AI Third-Party Risk Management: Vendors, Models, and Embedded AI

Vendor risk programs know how to assess a company that hosts your data or runs a service. AI strains them in three ways: model providers whose product changes under you, AI application vendors whose subprocessors include model providers you never contracted with, and AI features that appear inside software you already license, enabled by an update nobody reviewed. This guide covers extending the program to all three, drawing on FISTA Solutions' AI enablement practice. The operational side is in how to manage ai vendors and the questionnaire in the ai vendor security questionnaire.

What AI third parties belong in the program?

Type	Examples	Distinctive risk
Model providers	Hosted LLM and embedding APIs	Behavior changes; terms; retention; concentration
AI application vendors	AI-native products processing your data	Subprocessors; data handling; decision influence
Embedded AI features	AI added to CRM, HR, productivity, security tools	Arrives unreviewed; default-on; data to providers
Build partners	Consultancies and augmentation providers	Access to data and systems; IP; delivery quality
Subprocessors	Model providers and clouds behind vendors	Invisible without flow-down

Supply chain components beyond vendors are in ai supply chain security.

How do you find embedded AI features?

Survey vendor release notes and renewals for AI features; ask vendors directly what AI they have added, what data it uses, which model providers it sends data to, and whether it is on by default; add AI questions to renewal reviews; and monitor network egress for new model provider endpoints. Embedded features are the fastest-growing and least-controlled category in most inventories. Data leakage through them is covered in ai data leakage prevention.

How should AI third parties be tiered?

By the sensitivity of data they receive, the impact of decisions they make or influence, the degree of dependence, and regulatory coverage. A model provider receiving customer records for a servicing assistant is high tier; a drafting tool with no customer data is low. Tier sets due diligence depth, contract requirements, and monitoring intensity, consistent with the AI inventory's system tiers. Tiering is in ai model risk management.

What should due diligence cover?

Security certifications and practices; data handling including training use, retention, residency, encryption, and subprocessors; model governance including how the vendor evaluates, changes, and monitors models and handles incidents; explainability and oversight support for decision-influencing features; compliance attestations for applicable rules; and financial and operational stability. For build partners, add vetting, IP, and delivery practices. The full checklist is in the AI vendor due diligence whitepaper and the partner variant in how to choose an outsourcing partner.

What contract terms matter?

Prohibition on training with your data; retention and deletion; subprocessor disclosure and flow-down of obligations; notice periods for model changes and deprecations; security obligations and breach notification; audit rights; service levels including quality metrics where measurable; regulatory terms such as business associate or service provider clauses where required; and exit terms with data return and transition support. Verify that terms apply to the exact service tier in use. Contract practice is in how to negotiate an ai development contract.

How should AI third parties be monitored?

Continuously: quality on your own evaluation set for model providers and decision-influencing features, change notices and their measured effects, incident and breach notifications, cost against budget, configuration drift in embedded features, and periodic reassessment of certifications and terms. Uptime dashboards do not show quality regressions or silent model changes. Monitoring mechanics are in how to manage ai vendors.

What is concentration risk and how is it managed?

Dependence on one model provider, cloud, or vendor across many systems, so that an outage, price change, policy change, or exit affects the organization broadly. Manage it through a gateway that keeps providers interchangeable, tested fallbacks, evaluation that makes switching safe, contract exit terms, and periodic review of the dependency map in the risk register. Failover practice is in what is a fallback model and the register in ai risk register.

How does this connect to sector obligations?

Regulated organizations must bring AI third parties inside their service provider and business associate programs, with the sector's contract terms and oversight expectations. The financial services variant is in ai and glba compliance and the healthcare variant in ai and hipaa business associate agreements.

What mistakes are common?

Model providers treated as utilities outside the program; embedded AI features never inventoried; subprocessors invisible because flow-down was not required; due diligence done once at signing; monitoring limited to uptime; and no concentration view, so a single provider incident surprises the board.

What does sound practice look like?

An organization inventories its AI third parties including embedded features found through vendor surveys and egress monitoring, tiers them, completes due diligence proportionate to tier, contracts for training prohibitions, change notice, and exit, monitors quality on its own evaluation and configuration drift in embedded features, and maintains a concentration map reviewed quarterly. When a productivity suite enables a default-on AI feature sending data to a new provider, the review process catches it at the release note, and the feature is configured before it reaches users.

How FISTA Solutions helps with AI third-party risk

FISTA Solutions helps clients inventory and tier AI third parties including embedded features, run due diligence, structure contracts, and build the gateway, evaluation, and monitoring that make ongoing oversight and concentration management practical. The AI enablement practice leads governance and platform, AI agents run on managed providers, and forward deployed engineers embed with client risk and procurement teams. The record behind the approach is 150+ projects for 50+ companies.

To bring every AI third party inside your risk program, message FISTA on WhatsApp, or read the ai vendor security questionnaire for the questions due diligence should ask.