EnterpriseVal argues the GenAI ROI gap is a measurement failure and specifies the frozen configuration under test

arXiv·medium signalXBlueskyLinkedInCopy link

Submitted 2026-09-18, EnterpriseVal (2609.21841) starts from the mismatch between public benchmarks answering "what can the model do" and deployment decisions needing "is this workflow fit, reliable, safe and worth scaling, here, on our data, under our controls." It proposes a use-case-level evaluation system whose first component is a formal specification of the use case plus the frozen socio-technical configuration under test, naming the model, prompts, retrieval, tools and guardrails as part of the artifact being measured. For builders, the transferable idea is treating the whole harness configuration as the versioned unit of evaluation, not the model.