Five Questions to Ask Before Building an AI Agent

In this piece for Nesta, authors Seth Caldwell & Simon Wisdom share lessons from their experiments with agentic AI systems, focusing on how organizations can determine whether a problem is suitable for an AI agent solution. Using a case study involving heat pump installers overwhelmed by complex administrative processes with electricity network operators, the authors explain how agentic AI can help automate repetitive tasks such as handling emails, documents, and workflow coordination.

The article introduces a five-question framework for evaluating potential agentic AI projects. The first questions focus on whether outcomes can be verified, whether the underlying process is sufficiently understood, and whether the problem represents a meaningful operational bottleneck. The authors emphasize that AI agents work best when tasks can be broken into measurable steps, especially where work is repetitive and resource-intensive but still requires some human oversight and judgment.

Caldwell and Wisdom also caution that organizations must design these systems around real human behavior rather than assuming “human-in-the-loop” safeguards automatically prevent mistakes. In testing their DNO communications prototype, they found that users often approved AI-generated actions without carefully reviewing them, especially when the system appeared highly automated. This revealed the importance of carefully calibrating when agents should act autonomously, when they should request approval, and when genuine human expertise is required.

Finally, the piece stresses that successful deployment depends heavily on organizational trust and collaboration. Building effective agentic AI systems requires buy-in from leadership, engineering teams, and end users, particularly given concerns around privacy, security, hallucinations, and workflow integration. The authors conclude that organizations should approach agentic AI not as a universal solution, but as a targeted tool best suited for solving neglected operational bottlenecks where automation can meaningfully free human experts to focus on higher-value work.

CONTINUE READING…