Artificial intelligence (AI) is getting embedded in enterprises. Organizations are now deploying AI across software development, customer service, financial analysis, and executive decision-making. However, as AI adoption continues to accelerate, a prominent risk is surfacing which is way beyond the enterprise’s own tech stack. This is the third-party AI ecosystem.

Organizations are supposed to govern AI systems which they build themselves while they continue to rely externally on vendors for data services, foundation models, SaaS applications, APIs, as well as agentic platforms. This ends up creating a sort of governance conundrum. Although an enterprise may not own the model or operate the infrastructure beneath, however, it can be held liable for using it. Such a situation makes Third-Party AI Risk Management (AI TPRM) an important component of AI governance.

“There are two important things here. It is very important that the customer has full control over his configuration and data. The customer should be cognizant of what kind of data is going on his vendor platform and how it is getting stored,” says Dinesh Kamble, CISO, RBL Bank.

Gartner’s latest research augments this argument: “AI represents both increased cybersecurity risk to organizations via third parties, and increased opportunity to minimize the impact of third-party incidents.” This statement is drawn from Gartner’s report from Feb 2026. The research report says that cybersecurity leaders, including CISOs, need to shift towards monitoring that is integrated with GRC.

AI TPRM must go beyond the traditional security questionnaire

Research firm McKinsey goes on to make a similar point while analysing AI governance for financial institutions. It recommends a risk scorecard which is supported by technical review, vendor due diligence, as well as contractual controls. The report provides an extremely significant observation for vendor assessment:

“Procurement teams can use a mix of vendor due diligence, technical reviews of underlying models, and contractual safeguards to assign risk scores to third-party software and make more informed purchasing decisions,” observes McKinsey on vendor-assessment.

It is worth noting that a conventional vendor evaluation can establish whether a supplier has identity controls, vulnerability management, and certifications. Those controls continue to be essential; however, they are not enough for AI. The assessment needs to examine the technology in itself.

Enterprises need to be aware of several aspects, including how models are trained as well as fine tuned, what data is used and where it is stored, and whether the data can be utilized to improve upon models supplied by the vendor. They also have to establish how a vendor identifies hallucinations, explainability of outputs, and controls to detect model drift.

Cybersecurity teams should also consider attack surfaces that are AI-specific. These include cyber attacks and data leakage which have the potential to expose information that is sensitive. In fact, emerging technologies such as Agentic AI introduce added concerns due to the fact that third-party systems may possess the ability to access databases, applications, and business workflows. Agentic systems have the capability to take actions without human intervention.

The dilemma for organizations today is therefore not only: “Is this vendor secure?” But also: “Can we as an organization trust a particular AI system to operate safely within the specific business environment?”

Build an AI vendor risk classification

“Regulators and stakeholders are certainly paying attention; they are interested in how organizations are effectively managing their third-party risk activities,” says Antonia Donaldson, Director Analyst in Gartner’s Assurance Practice.

Effective AI TPRM needs to start with visibility. Enterprises will need an exhaustive inventory of all third-party AI systems that enter the organization. Shadow AI makes this situation even more challenging, since employees may be found adopting or using publicly available AI tools without security or procurement approvals.

Therefore, AI vendors should be classified and assessed on factors including:

Business criticality

Model transparency

Regulatory exposure

Degree of autonomy

Type and sensitivity of data processed

Access to enterprise systems

Financial and reputational risks and consequences

However, it is worth mentioning that not every AI vendor will require the same level of scrutiny. For example, a generative AI tool, which is used for low-risk marketing, may not face the same level of controls as an AI system that makes credit, recruitment or fraud-detection recommendations.

Join the community of 2M+ industry professionals.

Subscribe to Newsletter to get latest insights & analysis in your inbox.

All about ETCISO industry right on your smartphone!