In challenging times, ensuring financial integrity and fairer distribution of services by reducing disparities are among the top priorities for social and health-care systems globally. To deliver services at population-scale, governments and healthcare agencies are increasingly automating aspects of policy rule processing – e.g., automating activitiessuch as checking citizens’ eligibility to services as computer code. The resulting code becomes the 'effective', de facto policy, and is what most citizens experience in their everyday lives. Ensuring ‘program integrity’ through consistent application of policy rules at population-scale is critical to ensuring that scarce resources get to those in need and are not lost to fraud and waste. Regulatory text policies are complex, however, and the journey from policy intent to business requirements, and eventually to coded rules, is a long, multi-step translation process. Gaps, biases or errors may remain unnoticed through this journey, eventually increasing the risk of potential failure to reimburse healthcare providers for delivering necessary services or impacting delivery of services to vulnerable people. To address this, a recent global movement known as ‘Rules as Code’ envisages a multi-disciplinary approach, bringing together legislative drafters, policy analysts and software developers, to co-create policy and code. Given the complexity of this task, it is likely that AI will play a significant role in assisting policy experts scale the production of both human-readable and machine consumable policy rules creating ‘digital twins’ of the healthcare system. However, expecting AI tools to work “off the shelf” in this complex regulatory domain results in ineffective models that perform poorly when it matters most. In this talk, I will share a perspective from experiences in a large, multi-team social care project using Ontologies and Knowledge Graphs to model human expertise. In particular, through a combination of NLP, deep learning and rich semantics, we aim to design systems where human oversight and ability-to-correct are first-class design goals, supporting experts to check that extracted actionable rules faithfully represent the original policy intent. We propose that shifting the focus towards validation of policy and its digital expression will facilitate the production of better-quality, more humane policy; and that semantics have a key role on the creation and application of technology that drives positive impact at the intersection of business and society.