Meaningful human control needs anticipatory oversight
Human oversight is a cornerstone of AI governance, as set out in Article 14 of the EU AI Act. The ‘human-in-the-loop’ and ‘human-on-the-loop’ principles are widely recognised and effectively set the standard for human oversight. However, these terms often fall short when it comes to the use of agent-based AI, as they are based on the assumption of identifiable decision points at which a person can intervene. In practice, though, it is clear that human control in agent-based AI systems that plan, break down objectives, and execute long chains of actions occurs at the end of a chained output.
Research findings show that retrospective checks are ineffective.
Reasons for this include trust in AI, among other things. A survey of 319 knowledge workers revealed that those who trusted generative AI were less likely to scrutinise its results closely. Added to this is the issue of complexity: human predictive ability and accuracy decrease as complexity increases (Kahneman, 2011). In a sense, there is an 'automation bias' (trusting AI suggestions) as well as an 'authority bias' (technology appears more objective than humans). Only checking aggregated patterns comes too late for damage that only becomes apparent cumulatively.
This leads to a shift in perspective.
Rather than reacting to specific outcomes or being at the end of the process, human beings establish the conditions under which the agent operates. They define the objective, context, criteria, boundaries, level of risk and quality standards, and make the final decision. The AI then takes on repetitive tasks, providing support with structuring, variations, preliminary analyses, summaries, drafts and comparisons.
This is followed by several practical steps.
Define an agenda
Before deployment, establish the agent’s objectives, the limits of their authority, their guiding principles, and the triggers that will return control to a human. Base these on the responsibilities that your professional roles already entail.
Test the agenda
Run draft versions to identify cases that your specification would handle incorrectly, before real stakeholders are affected.
Reflexive and Agile Action
Examine patterns across multiple actions and use insights from this review to refine the agenda regularly. This transforms the oversight process into a continuous improvement cycle.
Document for traceability and learn
Maintain clear records of the agenda, reviews and revisions, so that you can explain and justify the behaviour of the system. If a system cannot be designed to be sufficiently transparent and adaptable, it is better to defer its high-risk deployment than to lower the standards.
Proactive oversight does not replace reactive interventions; rather, it determines when and how these interventions will take place in advance. This shifts the focus of work both ahead and retrospectively. It requires more preparation and scrutiny than simply faster results. The success of proactive oversight depends on the right tools, institutional support, and the genuine involvement of stakeholders.