The Wrong Unit of Automation: Why Task-Level Profitability May Fail at the Workflow Boundary

Artificial-intelligence adoption studies have advanced beyond technological exposure by asking whether automating a particular task is economically attractive. Yet an economically attractive task does not necessarily imply an economically improved process. A substitution may change exception rates, routing probabilities, waiting times, downstream verification, case acceptance, capacity constraints, and the allocation of work among people and systems. This paper develops the Workflow Boundary Hypothesis (WBH): a task-level investment recommendation can reverse when evaluated against the same counterfactual at the boundary of completed, independently accepted workflow cases. We define a task-to-workflow valuation bridge, an interaction residual, an accepted-case accounting convention, and bounded non-entailment propositions. Two transparent numerical constructions illustrate both a false-positive task business case and a genuine workflow-level gain; neither is presented as observed enterprise evidence. A preregisterable research design compares task-only screening, fully instrumented conventional business-process economics, and the proposed boundary-explicit representation under matched information, costs, and intervention rules. The primary outcome is out-of-sample decision regret relative to independently adjudicated workflow net value, complemented by accepted-case cost, throughput, cycle time, exception burden, tail risk, and measurement overhead. The proposal does not claim that existing process mining, simulation, activity-based costing, causal inference, or queueing theory is inadequate. Its incremental value exists only if explicit boundary accounting improves decisions beyond those strong baselines. No enterprise experiment, performance improvement, universal theorem of AI adoption, or demonstrated scientific Field is reported.