When Should an AI Agent Ask for Human Approval? An E-commerce Example

An AI agent’s ability to perform a task does not automatically mean it should have the authority to complete it independently. When designing agentic workflows, organisations need to decide which actions an agent may take and which require human approval.

E-commerce refunds offer a familiar illustration. Consider a customer requesting a refund for a damaged product. In a proposed workflow, an agent could retrieve the order, check the return policy, gather supporting information, and recommend a response. Whether it should also issue the refund depends on the authority the organisation has assigned to it.

Raisch and Krakowski (2021) distinguish automation, where machines take over a human task, from augmentation, where people collaborate with machines. They argue that these approaches are interdependent rather than neatly separable. This provides a conceptual basis for considering automated processing and human judgement together when designing AI-supported workflows.

For the hypothetical refund workflow, an organisation could establish three boundaries:

These are proposed design rules, not universal thresholds. The retailer would need to assess them against its policies, operating risks, and customer obligations.

An approval step should provide an opportunity to examine a recommendation, not merely accept it. In an experimental AI-assisted decision task, Buçinca et al. (2021) found that interventions encouraging deliberate consideration reduced some forms of overreliance on incorrect AI recommendations. However, they did not eliminate overreliance, and participants rated the most effective interventions less favourably. The findings were also task-specific: the reduction was significant for one component of the decision, but not for the complete two-part task.

Applying this principle to refunds, a proposed review screen could present the relevant order information, policy conditions, unresolved issues, and recommended action. For exceptions, it could ask the reviewer to check the supporting evidence and record a reason for the decision.

This interface is a design proposal, not an e-commerce solution validated by the study. Nevertheless, it illustrates how human review could be designed around critical evaluation rather than a simple approval button.

The architectural question is not only what an agent can do, but what it is authorised to do. A useful starting point is to specify permitted actions, approval conditions, and escalation paths alongside functional requirements.

The refund workflow presented here is illustrative. It draws on research about automation and human–AI decision-making rather than evidence directly validating autonomous e-commerce refunds.

Raisch, S., & Krakowski, S. (2021). Artificial intelligence and management: The automation–augmentation paradox. Academy of Management Review, 46(1), 192–210. DOI: 10.5465/amr.2018.0072.

Buçinca, Z., Malaya, M. B., & Gajos, K. Z. (2021). To trust or to think: Cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1), Article 188, 1–21. DOI: 10.1145/3449287.