How Analysts Use AI in High-Stakes Crime Linkage: An Industrial Study

Research Context and Approach

Crime linkage analysis is a cognitively demanding task where experts search large databases to identify series of offenses committed by the same individual. Because this process is time-consuming and involves exposure to disturbing material, researchers co-developed an AI-enabled decision-support tool with the UK’s National Crime Agency. This tool provides ranked predictions of potential crime links along with visual explanations of the underlying features (e.g., behavioral similarity, geographical proximity).

To evaluate the tool's usability and integration into real-world workflows, the authors conducted an industrial mixed-methods study. They combined direct observation, eye-tracking, mouse-tracking, and post-session surveys to observe six expert analysts as they performed realistic linkage tasks using the tool. This approach allowed the researchers to capture both objective interaction patterns and subjective user perceptions.

Key Findings

The study revealed that analysts do not blindly accept AI suggestions. Instead, they treat AI predictions as a starting point, consistently validating them against traditional, non-AI behavioral evidence. Analysts valued the transparency provided by the tool’s feature-level explanations, as these allowed them to understand the 'why' behind a prediction and align it with their own professional judgment.

Eye-tracking and mouse-tracking data confirmed that analysts actively engaged with the provided model features and frequently utilized the behavioral matrix—a tool mirroring their existing manual practices—to verify potential links. While the tool was generally perceived as easy to use, the findings highlight that for AI to be truly effective in high-stakes environments, it must be designed to complement, rather than replace, the established analytical methods that experts rely on for accuracy and accountability.

This research underscores that in high-stakes domains like law enforcement, the success of AI is not determined by predictive accuracy alone, but by how well the system integrates into existing human workflows. The study provides actionable evidence that designers of decision-support systems should prioritize transparency and the ability for users to cross-verify AI outputs with traditional evidence. By doing so, developers can build tools that are not only usable but also foster the trust necessary for operational adoption.