Key Takeaways on Agentic Trust Across Organizations

• AI systems increasingly operate across company boundaries, creating accountability gaps: When AI agents use tools, data, or services from multiple organizations, the relying party cannot fully assess whether delegated actions are trustworthy based solely on records from the original organization—a critical problem for responsible deployment.

• A new "trust evidence" framework is needed to track AI decisions across domains: The paper proposes that trustworthy cross-organizational AI requires an immutable record (a "trust-evidence envelope") that documents task authority, policy decisions, execution details, and provenance in a way that remains reviewable even after system failures, disputes, or revocations.

• Three specific conditions identify when cross-organizational AI trust becomes risky: The paper identifies propositions requiring special scrutiny: those depending on multiple domains, those needing independent verification (not just producer claims), and those that must remain auditable after problems occur—helping organizations recognize high-stakes scenarios.

• Explicit roles and transparent governance structures are essential: Rather than assuming trust, the reference model assigns specific functions (verification, policy enforcement, recovery) to clearly defined roles and trust domains, making it possible to identify who is responsible for what when things go wrong.

• Reproducible testing and metrics for AI trustworthiness are proposed: The TaaS-Eval protocol establishes standardized ways to test, measure, and report on whether cross-organizational AI systems actually maintain trustworthiness—moving beyond theoretical frameworks to practical accountability.