Your Vendor Didn’t Change. Its Identities Did: Agentic AI and Third-Party Risk

The pressure to adopt AI has become a time-sensitive business decision rather than a controlled technology experiment. Organizations want faster reporting, less repetitive work and more responsive, efficient operations while trying to keep up with rapidly evolving technology and market competitiveness.

Traditionally, growth, maturity and efficiencies were rolled out through the adoption of new technology, standards, policies and procedures. However, AI capabilities are quickly appearing inside products employees already use, bringing new functionality into familiar environments without necessarily introducing a new vendor. This creates a challenge for traditional third-party risk management (TPRM).

Programs built around vendor assessments, named users and known systems now need to consider autonomous actors operating through application programming interfaces, access tokens and service accounts. The supplier may be familiar, but the way its platform touches the business can change. The answer is not necessarily to abandon established controls or pause innovation, but to redefine governance standards and extend them beyond the vendor relationship to the nonhuman identities (NHIs) carrying out work within and across it.

NHIs are not new. Service accounts, API keys and robotic process automation have connected systems and performed repetitive tasks without a person signing in for years. However, the purpose of those arrangements was usually narrow in scope and controlled by administrators who understood the intended workflow.

Agentic AI changes this. An agent can now pursue an objective, use connected tools and take steps toward completing work with minimal human involvement. Pair that with the rapid deployment of AI and limited governance measures, and workforce users, such as employees and contractors, may be able to configure flexible agents inside existing applications, creating new identities that do not necessarily fit into traditional procurement or account provisioning processes.

This means that an NHI that once supported predictable, specific tasks can become a vehicle for more varied activities. But that does not mean existing identity and access management (IAM) controls are necessarily broken.

Now more than ever, your organization needs to scrutinize and validate existing controls and identify the gaps. The main question is whether those controls can properly manage NHIs, as they manage human identities, and regulate access when an autonomous agent is acting independently or is prompted by a human.

The subtler problem is capability drift. An agent approved to produce sales reports may, after a model upgrade, start combining data to answer broader questions and surface nonpublic forecasts or personal data without receiving a single new permission. In a case like this, access was approved for a purpose, but the agent’s capabilities grew beyond it.

A clearly defined NHI purpose matters because permissions alone cannot distinguish every acceptable use. Although data classification is far from a new concept, it is more important than ever in the agentic era. To classify data effectively, your organization must:

• Re-evaluate data classification levels and tagging parameters.
• Properly classify structured and unstructured data.
• Define which roles may access each classification type.
• Determine whether specific purposes justify access exceptions.
• Include data owners in approval decisions to align technical access with business expectations.
• Recognize that a technically valid connection may support the defined business use while still exceeding the original intent.

Organizations need a comprehensive view of agents because a simple inventory cannot provide enough detail to understand the true exposure. Visibility must also follow the chain of connections. An agent may operate within a vendor platform while relying on downstream services or partners. Understanding each action and connection helps reveal where information travels, where decisions occur and where responsibility changes hands.

Categorization can help prioritize higher-risk use cases, such as access to sensitive data or consequential workflows. At scale, cataloging and categorization are more sustainable than trying to manage every agent through manual spreadsheets. The catalog should connect agent activity to critical applications, data and business processes rather than treating each credential as an isolated object.

A practical inventory should capture:

• An agent’s purpose
• The owner of the agent, meaning the responsible party
• The identities the agent supports and uses
• Approved access and dependencies
• The infrastructure that supports the agent

Monitoring should support that inventory with a responsible path for review and escalation. Your organization can ask whether available records show what an agent accessed and changed, and whether its activity can be connected to the initiating human where applicable.

You should decide who reviews unexpected activity, how concerns reach the accountable owner and when an investigation becomes necessary. Otherwise, collecting records without assigning those decisions can create the appearance of control while leaving teams unsure what to do with the information.

Autonomy should never eliminate accountability. A human identity must still own the business process, approve acceptable uses of data and make decisions when behavior no longer matches expectations. When a human initiates an action, that process should be traceable. When work runs independently, a data owner should remain accountable.

A sustainable responsibility matrix, such as a RACI matrix, can make these decisions clear by identifying who is responsible, accountable, consulted and informed. Applied to agents, it should clarify ownership across business teams, security, identity administration, data governance and vendor management without making everyone responsible for everything.

Lifecycle management for agents should be as disciplined as it is for service accounts. However, considering today’s broader scope of capabilities and those predicted for the future, agent lifecycle management should be as disciplined as it is for human identities where possible. Review ownership and access when:

• An agent is created.
• The agent’s purpose changes.
• The agent loses its owner.
• The agent is retired.

Establish clear, accountable handoffs, such as automated reassignment and new-owner notifications, to prevent agents from continuing to operate when decision-making responsibility is unclear.

Risk-based triggers can make reassessments repeatable. New data connections, higher-impact workflows or material capability changes may warrant review. The goal is to reassess meaningful changes without subjecting every low-risk adjustment to the same approval process.

As new AI features are added to vendor products, it is important to verify that your vendors are performing due diligence.

Examine Guardrails

Alongside regulatory posture, security frameworks and audit evidence, organizations should examine the guardrails that constrain agent actions.

Include questions for the vendor about its agentic operating model, such as:

• What safeguards keep prompts and objectives within approved boundaries?
• How are those boundaries maintained as capabilities change?

It is also reasonable to request information about feature changes, available activity records and options to pause an agent or revoke its access. Answers should be assessed against the intended use case and your organization’s ability to implement and manage the controls.

Revisit Contracts

Agreements negotiated before agentic capabilities emerged may not address current or future uses. Hosting an agent on a provider’s platform does not shift all customer risk to the provider. Shared exposure must be clearly understood.

Legal teams should review how each contract assigns responsibilities and liability, while business and technical owners clarify the workflow, data involved and downstream dependencies. This collaboration is more effective than relying on a generic responsibility model to address every aspect of agent behavior.

Review and Reassess Risk

Most importantly, reassessment should not become a disconnected purchasing exercise. Share what vendor reviews reveal with the people configuring agents and approving data use. If a supplier cannot explain a relevant boundary, document the uncertainty and decide whether the proposed use remains acceptable within the organization.

There is no single tool or universal framework that truly removes agentic risk.

Every organization will have its own level of risk tolerance, resource constraints and operational needs, even compared with other organizations in the same industry. Tooling approaches will differ as well. The sustainable and scalable approach is to develop a repeatable process that connects visibility, accountability, data governance and access decisions in a centralized workflow.

Start with important workflows, establish ownership and determine what the existing controls cover today. Then build the catalog that documents those workflows, standard review practices and vendor conversations needed to manage changes over time. This lets TPRM mature alongside adoption rather than waiting for a perfect inventory or treating each new capability as a separate emergency.

Progress can be evaluated through practical questions, such as:

• Can teams identify an agent’s owner, explain its purpose and confirm which resources it should reach?
• Can they recognize meaningful changes and route them for review?

Answering those questions consistently defines the foundation of the operating process and how the technology should benefit it.

Agentic risk spans multiple disciplines, which means organizations need a coordinated response. CDW’s risk consulting, identity and access management (IAM) and AI governance experts can help modernize TPRM tiering, due diligence and contract language for agentic vendors, discover and govern NHIs, and align agent use cases with established AI governance frameworks. This enables useful autonomy while keeping nonhuman privileged access visible, purposeful and accountable.