Organizations have begun to collect employee input through conversational AI agents alongside structured surveys. This paper reports a field study conducted in a global management‑consulting firm that paired a pre‑survey with an adaptive AI voice interview covering the same topics in a single session. Forty‑four first‑session interviews yielded 132 matched theme observations. Depending on the favorability threshold, 20%–41% of sessions showed an initial positive rating followed later by a substantive concern. A Gioia analysis of 158 protective quotes drawn from 65 eligible sessions revealed that disclosures were rarely unguarded—employees softened their language, deflected accountability, and bounded the scope of their statements. This protective work correlated strongly with perceived legitimacy of the listening structure. From these findings we derived a bounded‑disclosure model and four propositions for voice, channel, and listening research: (1) positive ratings can coexist with genuine concerns; (2) the adaptive nature of the channel shapes disclosure depth; (3) organizational legitimacy moderates employee defensiveness; (4) silence may persist within apparent expression. The study suggests that designers of AI‑mediated listening systems must look beyond surface ratings to uncover hidden worries and avoid equating positive scores with true satisfaction.
Review