Data Governance as a Decision Variable: A Model-Driven Decision Support System for Budget Allocation in Generative AI Adoption

Corporate adoption of Generative Artificial Intelligence (GenAI) is routinely managed as a problem of technological scaling: firms decide how much to use the tool while treating the quality of the underlying knowledge base as a given feature of the environment. This paper shows that such a formulation is structurally incomplete and yields systematically biased recommendations. We present a model-driven decision support system (DSS) that endogenises data governance through an explicit budgetary variable and jointly solves the allocation between investment in information quality and the intensity of algorithmic use. The formal core is a non-linear, two-variable optimisation model in which data quality simultaneously modulates the value production function and the cost structure, through a variable cost coefficient and a convexity exponent that captures the agency cost of human oversight—the Babysitting Tax. We prove the existence of a scissor effect, whereby data degradation shifts marginal value downwards and marginal cost upwards at once; we characterise the Marginal Profit Destruction Zone, in which the firm remains profitable while destroying margin on every additional query; and we establish a viability corridor for governance investment bounded both from below and from above. We further prove that, for firms with severely fragmented data architectures, conditional optimal profit is non-monotonic, which gives the decision the character of a big push rather than of a marginal adjustment. On this core we build a five-layer artefact—telemetry, estimation, optimisation, policy, and interface—whose policy layer materialises in three executable budgetary decision rules. We show that these rules decentralise the centralised optimum between the Finance function and the Data function through the exchange of two scalars only, and that the convergence condition of the protocol coincides exactly with the second-order condition of the problem. The two-variable problem reduces to a scalar search of logarithmic cost, which makes the artefact computable in real time. Simulation over 10,000 replications and two calibrated organisational profiles quantifies the value of the artefact: a firm with fragmented data that applies the conventional static model over-uses the technology by 27.8%, believes it is obtaining a profit of +32.0 monetary units and in fact realises a loss of 36.7, whereas the DSS solution places it at +8.4.