
Do not preserve manual work for nostalgia
My view is simple: do not make people keep doing repetitive work because that is how previous generations learned.
If an agent can prepare a clean first pass, use it. Making a junior copy figures between systems does not become valuable merely because it is familiar.
But I would not confuse removing manual production with removing responsibility.
I would redesign the work like this:
- The agent prepares the reconciliation.
- The junior checks the evidence behind unusual entries.
- The junior diagnoses the exceptions instead of just clearing them.
- The junior explains what failed, what changed, and what should happen next.
- A senior challenges the explanation and decides whether the junior can own a harder case next time.
The agent handles more of the preparation. The human spends more time on diagnosis, context, and consequence.
This is a better use of both.
Four things the new role must replace
When you automate junior production work, check whether the new role still provides four things.
1. Exception exposure
People build judgment by seeing where the standard process breaks.
Do not route every unusual case straight to the most senior person. Give juniors a controlled exception queue. Let them investigate, form a view, and present the evidence before a senior decides.
The goal is not to let inexperienced staff approve high-risk actions. The goal is to let them practise the reasoning that comes before approval.
2. Feedback
An AI-assisted first pass can make weak work look polished. That makes specific feedback more important.
"Looks fine" teaches nothing. A useful review identifies the missed signal, the weak assumption, or the piece of evidence that should have changed the conclusion.
The junior should see the correction and then update the reasoning, not just accept the senior's final answer.
3. Customer context
Many operational decisions make sense only when you understand the person or business affected.
If agents handle every routine interaction and seniors handle every escalation, juniors may lose direct exposure to customers. They will know the workflow but not the reality around it.
Keep them close to selected conversations, complaints, handovers, and follow-ups. A CRM record is useful. It is not a substitute for hearing how a customer describes the problem.
4. Growing responsibility
Judgment develops when the consequences become real in controlled steps.
Start with diagnosis. Then let the junior recommend an action. Later, allow approval within a narrow limit. Expand the limit only when the evidence supports it.
A role with no increasing responsibility is not a development path. It is a permanent observation seat.
The pipeline problem
The previous article on this site focused on exception knowledge as a scarce AI-era skill. This article asks a different question: who is being developed to hold that knowledge in the future?
It is possible to agree that human judgment matters and still destroy the process that creates it.
That risk will not show up in an agent demo. It may not show up in the first year of cost savings either. Senior experts can keep the system running for a while because they already carry the context.
The weakness appears later. Fewer people can challenge the agent. Reviews become procedural. The same small group handles every difficult case. When one expert leaves, the business discovers that the operating knowledge was never transferred.
That is not an AI failure. It is a role-design failure.
One action for this week
Pick one workflow your business is automating now.
Write down which parts of the current workflow train people. Look for the moments where a junior sees an exception, receives correction, speaks to a customer, or takes on more responsibility.
Then assign those training moments explicitly in the redesigned workflow.
Do not leave them as accidental by-products of manual work. Make them part of the role, the review process, and the progression criteria.
Use the agent to remove unnecessary effort. Keep the experiences that produce people capable of questioning the agent.
Otherwise, the business may get faster every quarter while becoming less able to judge its own work.
Sources
[1] https://bluegain.com/insights/whitepaper/the-human-operating-model-for-the-agentic-enterprise — The Human Operating Model for the Agentic Enterprise