Third-Party AI Vendor Risk: Board Due Diligence Beyond the Contract

Here’s a question worth asking at your next vendor review, and one most companies genuinely can’t answer: the AI feature sitting inside a tool you already approved — whose model is actually running underneath it? Not the vendor’s name on the invoice. The actual model provider, several layers down, that you never contracted with, never vetted, and quite possibly never heard of. That’s a fourth party, and it’s already inside your systems.

Third-party involvement in breaches doubled year-over-year, reaching roughly 30 percent according to Verizon’s 2025 data. AI didn’t cause that shift on its own, but it’s actively widening the gap, because your vendors’ AI features process your data using models that may sit entirely outside the relationship you actually signed.

Executive Summary

Third-party AI vendor risk is the exposure a company takes on when it relies on outside vendors’ AI tools, features, or embedded models — exposure that a signed contract, on its own, does almost nothing to manage. AI adoption is now mainstream: 88 percent of businesses use AI in at least one business function, with global AI spending approaching $2 trillion in 2026. Here’s the governance problem underneath that growth. Companies remain accountable for the outputs their vendors’ AI tools generate, and for how those vendors process the business data fed into them, regardless of who actually built the underlying model. Traditional vendor risk management — a due diligence questionnaire at onboarding, filed away until renewal — was already a thin approach for ordinary software. For AI, it’s structurally inadequate, because models drift and get silently updated long after the ink on the contract has dried. A risk assessment from twelve months ago tells a board almost nothing about what that tool is actually doing today.

Quick Answers

What is third-party AI vendor risk? 

It’s the risk introduced when a company depends on an outside vendor’s AI-powered tool or feature — covering everything from data handling and output accuracy to legal accountability for decisions the AI tool influences or makes.

How is AI vendor risk different from ordinary software vendor risk?

 Traditional software behaves consistently once it’s deployed, until a company deliberately updates it. AI models can change behavior on their own over time — through retraining, silent updates, or what’s often called model drift — meaning the tool you assessed at signing may not be the tool actually running a year later.

What is “fourth-party” AI risk?

 It’s the risk carried by a model provider or sub-vendor sitting underneath your actual contracted vendor — a company you never directly engaged with, but whose model or infrastructure is quietly embedded inside the product you did sign for.

Are companies actually liable for what a vendor’s AI tool does? 

Generally, yes. Businesses are accountable for outputs generated by third-party AI tools and for how vendors handle the prompts and business data fed into them, which is precisely why boards can’t treat vendor AI risk as something a procurement team alone can fully own.

Why the Contract Was Never the Finish Line

Here’s an uncomfortable detail worth sitting with: many vendor agreements already include audit clauses giving the company the right to inspect and verify vendor practices. Procurement and legal teams routinely never invoke them. That gap existed well before AI became a factor — it’s a longstanding weakness in how vendor risk gets managed generally, where the contract gets negotiated carefully and then largely forgotten until renewal.

AI doesn’t create this gap. It makes the consequences of ignoring it considerably worse. Ordinary software, once tested and deployed, behaves the same way tomorrow as it did today, barring a deliberate update a company can review and approve. AI models don’t offer that same stability. They drift. They get retrained. Vendors push updates that change a model’s behavior without necessarily renegotiating anything contractually, because from the vendor’s perspective, it’s just normal product maintenance. From the client’s side, though, it means the risk profile of that tool can shift meaningfully without anyone on the client’s side actually noticing, let alone reassessing it.

That’s the real argument for going beyond the contract. A contract captures a moment in time — the vendor’s practices, the model’s behavior, the data flows, all as they existed at signing. AI vendor risk is a living condition, not a fixed state, and treating it like a one-time legal formality misses exactly the part that’s actually dangerous.

What Real Board Oversight Looks Like

Risk committees need to move past “did we vet this vendor” toward a sharper, more specific set of questions. Who inside the organisation actually owns re-assessment of this vendor relationship, and how often does that reassessment happen — annually, at contract renewal, or does it just… not happen? If this AI tool produces a harmful or clearly wrong output tomorrow, who has the authority to suspend the integration immediately, and who is responsible for notifying anyone affected? Does anyone in the organisation actually know which underlying model or model provider sits beneath this vendor’s product, or has that always simply been assumed to be the vendor’s problem to manage?

These aren’t legal box-ticking exercises. They’re the actual incident response chain a company needs in place before something goes wrong, not scrambled together while it’s happening. Most organisations, even sophisticated ones, don’t currently have this chain clearly defined specifically for AI vendors — it tends to get folded loosely into general vendor management processes that were designed for a much more static kind of software relationship.

A Practical Framework for Oversight Beyond the Contract

A handful of concrete practices distinguish organisations actually managing this risk from ones that believe a signed agreement has already handled it.

Build a real vendor and model inventory, including the fourth-party layer. This means mapping not just the AI vendors a company directly contracts with, but the model providers and sub-processors embedded inside those vendors’ products. Risk increasingly extends into what’s being called fourth- and fifth-tier relationships — dependencies most organisations genuinely cannot see without deliberately going looking for them.

Shift from point-in-time assessment to continuous monitoring. A questionnaire completed at onboarding is a snapshot, not an ongoing safeguard. AI systems require more intensive ongoing monitoring than traditional software specifically because of model drift, frequent updates, and evolving risk profiles — which means sustainable monitoring processes, with clear automation and escalation paths, need to replace the old annual-review cadence.

Actually exercise the audit rights already sitting in most contracts. Quarterly reviews of vendor evaluation results and incident reports aren’t optional overhead — they’re the evidence trail that proves genuine due diligence occurred if an AI output later causes real harm. An audit clause nobody invokes provides exactly the same protection as no audit clause at all.

Define the incident escalation path before an incident happens. Who contacts the vendor. Who has authority to suspend the integration. Who notifies affected customers, employees, or regulators. Deciding these questions in the middle of an active incident is how a manageable problem turns into a genuinely damaging one.

A Real-World Example

The pattern shows up clearly across sectors that rely heavily on outsourced services, mortgage lending being a particularly well-documented case. Third-party AI use is now treated as one of several dimensions financial institutions are expected to actively evaluate, and firms that underestimate their own maturity here tend to miss critical controls specifically around data sensitivity, vendor oversight, and how AI decisions or outputs get governed once they’re facing the customer. Outsourced services now touch nearly every critical function in these organisations — from loan processing and underwriting automation through to servicing transfers and document handling — and vendors that drive genuine efficiency simultaneously introduce layered risk that many institutions still manage as a one-time checklist item rather than an active, ongoing oversight obligation. Analysts covering this space have been explicit that risks extending into fourth- and fifth-tier vendor relationships mean one-time assessments and purely contract-based oversight are no longer sufficient on their own — the recommended shift is toward continuous monitoring and genuinely comprehensive vendor inventories, built on a model of shared responsibility rather than a signed document filed away and revisited only at renewal.

FAQs

Key Insights

• An audit clause nobody invokes offers exactly as much real protection as having no audit clause at all — the legal language matters far less than whether anyone actually acts on it.

• AI vendor risk is a living condition rather than a fixed state, because models drift and get silently updated in ways ordinary software licenses were never designed to account for.

•  Fourth-party exposure — the model provider sitting underneath your actual vendor — is often invisible until an organisation deliberately maps it, and most currently haven’t.

• The shift the strongest organisations are making isn’t more paperwork at signing, it’s continuous monitoring and a pre-defined incident response chain that exists before, not during, a crisis.

Key Takeaways

A signed contract with strong-sounding audit language feels like protection, and functions as almost none unless someone actually uses it. Boards overseeing companies that rely on AI-enabled vendors — which, given current adoption numbers, is nearly every board — need to push past “did we do diligence at signing” toward “do we know what this tool is doing right now, and do we know exactly who acts if it goes wrong.” Start with a real inventory that includes the fourth-party layer most companies have never mapped. Build continuous monitoring instead of an annual questionnaire. And define the incident escalation path on a calm Tuesday, not in the middle of the first real problem. The contract was always just the starting point. Treating it as the finish line is exactly the gap this risk lives in.

Stay Ahead in AI Governance

Explore emerging governance challenges and strengthen your boardroom expertise with the learning opportunities offered by Directors Institute.

Register for a Webinar with Directors Institute