Automation leaves people with the hardest work

Essay

The business case is approved on the promise that AI will remove repetitive work. The first reports look encouraging: fewer manual steps, faster throughput and a cleaner queue.

Then the remaining work changes.

The straightforward cases pass through the system. People receive the exceptions: the ambiguous customer, the conflicting evidence, the case that does not fit the rule and the decision that will be difficult to undo. Volume may fall while judgement, emotional load and consequence rise.

The leadership decision is not simply what to automate. It is what authority and support must move towards the people who inherit the hard parts.

Efficiency can hide a transfer of difficulty

Automation is often measured through the task it removes. That view can miss the work it concentrates. When a system handles predictable cases, the human role becomes less frequent and more exceptional. The person may intervene only when the normal process has failed, the evidence is incomplete or the stakes are unusually high.

This pattern has a long history. In her 1983 paper “Ironies of Automation”, Lisanne Bainbridge examined how automated systems can leave people responsible for monitoring and recovery while reducing the routine involvement through which they maintain understanding. The paper concerns industrial automation, not today’s generative AI. Its relevance here is interpretive: removing routine participation can make the remaining human task harder, not easier.

A contemporary leader therefore needs to look beyond the number of tasks automated. Ask what kinds of work now reach a person, what information they receive and what happens when they disagree with the system.

The problem appears at the boundary

A well-performing model can still create a poorly designed job. The risk appears where the automated flow hands work back to a person.

The handover may arrive without the context needed to understand why the case was flagged. The employee may be accountable for the outcome but unable to change the recommendation. Targets may still reward speed even though the remaining cases require care. Escalation may exist on paper but feel unsafe in practice.

None of these is a model-accuracy question. They are choices about work, authority and responsibility. If leaders leave them implicit, frontline teams will resolve them through workarounds. Some will approve recommendations they do not trust because challenging them takes too long. Others will build parallel notes and spreadsheets to recover the context the formal system has lost.

Those behaviours should not be dismissed as resistance. They are evidence about the design of the boundary between human and machine.

Four leadership choices

1. Decide what must remain visible to the person

Someone asked to review an exception needs more than an output. They need the relevant evidence, the system’s limits and a clear account of what has already happened. The appropriate level of explanation depends on the decision and the consequences; a generic confidence score will not always be enough.

2. Match authority to accountability

If a person remains accountable, they need a meaningful ability to pause, change or escalate the decision. “Human in the loop” is weak protection when the person has neither time nor permission to exercise judgement.

Leadership should be able to name who may override the system, on what grounds and what happens next. An override should create learning, not suspicion of the employee who raised it.

3. Redesign measures around the remaining work

A queue of exceptions cannot be managed as if it were the old mix of routine and difficult cases. Time-per-case may rise for a sound reason. Quality may depend on consultation across teams. Emotional labour may increase when automation routes the most sensitive cases to people.

Measures, staffing and recognition need to reflect that shift. Otherwise the organisation celebrates efficiency at system level while making the human role less sustainable.

4. Keep enough participation to maintain judgement

People asked to intervene rarely can lose the familiarity that good intervention requires. Leaders should decide how reviewers will maintain context: through sampling routine cases, structured review, simulations, peer discussion or another method suited to the work.

The aim is not to preserve manual activity for its own sake. It is to protect the capability the organisation still expects people to provide when the automated path is no longer adequate.

A conversation before the next approval

Before approving another automation proposal, put one real workflow on the table. Follow a normal case and an exception from beginning to end. Then ask:

• Which work disappears, and which work becomes more concentrated?
• What does the person handling an exception need to see?
• Can that person challenge the recommendation without penalty or delay?
• Who owns the result when human and system judgement differ?
• How will leaders know whether the remaining job is becoming harder?

The answers may change the design, the measures or the pace of rollout. That is not an argument against automation. It is the work of making the promised improvement real for the organisation as well as the process.

Authority has to move with the difficulty

AI can remove drudgery. It can also concentrate uncertainty. Both may be true in the same workflow.

The senior team does not need to predict every exception. It does need to set a position people can rely on: which judgements remain human, what authority accompanies them and how the organisation will learn when the boundary is wrong.

If the remaining work demands more judgement, the remaining role needs more than a new interface. It needs a leadership decision.