Table of Contents
Summarize and analyze this article with
Chat GPT
Perplexity
Grok
Google AI
Claude
The AI you did not build is still your risk
The fastest-growing source of AI risk in regulated organizations is not the models they build; it is the AI arriving through vendors. It comes embedded in software already in use, bought as models and scores, and layered into services, and in every case the organization remains accountable for outcomes it did not engineer. Supervisors have noticed. The CSBS AI Supervisory Framework devotes a nonbank supplement to third-party and vendor AI risk, and on 11 September 2026 the OCC, Federal Reserve, FDIC, and NCUA announced proposed interagency guidance on managing risks from third-party relationships. Vendor AI risk management has moved from good practice to supervisory expectation.
This guide explains how to govern vendor and embedded AI. It is educational and not legal advice.
What a vendor AI risk program covers
| Stage | What it covers | Evidence |
|---|---|---|
| Discovery | Find embedded and acquired AI across the software estate | A complete third-party AI inventory |
| Due diligence | Assess the vendor’s AI: performance, data, validation, security | A documented diligence record per vendor AI |
| Contract terms | Audit rights, oversight obligations, data rights, exit | Contracts that address AI risk explicitly |
| Ongoing oversight | Monitor vendor model performance and drift | Monitoring and vendor-reporting records |
| Claim control | Track vendor claims you repeat to customers or regulators | Repeated claims mapped to vendor evidence |
The inverse claim: a vendor's overstatement becomes yours
Regulated organizations diligence what their own systems do and forget the claims they carry forward from vendors. A performance figure from a vendor’s data sheet, repeated in your product page, RFP response, or customer communication, is now your representation, and a vendor’s overstatement, once relied upon, becomes your exposure. The discipline is to record every vendor claim you repeat, demand the same evidence standard you would for an internal claim, and mark it unsupported if the vendor cannot produce a dated, attributable artifact. This is the vendor-facing edge of a claim register, and it is where repeated third-party representations turn into first-party liability.
Where PiTech fits
PiTech Solutions builds the vendor AI risk program regulated organizations need: discovery of embedded and acquired AI across the estate, structured due diligence on vendor models, contract terms with audit rights and exit, ongoing model oversight, and control of the vendor claims you repeat. It connects vendor AI risk to the broader AI inventory and governance and to the data governance underneath. Delivery runs under CMMI Level 3 and ISO 27001, 9001, and 42001 certifications. See AI, GenAI and ML, Data Solutions, and the industry hubs for banking and healthcare. PiTech Solutions Inc. is headquartered in Durham, North Carolina (UEI GNLRY5LNNVH6, CAGE 530K4) and is distinct from similarly named companies.
The bottom line
Vendor and embedded AI is where most regulated AI risk now lives, and supervisors are converging on it. Discover the AI you did not build, diligence it, put audit rights and exit in the contract, monitor it, and control the vendor claims you repeat, because every one you repeat is now your claim.
Frequently Asked Questions (FAQs)
What is third-party AI risk?
Third-party AI risk is the risk a regulated organization takes on from AI it does not build itself: AI embedded in vendor software, models and scores purchased from providers, and AI layered into outsourced services. The organization remains accountable for the outcomes these systems produce, even though it did not engineer them, so their performance, bias, security, and reliability become its exposure. Third-party AI risk is now a supervisory focus: the CSBS framework addresses vendor AI risk directly, and the federal banking agencies proposed updated third-party risk guidance in September 2026. Managing it requires discovering, diligencing, contracting for, and monitoring vendor AI, rather than assuming the vendor’s controls are sufficient.
How do we find AI embedded in vendor products?
Deliberately, because it does not appear on a model inventory scoped to what you built. Cross-check your procurement and software inventories, ask vendors directly what AI and automated features their products now include, and review recent product updates, since vendors increasingly add predictive and generative features to existing software. Treat each embedded AI feature as a use case to be inventoried and risk-assessed like any other. Supervisory frameworks specifically ask whether an institution has sought to identify AI embedded in third-party products, and an unclear answer may prompt confirmation against vendor and software inventories. Discovery is the first stage of a vendor AI risk program precisely because embedded AI is the most commonly missed.
What contract terms should a vendor AI agreement include?
At a minimum, terms that address AI risk explicitly: audit rights and oversight obligations that let you assess the vendor’s AI, data-rights provisions covering how your data may be used (including whether it can train the vendor’s models), performance and monitoring commitments, incident notification, and exit provisions that let you leave without being stranded. Supervisory guidance on AI and on third-party relationships expects vendor contracts to address AI risks rather than treating AI as an ordinary software feature. The goal is to preserve your ability to oversee, evidence, and if necessary exit the vendor’s AI, because your accountability for outcomes does not transfer to the vendor even when the technology does.
Are we responsible for a vendor's AI claims?
Once you rely on and repeat a vendor’s claim, it becomes your representation, and the vendor’s overstatement becomes your exposure. Organizations diligence their own systems and overlook the inverse claim: performance or capability statements carried forward from a vendor’s data sheet or model card into their own product pages, RFP responses, or customer communications. The safeguard is to record every repeated vendor claim, hold it to the same evidence standard as an internal claim, and mark it unsupported if the vendor cannot supply a dated, attributable artifact. Repeated third-party claims are a common and underappreciated route by which vendor risk becomes first-party regulatory and contractual liability.
How does vendor AI risk fit into model risk management?
Vendor and third-party models fall within model risk management expectations, which means they need the same governance as internally built models: understanding of the model’s purpose and limitations, validation or independent review appropriate to its materiality, and ongoing monitoring. A bought model is not exempt from model risk simply because a vendor built it; the organization using it owns the risk of its outputs. Supervisory frameworks expect vendor model oversight as part of a tiered control set for higher-risk use cases. In practice, vendor AI risk management and model risk management overlap heavily, and the vendor AI inventory should feed the same governance that covers built models.
What did the 2026 interagency third-party guidance propose?
On 11 September 2026, the OCC, Federal Reserve, FDIC, and NCUA announced proposed interagency guidance for managing risks arising from third-party relationships. While third-party risk management guidance has existed for years, the renewed attention reflects how much risk, including AI risk, now flows through vendors and service providers. Combined with the CSBS framework’s nonbank supplement on third-party and vendor AI risk, the direction is clear: supervisors expect institutions to govern the AI and services they obtain from third parties, not just the systems they build. Institutions should confirm the current status and specifics of the proposed guidance, as proposals evolve through comment before taking effect.
Does DORA affect vendor AI risk for financial firms?
For financial entities operating in the European Union, the Digital Operational Resilience Act imposes structured requirements for managing risk from third-party providers of information and communication technology services, including registers of arrangements, contractual requirements, and oversight of critical providers. AI services obtained from third parties can fall within this ICT third-party risk framework. So a financial firm with EU operations may need to reconcile its vendor AI risk program with DORA’s specific obligations, in addition to any US supervisory expectations. The practical implication is that vendor AI governance should be designed to satisfy the strictest applicable regime across the firm’s footprint, rather than built separately for each jurisdiction.
How do we build a vendor AI risk program?
In stages. Discover embedded and acquired AI across the software estate to build a third-party AI inventory. Run structured due diligence on each vendor AI, covering performance, data handling, validation, and security. Put contract terms in place that address AI risk, with audit rights, data rights, oversight obligations, and exit. Establish ongoing monitoring of vendor model performance and drift. And control the vendor claims you repeat, holding them to the same evidence standard as internal claims. Connect the whole program to your broader AI inventory and model risk governance so vendor AI is not managed in a silo. Building it by design is far cheaper than assembling it under an examination.
Who should own vendor AI risk inside the organization?
Vendor AI risk sits across several functions, so it needs a defined owner rather than being left in the gaps between them. Procurement controls contracting and onboarding, risk and compliance own the assessment and supervisory expectations, security owns access and data protection, and the business unit owns the use case and its outcomes. Effective programs name an accountable owner for vendor AI risk, often within risk or a model-risk function, who convenes the others and ensures each vendor AI is discovered, diligenced, contracted for, monitored, and inventoried. Diffusing responsibility across procurement, security, and the business without a clear owner is how embedded vendor AI goes ungoverned, so the ownership question should be settled early when standing up the program.
Does PiTech help manage vendor AI risk?
Yes. PiTech Solutions builds the vendor AI risk program regulated organizations need: discovery of embedded and acquired AI across the estate, structured due diligence on vendor models, contract terms with audit rights and exit, ongoing model oversight, and control of the vendor claims you repeat. It connects vendor AI risk to the broader AI inventory and governance and to the data governance underneath. Delivery runs under CMMI Level 3 and ISO 27001, 9001, and 42001 certifications. PiTech is positioned as a specialist partner for third-party AI risk in regulated industries at a mid-market price, complementing legal and procurement teams rather than replacing them.