Advancing Responsible AI Innovation A Playbook 2025

Unlock AI innovation with trustworthy data governancePlay 2 Successful AI innovation depends on secure, high-quality and compliant data access with controls on processing, consent, cross-border transfers and AI deployment. Therefore, a modern data foundation must embed security into data workflows and AI systems as well as upgrade traditional security models. Organization leaders Key roadblocks that arise within the organization Low data quality and legacy processes, undermining the reliability of company systems Isolated data ecosystems and fragmented governance, limiting cohesive data strategies and enterprise- wide insight generation Complex approval processes, hindering internal and external data sharing Data scarcity, especially in categories that are prone to underrepresentation, impacting model training and risk mitigation strategies Actions for organization leaders –Implement an enterprise-wide data governance strategy: Establish guardrails for integrity and compliance that ensure data quality, interoperability and traceability across business units. Deploy data stewards to bridge centralized and decentralized governance approaches (see Case study 1). –Reduce silos and streamline approvals: Enable greater internal access to data insights and the external sharing of data by reducing legacy silos through data mapping exercises and simplifying policy approval processes. –Explore approaches to address data scarcity: Ensure company access to a sufficient volume of high-quality and representative data. Consider these approaches: –Share data between organizations: Explore the variety of sharing models and assess trade-offs. For example, data trusts are managed by a third party with a fiduciary duty to protect contributors’ interests, whereas data cooperatives are member-owned and governed. Organizations can reduce data training concerns from potential partners by contributing to efforts that standardize AI-related contractual provisions e.g. the Bonterms AI Standard clauses.21 Proactive communication with the public on the goals and limitations of a data-sharing initiative is needed to secure trust and buy-in. –Collaborate on data analysis without sharing raw data: One such method is federated learning, where a shared AI model is trained locally using data from decentralized edge devices or servers and only the model updates are shared with a central server for aggregation. Another technique employs data clean rooms – controlled environments where organizations upload their data for the retrieval of aggregated and anonymized insights.22 –Synthetic data: Examine how synthetically produced data may provide an alternative to data scarcity. Care is needed to proactively address governance challenges that synthetic data introduces, such as realism validation with non-replication of private data, provenance documentation and bias replication. Advancing Responsible AI Innovation: A Playbook 12