AI systems are increasingly embedded into the everyday infrastructure of modern workplaces, from resume screening systems to AI features built directly into productivity tools. Many AI tools—such as automatic speech recognition (ASR) tools or writing assistants—do not make explicit decisions. Yet, these tools can meaningfully influence how people are evaluated by mediating how work is produced and communicated. Understanding these effects requires expanding beyond the narrow scope of much existing fairness research, which has focused primarily on reducing model errors, mitigating bias in training data, and optimizing performance metrics between groups. To build AI systems that are equitable and inclusive, it is equally essential to understand the human experience—how AI tools shape what people see, how they evaluate others, and how individuals feel seen within AI-mediated environments. This dissertation investigates how AI tools can produce differential outcomes even when they are not explicitly making decisions. In the first study, we examine how the audience evaluates the speaker and their content when there are inaccurate ASR generated subtitles, demonstrating how technical errors can quickly become social judgments. The second study illustrates how the mere suspicion of using an AI tool—regardless of whether it was actually used—can negatively influence job performance and hiring outcomes. The third study demonstrates how the stylistic mismatches between an AI tool’s suggestions and a user’s voice can undermine feelings of inclusion, control, and ownership. Taken together, these studies show that AI systems function not merely as productivity aids but as sociotechnical infrastructure that shapes perception, evaluation, and lived experience while completing work tasks. By foregrounding human experiences, this dissertation identifies harms that remain invisible to model-centric evaluations and contributes a framework for supporting agency and fostering inclusion. In doing so, it charts a path toward more equitable, accountable, and human-centered measurements for AI tools.