Check out our upcoming events | Learn More
Articles

Third-party vendor risk management in an AI world (part 2)

Build an AI vendor-risk program all at once and it fails. Build it in phases, proportionate to the risk, and it survives.

  • August 14, 2026
  • 6 Minute Read
Back to All Insights

Part 1 made the case that AI breaks how vendor risk usually works. The vendor you assess keeps changing after you sign, and the real risk often sits a layer below it, with a provider you never evaluated. This piece is about the program you build in response, and why it starts with your third parties and their ever changing risk profile

AI is entering most organizations faster than governance can track it; not through a single procurement decision, but through AI features switched on in tools you already own, point solutions adopted by individual teams, and vendors quietly routing your data through models you never evaluated. For the majority of enterprises, third parties are where AI actually shows up, which makes third-party risk management (TPRM) the natural place to manage it. 

But TPRM has an uncomfortable gap. It was built to assess a vendor’s financial stability, security posture, and data handling, and never for how an AI model behaves, what data it was trained on, or the ways an AI feature can drift after you’ve signed the contract.

The instinct is to close that gap with a comprehensive new process. That instinct is usually what kills the effort. An AI TPRM program lives or dies not on how sophisticated its framework is but on whether the business will actually route decisions through it. 

The rest of this piece lays out the stages of standing one up, but  they only work if three principles govern every choice along the way.


Principles that decide whether your TPRM program survives

Proportionality. Match the depth of review to the risk. The program dies if it’s pure friction, so give low-risk tools light, fast checks and reserve deep review for the vendors that genuinely warrant it. A program the business trusts to move fast on low-risk tools earns the right to be slow and demanding on high-risk ones.

Early, visible wins. Show value before the program looks like a tax. Governance, risk, and compliance (GRC) functions lose their funding the moment they’re seen as overhead, So, bank a concrete win early, whether it’s a real risk you caught, or a deal that cleared faster because the AI review path was ready. A story about what the program made possible protects it far better than a story about what it blocked.

Crawl, walk, run. Don’t try to build the whole program at once. Begin with discovery and a manual questionnaire for your highest-tier vendors, prove the model works, then layer in tooling, monitoring, and certification over time. Attempting to stand up the mature end-state program on day one is the single most common way these efforts collapse under their own weight.

With those three principles in place, the build itself runs in six phases.

Six phases, in the order that survives

Phase 1 — Visibility before governance

You can’t manage what you can’t see, and with AI your inventory is already wrong because of the two dynamics from before: shadow AI (teams adopting tools directly) and existing vendors switching AI features on. 

So the first build is a repeatable discovery mechanism, not a one-time census. Practically speaking, reconcile procurement records, SSO/OAuth grants, expense data, and network/CASB signals to find where AI is actually in use. Then make discovery ongoing; the census you take today decays within a quarter. 

This phase also gives you the evidence you’ll need to justify resourcing to leadership (“we found 40 AI tools processing data, we assessed six of them”).

Phase 2 — Decide who owns it

This is the organizational problem that determines whether the program survives. AI vendor risk has no natural home; security, privacy, legal, procurement, and data governance each own a piece and none owns the whole. If you don’t resolve this early you get gaps and turf friction.

You don’t need a new department but you do need a named accountable owner and a cross-functional review path (often a lightweight AI risk council or an extension of an existing vendor/security review board) with a clear RACI. The person who owns it needs enough authority to hold a procurement gate. Pick the owner deliberately rather than letting it default to whoever complains loudest.

Phase 3 — Branch your existing intake

Add a trigger question to your standard vendor intake: does this vendor use AI, or process our data through a model? A “yes” routes to an AI-specific questionnaire module and an AI-aware risk tier; a “no” stays on your normal flow. 

Tier on use case and data sensitivity, not just the vendor, since the same vendor can be trivial or high-stakes depending on what it touches. Keep the AI module proportionate, a short screening set that escalates to deep diligence only for higher tiers. A 200-question AI assessment applied to everything guarantees the business goes around you.

Phase 4 — Adopt a framework as scaffolding, not the work

Anchor to NIST AI RMF and/or ISO/IEC 42001 so you’re not inventing a taxonomy and so you have something defensible to point auditors and leadership at. 

But treat them as scaffolding.  They tell you the categories, not how to run your process. Map your questionnaire and controls back to the framework so everything is traceable, then move on. Industry frameworks make the process predictable for the people running it today and the people who join later.

Phase 5 — Build the monitoring engine (the hard, sustained part)

Point-in-time assessment is easy; the continuous piece is where programs stall. Operationalize the reassessment triggers from before (model-provider change, training-use change, new region, added agentic capability, material model update) into something that actually fires. 

In practice this means contractual notification rights (so a trigger is even visible to you), a defined cadence for higher-tier vendors, and monitoring for the two silent-change problems: a vendor swapping its model provider, and a model updating underneath a workflow you already validated. This is genuinely the least mature area across the industry, so plan for it to be iterative rather than solved at launch.

Phase 6 — Close the skills gap

Under-appreciated and often the real bottleneck. Your TPRM analysts are excellent at SOC 2 scoping and DPAs, but reviewing a model card, judging whether “we don’t train on your data” actually holds, or spotting the absence of prompt-injection testing requires knowledge you can’t assume they have. Options in rough order of velocity; build a review playbook with worked examples and escalation criteria, embed a security/ML-literate reviewer into the review path for high-tier cases, and upskill the core analysts over time. The playbook matters most by letting ordinary analysts handle the routine 80% and flag the rest.

Read as a sequence, the six phases share a logic; see what you actually have, decide who’s accountable for it, route it through a process that scales to the risk, anchor that process to something defensible, keep watching after the contract is signed, and make sure the people doing the work are equipped to do it. Visibility comes before governance because you cannot tier, assess, or monitor what you haven’t found. Ownership comes before tooling because the sharpest questionnaire in the world fails if no one has the authority to hold the gate.

None of this has to launch at once. The phases are a maturity path, not a checklist to clear before go-live. A credible starting position is Phase 1 discovery plus a manual questionnaire for your highest-tier vendors, just enough to catch real risk and to show leadership what’s out there. Continuous monitoring and framework certification can follow once the core is proven. The organizations that stumble are almost always the ones that tried to stand up the full apparatus on day one and collapsed under the weight of their own process.

That is the through-line back to where we started. Keep the program proportionate, so the business routes decisions through it rather than around it. Bank early, visible wins, so the program is remembered as an enabler rather than a tax. And crawl before you walk, so the thing survives long enough to mature. A program that is trusted, funded, and still standing in two years will manage far more real AI risk than a comprehensive one that the business quietly abandoned in its first quarter.

Most enterprises meet AI through their vendors, but traditional third-party risk management was never built for how AI models behave or drift after you sign. Part 2 lays out a six-phase, risk-proportionate way to build an AI vendor risk program the business actually uses, instead of a comprehensive one it quietly abandons in the first quarter.