Generative AI and the architecture of knowledge work: Expertise bypass, exchange governability, and shifting bottlenecks

Research Summary Research on GenAI and work has largely focused on tasks. We ask when task‐level effects alter knowledge‐work industry architecture: the allocation of activities across firm boundaries. Using a three‐stage abductive design—an executive roundtable, a panel of 626 firm websites (2017–2025), and 33 interviews plus two further roundtables with C‐level executives and corporate directors—we document divergent repositioning: professional services firms foreground relational value, while interface‐oriented technology firms emphasize scalable interfaces. We find that activity reallocation is shaped by two independently varying dimensions: expertise bypass and exchange governability. Their intersection yields four configurations: contestable reallocation, persistent bundles with selective insourcing, federated expert provision, and integrated expert provision. We extend industry architecture theory by explaining change when production and exchange conditions shift separately. Managerial Summary GenAI is not simply making knowledge work more efficient—it may be changing which organizations perform which work. We document that professional services firms are emphasizing relationship‐driven, judgment‐intensive work, while interface‐oriented technology firms are extending scalable delivery. Our framework explains why task exposure alone does not determine where work moves. We distinguish between expertise bypass—whether GenAI can produce an output without the expertise previously required—and exchange governability—whether that activity can be specified, verified, priced, and warranted across firm boundaries. Where both are present, work becomes contestable and may be reallocated. Where they diverge, highly exposed activities may remain bundled, while less exposed, human‐intensive work may move through federated specialist models. These conditions shape which organizational forms become viable.