IDSA’s new position paper, Data Spaces and AI: Trustworthy Agentic Participation in Data Spaces, sets out why these two developments belong together and how each supplies what the other needs.
The relationship runs in both directions, and that bidirectionality is what makes it worth examining carefully.
The governed data foundation AI needs
Most valuable data sits inside organizations that will not share it without legal clarity about permitted uses, liability and the boundaries of data sovereignty. Web scraping and generic platform terms of service cannot reliably reach this data. A data space can. It lets providers and consumers negotiate access terms directly, for each dataset and each intended use, under machine-readable usage policies that make those terms explicit and enforceable.
For AI development, this means access to curated, domain-specific datasets with verifiable provenance. Shared catalogues make data findable across organizational boundaries. Federated identity mechanisms extend trust between parties that have no prior relationship. Usage policies expressed in open standards such as the Open Digital Rights Language (ODRL) specify what data may be used for and under which conditions. These are the governance properties a general AI stack assumes are already settled. In most cross-organizational settings, they are not. Data spaces provide them.
The automation data spaces need to scale
Building a data space means integrating its components into the existing systems of many different participating organizations. That integration work is expensive. AI substantially reduces the cost by generating metadata, aligning schemas across heterogeneous backends, translating human-readable legal terms into machine-readable policies and monitoring policy compliance across participants. These are tasks that would otherwise require sustained manual effort for every new participant onboarded and every new dataset described.
The FAIR principles — Findability, Accessibility, Interoperability and Reusability — provide a practical way to map out where each direction of the relationship applies. On the findability dimension, AI supports metadata generation and semantic search, making catalogues more useful. On the accessibility dimension, the Model Context Protocol (MCP) provides a standardized interface through which AI systems can connect to heterogeneous data services. On interoperability, AI assists in aligning vocabularies and bridging terminology differences without requiring full upfront standardization. On reusability, AI can help translate legal terms into enforceable policies and assess whether datasets are fit for a given purpose.
Three patterns of collaboration
Research by Fujitsu and Fraunhofer ISST describes three patterns by which AI workloads draw on a data space. In collaborative model development, organizations train a shared model by exchanging data or model parameters, with the data space providing the usage policies and provenance records that govern what enters training. In inter-organizational model inference, one organization enriches its own model at inference time with data held by others, with the data space providing discovery, access control and usage conditions. In inter-organizational agent collaboration, autonomous agents from different organizations accomplish a task together, with the data space providing participant identity, contract negotiation and an auditable record.
These patterns recur across the pilots documented in the paper, from AI-enabled robotics in Germany to data marketplace negotiation in Japan. They are not theoretical. They are working in practice.
Implications for organizations
For AI practitioners, the paper is a caution against treating API connections as automatically trustworthy. A model-connected interface is not automatically compliant. A retrieval-augmented generation pipeline is not automatically authorized. Data spaces supply the missing layer: identity, contract, usage control, semantic interoperability and provenance.
For data space practitioners, the paper is a preparation guide. The familiar primitives — participants, credentials, catalogues, data products, Connectors, usage policies — are sufficient. What changes is that some participants will now be autonomous agents. Handling that requires making agent identity, scope, tool access and audit explicit. The foundations are already in place.
The next post in this series looks at how data spaces make AI trustworthy across organizational boundaries — covering the trust framework, verifiable identity and what the regulatory landscape requires.
