Building AI products that earn trust requires treating AI as a design material rather than a feature. Key principles include exposing intent before output, surfacing uncertainty signals, designing for human participation, and decomposing outputs into stages. Drawing from experience building AI features in Jira Service Management, the author argues that dependability comes from deliberate design choices—checkpoints, confidence signals, correction paths—not just model quality. Content design shapes AI behavior through system prompts, and evaluation metrics should focus on output structure, tone consistency, verbosity, and correction cost rather than just latency and accuracy.
Table of contents
AI is a material, not a featureThe design levers that matterFrom black box to collaboratorLanguage is behavior designBounding creativity without killing itMeasuring what actually mattersDependability is designed, not assumedSort: