Research

Organizations increasingly use information systems to help knowledge workers manage portfolios containing many interconnected products and customers. Drawing on complexity science, we distinguish portfolio complicatedness, the number of interconnected components a worker carries, from complexity, which can emerge when interactions among those components become ad hoc and nonlinear. We theorize that deterministic CRM features impose predefined structure that should keep increasingly complicated portfolios tractable, whereas probabilistic GenAI features generate context-dependent outputs and iterative exchanges that can fuel complexity dynamics. Using agent- and deal-level data from U.S. independent insurance distribution and a shift-share instrument identification strategy that addresses endogenous portfolio composition and technology use, we find patterns opposite to these expectations. CRM use negatively moderates the performance effects of both product and customer complicatedness, while GenAI use positively moderates them. The shapes of these relationships reveal boundary conditions on the theorized mechanisms: predetermined structure is valuable when complicatedness creates interaction demands that require coordination and retrieval, but can constrain performance when those demands remain tractable; probabilistic interaction can impose costs at low complicatedness yet become valuable as informational demands exceed what a worker can readily maintain. The findings extend complexity science by showing that information systems can improve performance not only by taming complexity through structure but also by increasing a worker’s capacity to perform as complicatedness creates increasingly demanding interaction conditions.

Digital platforms evolve through continuous feature releases into systems knowledge workers already depend on, making innovation and reliability inseparable. Reliability research treats resulting failures as costs offsetting innovation, but this view is incomplete. Using Real User Monitoring traces of 34 million actions from an InsurTech CRM serving insurance agencies, we estimate a causal mediation system in which feature use and IT failure are separately endogenous, each with its own shift-share instrument. Feature use raises conversion. It also raises IT failure. Failure in turn raises conversion, and this indirect pathway is 22% of the total effect. Decomposing failure across five categories of knowledge work shows why. Recoverable internal failures are followed by retry and fast resolution and contribute positively. Customer-facing failures are followed by abandonment and contribute negatively. Recoverability determines whether disruption creates or destroys value.

Knowledge work is now routinely delegated through agentic AI to search, read, extract and compose on a worker’s behalf. The outputs read fluently and are often sycophantic, shaped to please the worker rather than to establish that a claim is true. Verifying them requires provenance that lasts: a record of the documents the AI could consult, the passage it relied on, and the path from that passage to the claim, still reachable long after the work. Today that record is lost by default, because the passage sits inside a session that closes.

We call this delegation without durable provenance. It is distinct from hallucination and model opacity, since it arises even when a claim is accurate. Attribution, data provenance and scholarly policy each address part of the problem, but each settles the question at composition. We argue that evidentiary standing is not settled there: it must be sustained for as long as the claim stands.

Following a design science approach, we develop a design framework for such an environment and build ProRead, an open-source, local-first environment in which agentic AI reads and composes over a collection the worker holds. Each passage receives a fixed address that outlasts the session and travels into the notes, summaries and maps built upon it. This makes verification affordable, and affordable verification is more likely to be performed. And because the agent must attach an address to what it asserts, it curbs sycophancy.