A practitioner's guide to designing reliable AI/LLM-powered workflows for official statistics and research in high-accountability environments. Covers the full lifecycle from design through deployment: classification and coding workflows, data wrangling, extraction, ensemble and multi-model architectures, pipeline infrastructure, evaluation frameworks, state management and validity, security and supply chain, institutional governance, and cost analysis. Written for the interdisciplinary teams that build and operate AI systems in federal statistical agencies and similar organizations: data scientists, statisticians, research methodologists, software engineers, IT and security professionals, and the program managers who coordinate them. Each role will find chapters that speak directly to their work and chapters that help them understand what the rest of the team needs. The book introduces State Fidelity Validity (SFV), a framework for identifying how LLM-specific failure modes threaten classical research validity, and demonstrates design patterns through the Federal Survey Concept Mapper case study (6,954 survey questions, dual-model cross-validation, Cohen's kappa = 0.839). This is a living document. The design principles are durable; tool-specific guidance will be updated as the field evolves. Corrections and new content will appear in subsequent versions. The latest edition is always available at https://brockwebb.github.io/ai-workflow-design/. Companion to AI for Official Statistics (DOI: 10.5281/zenodo.19206379). Where that book introduces AI concepts for statisticians, this book teaches how to build AI pipelines that work reliably when the work matters.