Designing reviewable human–AI workflows for professional safety decisions

Authors:

Abstract
,AI assistance in professional work requires interfaces through which people can examine recommendations, determine their disposition, and preserve a record of that decision. Evaluating these interfaces is difficult when ostensibly similar logged actions have different meanings across tasks. We report a design science study of reviewable human–AI workflows in a Brazilian industrial safety platform. Formative interviews with nine professionals contextualized the design of task-specific evidence displays, intervention controls, manual alternatives, and linked decision records. A naturalistic evaluation examined 2,012 records from 65 users across 64 client accounts and six workflows between 24 April and 9 July 2026. The records contained 1,940 approvals, 23 edits, and 49 rejections; 97.6% of outputs were retained unchanged or after modification. All registered edits occurred in the textual practical-recommendation workflow. An earlier interface illustrated a consequential measurement difference: an unchanged saved text indicated retention, whereas a non-edited record in the later interface could indicate either approval or rejection. A separate risk-rating analysis reconstructed one professional override but largely comprised approval-defined value pairs. The contribution is a situated design account connecting task evidence, professional intervention, and retrospective review, together with an evaluation approach that interprets action metrics through the controls and recording rules that produce them. The findings establish deployed intervention opportunities and recorded use; they do not establish appropriate reliance, independent decision quality, or improved safety outcomes.

This publication has 14 references indexed in Scilit: