AI-Enabled Synthesis of Stakeholder and Survey Data in Operations Research: A Review of Advances and Applications

Operations research depends on inputs that are not naturally numerical. Objectives are negotiated, constraints are contested, criterion weights are argued over, and the parameters that finally enter a model are usually distilled from interviews, facilitated workshops, questionnaires and free-text commentary. Converting that material into a usable model has long been the slowest, least reproducible and least auditable part of the analytical process. Between 2023 and 2024, with unusual density in the latter year, a body of work emerged that applies large language models and related generative systems to precisely this conversion problem. This paper reviews that work and situates it within the wider applied analytics literature on data-driven decision systems, procurement and supply chain optimisation, safety and risk management, healthcare operations, energy transition planning, and enterprise financial decision-making. The paper develops a conceptual model that decomposes the path from stakeholder situation to decision model into six functional layers: elicitation, qualitative coding and synthesis, synthetic respondent generation, autoformulation from natural language, preference and weight elicitation for multi-criteria methods, and explanation, verification and governance. Six propositions derived from the model are examined against the literature, and applications are then surveyed across eight domains in which stakeholder and survey data feed directly into operational models. Three findings structure the argument. First, progress has been fastest at the two ends of the pipeline, in text-to-model translation and in automated deductive coding, while the middle, where stakeholder meaning is negotiated and converted into legitimate model structure, remains comparatively underserved. Second, the evaluation base is thin relative to the enthusiasm, with most reported gains measured against convenience benchmarks rather than controlled comparison. Third, and most consequentially for operations research specifically, the errors these systems make are structured rather than random, since persona conditioning amplifies group differences and models drift toward socially acceptable answers on sensitive items, which converts a noise problem into a bias problem that optimisation then propagates into unfair allocations. The discussion characterises this pattern as an inversion, in which the capability profile of the technology is the reverse of the difficulty profile of the practice. Theoretical, methodological, practical and policy implications are drawn out, limitations of the review, the evidence base and the model are stated, and the paper closes with a research agenda organised around verification, participatory legitimacy, and the epistemic status of synthetic evidence in decision models that carry real consequences.