Knowledge work has relied on people to turn messy reality into finished outcomes. People decide what is needed and put solutions into practice. AI agents are automating parts of this work.

AI agents can now orchestrate actions to pursue an objective. A system can interpret a goal, select tools, observe the results, and decide the next step.

Agents can therefore undertake work that has depended on human flexibility: interpreting information, transforming and creating outputs, applying judgement within defined boundaries, and coordinating actions.

This raises questions about governance, expertise, and human relationships.

When organisations delegate execution to agents while leaving a person formally responsible, they risk creating what Madeleine Clare Elish calls a “moral crumple zone”. Without time or visibility to intervene, that person’s oversight becomes ceremonial while their responsibility remains real.

Accountability must therefore be paired with legibility and authority to intervene. Rather than asking staff to sign off on finished deliverables, organisations should ensure the agent’s inputs, sources, actions, exceptions, and points of uncertainty are inspectable. Oversight also requires the ability to challenge or stop the process.

Professional development often depends on people building tacit knowledge through routine execution. If agents absorb those tasks, organisations need new ways for people at any career stage to acquire the expertise required to set direction, exercise authority, and assume accountability.

Development should combine practice with exposure to successful outcomes, edge cases, competing approaches, and flawed agent outputs. Expertise will be built through practice, comparison, and evaluation. Over time, this provides the basis for greater responsibility in setting goals, planning work, and directing agent-supported processes.

In many forms of knowledge work, the relationship is part of the work. A recommendation, decision, or explanation gains meaning and legitimacy partly from the person who offers it, responds to others, and stands behind the outcome.

As agents reduce the time needed to produce routine outputs, more value may shift towards problem-framing, deliberation, and consensus-building. Time released by automation can be used to create the shared understanding needed for solutions to be trusted and adopted.

Taken together, these shifts change the conditions under which authority is exercised, judgement is formed, and relationships are sustained.

Rebundling knowledge work requires deciding what to delegate, where authority sits, how future expertise will be developed, and where human presence is part of the work itself.