AI Governance Institute
← News
Research2026-11-12

Vendor Custody, Delegated Authority, and the Agentic AI Accountability Gap

What happened

The paper "Who holds the judgment? Agentic AI, vendor custody, and the accountability gap" was published in a peer-reviewed Springer Nature journal on 5 November 2026. It analyzes accountability when an AI agent acts with delegated authority. Covered actions include sending communications, initiating transactions, and making recommendations affecting third parties. The research finds that vendors define the behavioral constraints built into the model, but deploying organizations cannot independently verify those constraints or inspect the reasoning behind specific decisions. When an agent causes harm, neither the vendor contract nor existing liability doctrine cleanly assigns responsibility. The authors argue this creates a structural accountability gap that sits outside what current incident response programs, model governance frameworks, and third-party risk management programs are designed to close. The findings reinforce regulatory signals from the Five Eyes Guidance on the Careful Adoption of Agentic AI Services and the EU AI Act (Regulation (EU) 2024/1689). They also add academic weight to concerns in Deloitte's finding that 40% of enterprises let AI act without human approval.

Why it matters

  • ·Regulatory frameworks including the EU AI Act (Regulation (EU) 2024/1689) assign accountability to deployers, not vendors. If a vendor's alignment controls fail and an agent causes harm, the enterprise deploying the agent is the named responsible party. Compliance teams cannot assume vendor safety commitments transfer that accountability.
  • ·Because vendors retain custody of the behavioral controls embedded in AI models, enterprises have limited ability to verify that an agent will behave as contracted across all conditions. This makes standard vendor due diligence, which typically relies on documentation and certifications, insufficient for agentic deployments where behavior can vary based on inputs the enterprise never reviews.
  • ·Incident response programs assume investigators can reconstruct what happened and why. When an agent's decision logic is opaque and controlled by a third party, post-incident review is constrained to what the vendor chooses to disclose. This creates a notification and audit trail gap that regulators are increasingly examining, as seen in Banking Examiners Are Now Asking Banks to Prove AI Agents Can Be Audited.

Stay ahead of stories like this

Get every Global AI governance development like this one, plus the rest of the week's developments. Every Thursday.

Free every Thursday. Unsubscribe anytime.

Governance controls affected

AGT-021Human Oversight Classification Rationale LogAGT-006Agent Action Audit TrailHOC-002Human Approval Gate for Consequential AI DecisionsPRC-001AI Vendor Due DiligenceIRC-002AI Incident Response Playbook

What to do now

  • ☐Review every active agentic AI deployment and confirm in writing which party, the vendor or your organization, is responsible for each category of harm the agent could cause, including errors, unauthorized actions, and data disclosures.
  • ☐Ask your legal and procurement team whether current vendor contracts give your organization the right to obtain logs, decision rationales, or behavioral audit records when an AI agent causes a compliance event or regulatory inquiry.
  • ☐Require your AI governance team to document, for each agentic deployment, what human oversight checkpoint exists before the agent takes an irreversible action, and confirm that checkpoint is functioning and recorded.
  • ☐Assess whether your incident response plan covers scenarios where the root cause of an AI agent failure is inside the vendor's model and cannot be independently verified by your team.
  • ☐Identify any agentic AI deployments where your organization is the regulated entity under sector rules (financial services, healthcare, employment) and confirm those deployments meet the specific human oversight and auditability requirements in those sectors.

What to watch next

Regulators are moving quickly to formalize deployer accountability for agentic AI. The EU AI Act (Regulation (EU) 2024/1689) enforcement apparatus is now active. The Five Eyes Guidance on the Careful Adoption of Agentic AI Services signals coordinated supervisory attention across multiple jurisdictions. Compliance teams should monitor whether sector regulators in financial services and healthcare issue specific agentic AI accountability guidance in 2027. Teams should also track whether vendor contract standards evolve to include explicit inspection rights and incident notification obligations for agent behavior. Litigation against AI vendors over agent-caused harm will begin to shape liability doctrine. Those outcomes will affect how deployer accountability is interpreted in practice.

Related Coverage

Corporate Policy2026-11-04

CrowdStrike's Agent Kill Switch Sets a New Containment Bar for Rogue AI

CrowdStrike has added an AI kill switch to its Falcon Guardian platform. It lets authorized administrators stop a misbehaving autonomous AI agent, block it from restarting, and revoke its access to enterprise systems. Every action the agent takes is recorded in an audit trail. Security analysts say the tool is a useful emergency brake but must be paired with agent discovery, identity management, and incident response controls to be effective.

Research2026-11-12

OCC Supervisory Framework Puts Generative and Agentic Bank AI Under Examiner Scrutiny

The OCC issued a supervisory framework on November 4, 2026. It guides examiners reviewing how national banks and federal savings associations use AI in credit decisioning, payments, fraud detection, customer service, and risk management. The framework is non-binding and sets no new enforceable standards. It fills a governance gap left by the April 2026 interagency model risk management guidance, which explicitly excluded generative and agentic AI as 'novel and rapidly evolving.' Banks with material AI deployments in those categories now face structured examiner scrutiny without previously having been required to build matching controls.

Research2026-11-11

AI BOMs Are the Missing Link in Enterprise Supply Chain Transparency

SBOM expert Allan Friedman argues in a SecurityWeek podcast interview that software bills of materials must move from static inventory documents to active inputs for vulnerability management and risk response. AI systems complicate this by introducing model provenance, training data, runtime data sources, and ephemeral agent actions that traditional software inventories do not capture. Friedman describes AI BOMs as an emerging necessity that converges software inventory concepts with logging and third-party risk management.