Frugal AI Models Show Uneven Accuracy Loss That Threatens Reliable Environmental Decision Making

Frugal, compressed, and edge AI models produced through pruning, quantization, knowledge distillation, and TinyML design are increasingly promoted as the accessible route by which resource-constrained firms can adopt AI-driven environmental decision tools without the computational cost of full-scale models. Yet compression degrades model accuracy unevenly and unpredictably across architectures, raising concern that the same resource constraint making frugal AI attractive as an accessibility solution also makes its accuracy risk hardest to detect. This paper reports a PRISMA 2020-compliant systematic review of the literature at the intersection of frugal AI and environmental sustainability. Searches of Scopus and Web of Science returned 283 records, reducing to 224 unique records after deduplication and 31 included studies after title/abstract and full-text screening. The corpus was characterised descriptively and coded against seven a priori constructs spanning accuracy-resource trade-offs, resource-constrained context, decision-consequentiality, the AI system's own environmental footprint, uncertainty/reliability treatment, field validation, and theoretical grounding in the Green/frugal-AI literature; an exploratory topic model (Latent Dirichlet Allocation, K = 5) provided an independent, corroborating lens. Accuracy-resource trade-offs and resource-constrained context were reported almost universally, as required by the review's own eligibility criteria, but decision-consequentiality (54.8%), the AI system's own footprint (32.3%), field validation (48.4%), and theoretical grounding (19.4%) were each reported in fewer than three in five studies. These reporting gaps motivate the Frugal AI Environmental Efficacy (FAEE) framework, presented here as a conceptual synthesis not an empirically validated typology for locating frugal-AI-for-environment applications along axes of compression aggressiveness and decision consequentiality.