Blog

AI Third Party Risk Management: A Guide for Leaders

How leaders assess AI third party risk: mapping the supplier chain, EU AI Act buyer deadlines, evidence to demand, hidden fourth parties and ongoing monitoring.

· By

A glass wall with five rows of coloured sticky notes linked by string into a descending chain, with a printed dependency diagram and contracts in the foreground.
An AI supplier is a chain of model provider, cloud host and subprocessors, not a single party under one contract.Image: IamVera.ai — original editorial illustration

Someone on your team has already switched on an AI feature inside a tool you pay for. It drafts faster, nobody objected, and no one has checked where the documents go once they leave the browser tab. Then a board paper asks a plain question: which supplier did we approve, when, and what do we hold on file to prove it? The answer is harder than it used to be, because the supplier isn't one party. The assistant on your colleague's screen sits on a model from one company, runs in a cloud tenancy from another, calls a plug-in maintained by volunteers, and may send material to a review team you've never met. An approval signed nine months ago can describe a model that no longer exists in that form. Treat an AI supplier as a moving dependency chain rather than a one-time purchase, and the practical questions change.

Treat an AI supplier as a moving dependency chain rather than a one-time purchase and the practical questions change. What evidence do you hold, who owns each link, and how would you know if something underneath moved?

What Makes AI Third Party Risk Different From Vendor Risk?

AI third party risk management is the process of identifying, assessing, monitoring and documenting the risks an organisation takes on when it relies on external AI models, providers, datasets, plug-ins and hosting. Two things separate it from ordinary software vendor review. First, the supplier's output can change without any change to your own code, because weights, training data, API behaviour and hosting all sit outside your version control. Second, your prompts and documents may reach parties you never contracted with and cannot audit. A contract with the vendor at the top of the chain binds that vendor only, not the model provider, the cloud host or the labelling team underneath it.

The dependency is a chain, not a single supplier.

Where the chain actually runs

A single assistant can sit on five layers, and each one fails differently. The table below is the reference list for the rest of this guide.

  • Model provider — What It Supplies: Model or API output; Exposure If It Fails or Changes: Behaviour shifts after a version or weight change
  • Cloud or hosting provider — What It Supplies: Compute and storage for prompts, documents and logs; Exposure If It Fails or Changes: Data location, outages and unclear retention
  • Subprocessor — What It Supplies: Processing of prompts, documents or embeddings; Exposure If It Fails or Changes: Unknown transfers and retention periods
  • Plug-in or open-source component — What It Supplies: Tools, libraries and connectors; Exposure If It Fails or Changes: Compromised or unmaintained dependencies
  • Human review vendor — What It Supplies: Labelling or review of outputs; Exposure If It Fails or Changes: Confidentiality of material shared for review

The financial case for mapping that chain is older than the current model cycle, so it should be quoted with its date attached. IBM's 2024 Cost of a Data Breach research put the average cost of a data breach at USD 4.88 million, up from USD 4.45 million in 2023, a 10% increase and the highest rise since the pandemic.

The same research found that 40% of breaches involved data distributed across multiple environments, which is precisely the condition external APIs, managed services and interconnected suppliers create by default. Read it for what it is: a general breach-cost baseline dated 2024, not a measurement of AI vendors. It explains why the control question is broader than a security certificate. You need to know where prompts, documents, embeddings, logs and outputs are processed, which subprocessors receive them, how long they're retained, whether your data trains a model, and whether access can be revoked and data deleted.

The test that follows takes about ten minutes with your own team. List every layer your AI use depends on, including the ones nobody signed a contract with, and write a name next to each. Where a layer has no name and no evidence behind it, you've found the first gap.

What Does the EU AI Act Change for Buyers?

The Act converts vendor due diligence from a procurement questionnaire into an evidence requirement, and it does so on a staged timetable. For a buyer, the timetable matters as much as the duties, because two suppliers can be subject to different deadlines on the same day. The dates below come from the EDPS timeline of the EU AI Act, which tracks when each group of obligations applies.

  • 1 August 2024 — What Applies: AI Act enters into force; What Buyers Should Record: Confirm the supplier's EU exposure and applicable obligations
  • 2 February 2025 — What Applies: Prohibited practices and AI-literacy obligations apply; What Buyers Should Record: Confirm the use case is not prohibited
  • 2 August 2025 — What Applies: General-purpose AI model obligations apply; What Buyers Should Record: Provider classification of the model
  • 2 August 2026 — What Applies: Commission enforcement powers for GPAI obligations begin; What Buyers Should Record: Technical documentation and information on capabilities and limitations
  • 2 August 2027 — What Applies: GPAI models placed on the market before 2 August 2025 must comply; What Buyers Should Record: Applicable transition deadline for each model

What each date asks of a buyer

The Act entered into force on 1 August 2024 with a risk-based structure that reaches along the value chain, from providers to deployers, downstream integrators and users of general-purpose models. The first live date was 2 February 2025, when prohibited practices and AI-literacy obligations took effect. Before any deeper review, confirm that the intended use isn't one of the prohibited practices, because no volume of vendor evidence cures a use case the law doesn't permit.

2 August 2025 is the date that changed the supplier conversation. Obligations for general-purpose AI models became applicable then, and the European Commission's AI regulatory framework page sets out the transition: providers of GPAI models placed on the EU market before that date must comply by 2 August 2027. The Commission's enforcement powers for GPAI obligations begin on 2 August 2026, covering transparency, copyright, and security and safety requirements for the most advanced models. The high-risk tier arrives after that, which gives buyers a window rather than a pass.

The consequence is documentary. A questionnaire asking whether a supplier "complies with the AI Act" produces nothing you can put in front of a board or a supervisor. The Act's structure implies specific artefacts: model documentation, testing, traceability, human oversight, cybersecurity, monitoring and incident escalation across the supplier lifecycle. If your team is working through what an approved model does and doesn't settle, the analysis in what buying a compliant AI model does not cover follows the same logic onto the deployment side.

The systemic-risk question most questionnaires miss

Providers of the most advanced models carry additional duties. Commission guidance for general-purpose AI providers sets out that providers who know they're developing a model with systemic risk must notify the AI Office, assess and mitigate those systemic risks, and meet security and safety requirements addressing large-scale harms. The examples given include chemical, biological, radiological and nuclear risks, loss of control, cyber-offence and harmful manipulation.

A questionnaire that stops at accuracy and uptime won't surface any of that. Ask whether the supplied model carries a systemic-risk classification, what assessment and mitigation the provider performed, what security evidence exists, and how material changes or newly identified risks reach customers.

Keep four records per supplier: market-entry date, provider classification, applicable transition deadline and systemic-risk status. All of this is EU law, other jurisdictions set different duties, and whether a given obligation reaches you depends on your role and use case. Those four records are still worth holding, because they're the first thing any supervisor asks for.

Which AI Uses Need the Strictest Review?

Review depth follows the use case and the model's classification, not one corporate threshold applied evenly to every supplier. A drafting assistant working on public material and a system holding medical records shouldn't face the same test, and treating them identically spends the strictest scrutiny on the use that needs it least.

  • Internal drafting on non-sensitive material. A standard questionnaire, a current subprocessor list and a dated version record are proportionate. Keep the questionnaire short enough that suppliers answer it fully.
  • Customer-facing content, where errors reach the public. Add extended adversarial testing and a named reviewer for anything published, alongside the standard pack.
  • Confidential legal, medical or personnel records. Add data-flow analysis with contractual audit rights, so you know where the material is processed, which subprocessors receive it and how long it's retained.
  • Decisions affecting a person's rights, money or access. Require independent human review before anyone relies on the output, plus the supplier's evidence behind any risk classification it claims.

Who carries the duty to review

Most supplier terms move review duties onto the customer, and the wording is worth reading closely before you sign. Our own terms state the position plainly: customers must review Output before relying on it, acting on it, filing it or sending it to a third party. That's a fair example of the clause to look for. If a supplier's terms place a review duty on you, your internal process has to name who performs that review and how it's recorded.

Separate the three layers of this argument cleanly. The statutory timetable comes from EU legislation. The judgement that a high-impact use should carry a written, named review duty is our reading of how accountability survives an audit. How far such duties are enforceable, and on whom, varies by jurisdiction and by the role you occupy in the chain.

What this means for approval: the more an output affects a person's rights, money or access, the more evidence the review needs before anyone relies on it.

Which Evidence Should You Demand Before Signing?

Ask for a named document, a test result or a contract clause against each risk area, tied to the specific use case, and record what arrived and what didn't. Where a supplier replies with a narrative about its commitment to safety, mark the row open rather than satisfied.

  • Data location and retention — Evidence to Request: Processing locations, retention periods and deletion process; Why It Matters: Shows where confidential input sits and for how long
  • Training use — Evidence to Request: Written statement on whether customer data is used to train models; Why It Matters: Prevents silent reuse of sensitive material
  • Subprocessors — Evidence to Request: Named list of subprocessors with change notice; Why It Matters: Reveals the hidden part of the chain
  • Security testing — Evidence to Request: Dependency scans, signed releases and adversarial test results; Why It Matters: Tests prompt injection, tampering and unsafe tool use
  • Availability — Evidence to Request: Service-level objectives, fallback options and rate-limit test results; Why It Matters: Measures exposure to outages and abuse
  • Incident handling — Evidence to Request: Notification deadline and named contact route; Why It Matters: Sets the clock for your own response
  • Exit — Evidence to Request: Data return and deletion commitments on termination; Why It Matters: Allows the organisation to leave safely

Why each row exists

The threats are not ordinary software bugs. NIST's AI risk guidance identifies poisoning, malicious inputs or outputs, denial of service from adversarial prompts, unauthorised information disclosure, model-weight theft, compromised dependencies and misconfigured data pipelines as relevant to AI systems. Each maps to a specific request rather than a general assurance. Dependency and artefact scanning covers compromised components. Signed releases and provenance records make unauthorised changes detectable. Adversarial testing exposes prompt-injection and unsafe tool-use paths. Data-flow analysis shows whether confidential inputs are retained, transferred or used for training. Rate-limit and fallback tests measure availability under abuse or provider outage.

The contract carries the rest. Request incident-notification deadlines, audit and evidence rights, data-location and deletion requirements, a subprocessor list, training-use restrictions, version-change notice and exit assistance. Set the depth by the tier the use case sits in, since a public marketing assistant and a system holding medical records shouldn't face identical tests.

In our analysis, disclosure and assurance get confused constantly. A supplier naming its model provider has given you transparency. It hasn't shown that outputs are accurate, that data handling is safe, or that incidents get reported on time.

What an audit report proves

An audit against an international standard is useful evidence and not proof. The scope of the audit, the date, the version audited and the exceptions all matter. If you want to know how to word the inspection clause that makes such evidence obtainable, the structure of audit rights in AI contracts covers the wording.

A row without evidence stays an open question. It never becomes a green tick because a renewal date arrived.

How Do Hidden Fourth Parties Change Your Risk?

A trusted direct supplier can inherit material risk from parties you can't audit or contract with. Foundation-model providers, cloud infrastructure, data brokers, external evaluators, plug-ins and human review vendors all sit behind the vendor whose name is on your invoice, and the first section's layer table is the list to work from.

The gap between awareness and action is measurable. Venminder's 2025 survey found that 58% of organisations review third parties' risk-management programmes for fourth-party oversight, while 25% focus only on critical or high-risk fourth parties. Retain the 16% figure only if the cited survey page explicitly identifies it as the share using tools to monitor external posture. Most programmes have the policy and not the practice.

A playbook for chains you can't audit

  1. Name each upstream party. Ask the direct supplier to identify every model, host and subprocessor in the chain, in writing and by role.
  2. Rank concentration. If four suppliers run on the same model and the same cloud region, you have one dependency with four invoices, and an outage is a correlated event rather than an isolated one.
  3. Set notification clauses that cover the whole chain. A notice duty that stops at the direct supplier's own changes leaves the largest risk unmonitored.
  4. Require a contingency option. A second provider, a cached local model, or a documented manual procedure for high-impact work.
  5. Keep human review as the fallback. Where the chain can't be verified, a person has to own the decision before it takes effect.

https://www.youtube.com/embed/Up27fcrhc1E

When disclosure is refused outright, the decision is still yours to make, and it should be recorded. A business can prohibit a use case because the chain isn't auditable, and the reasoning belongs in the risk file with a date and an owner. The mechanics of naming and tracking those parties are set out in how subprocessors extend the AI chain, which is the clause set to require before signing.

How Do You Monitor an AI Supplier After Approval?

Approval is where monitoring starts, not where it stops. A vendor can change model weights, training data, API behaviour, dependencies or hosting after approval, and your application code won't move an inch. Without a routine that detects those changes, the assessment you signed describes a system that no longer exists.

The lifecycle below is the routine in one view.

The routine, step by step

  1. Keep a versioned AI bill of materials. List each model, provider version, data source, plug-in and hosting location. NIST's work on third-party AI risk recommends an inventory of third-party materials, including foundation models, open-source software, proprietary data and hardware. Without the inventory, nothing else in this routine is possible, because there's no baseline to compare against.
  2. Attach evidence to each entry. Model card, contract terms, test results and deployment configuration sit beside the dependency, not in an inbox. When someone asks why a tool was approved, the answer is one row away.
  3. Set change triggers. A version-change notice, a new subprocessor, a new hosting region or revised terms each trigger a reassessment. Treat these as signals, not housekeeping.
  4. Re-run regression tests after each trigger. Cover accuracy, harmful outputs, prompt injection, data leakage, availability and policy compliance. Testing the same six dimensions every time is what turns a change into a comparison.
  5. Record the run itself. Model version, provider, retrieved sources, reviewer outputs and cost belong in the record. A recorded run makes changes easier to detect, investigate and, where a rollback or fallback exists, reverse., and it distinguishes drift from noise when behaviour shifts.
  6. Review on a fixed schedule and set the exit point. NIST also points to provenance mechanisms, documented fallback options and decommissioning when risk exceeds organisational tolerance. Write down in advance what would make you leave.

Practical rule: if a change can't be seen in the inventory, it can't be assessed, and if it can't be assessed, the approval no longer describes the system you're running.

Here's a question you can answer before the end of the week. Which entry in your inventory has changed since the last review, and who was told? If nobody can answer, you have a document rather than a control.

Who Owns Third-Party AI Risk and What Goes in the Risk File?

One named executive owns each material supplier. Alongside that, name who approves exceptions, who watches for changes and who reports to the board. Shared ownership produces a file nobody maintains, and the failure surfaces at the worst moment, when a supervisor or a customer asks a specific question and the answer lives in three people's memories.

NIST's AI Risk Management Framework treats third-party AI risk as an explicit governance responsibility rather than a vendor onboarding task. Govern 6.1 requires policies and procedures addressing AI risks associated with third-party entities, including infringement of third-party intellectual property or other rights. Manage 3.1 requires organisations to regularly monitor risks and benefits arising from third-party resources and to apply and document risk controls.

The file in four parts

  • Dependencies identified. The external software, models, data and other supply-chain relationships you rely on, at every layer.
  • Controls defined. What was required of each dependency, and on whose authority.
  • Monitoring schedule. When each entry is reviewed, and what triggers an off-cycle review.
  • Evidence retained. The documents and test results that show the controls operated, not just that they were designed.

Under Manage 3.1 the file has to stay current rather than be archived at approval, which means it's a working record maintained by people. That matters for how you build it. Crowe's 2025 TPRM Benchmark Study reported that approximately 63% of respondents assessed third parties for AI or machine-learning-related risks. Retain the 71% figure and the financial-services qualification only if the cited study explicitly supports both details. A file that only works with new software is a file most organisations can't keep.

The board test is simple. For each supplier, can the file show who owns the risk, what changed last, and what evidence supports the current position? If any of the three takes more than a minute to produce, the file isn't doing its job.

What Should You Decide Before Approving an AI Supplier?

Three outcomes cover most decisions, and each one should be written down. Approve with conditions when the evidence table is complete except for items the supplier can still fix, such as a subprocessor list or a version-change notice, and set a dated deadline for each open item. Choose a fallback when the supplier is sound but could fail, such as a second provider or a manual procedure for high-impact work. Refuse when the chain can't be audited and the use involves confidential or high-impact data, and record the reasoning so the decision can be reviewed rather than relitigated.

Five questions to put to the supplier this week:

  • Where do prompts, documents and outputs get processed, and for how long? Ask for locations and retention periods, not reassurance.
  • Is customer data used to train models? Get it in writing, and check it against the contract.
  • Which subprocessors and models sit behind the service? A named list with a change-notice duty.
  • How and when are version changes notified? Notice before the change takes effect is worth more than a changelog afterwards.
  • Do external providers receive only context released after pseudonymisation and outbound verification? Here a privacy architecture becomes inspectable rather than aspirational.

On the last question, our design is a benchmark you can hold other suppliers to. That is a control description, not a legal guarantee of GDPR compliance, and judgement plus the final decision remain with the user.

Compare the supplier's answers against the evidence table before the next renewal. Where a row is still empty, you know precisely what to request, and from whom.

You can review real runs, configurations and recorded model events instead of accepting a questionnaire at face value.

More articles on this subject are collected in the AI governance overview.

Sources and references

  1. EDPS timeline of the EU AI ActEuropese Unie
  2. European Commission's AI regulatory framework pageShaping Europe’s digital future · 2026-11-15
  3. Guidelines for providers of general-purpose AI modelsShaping Europe’s digital future
  4. Venminder's 2025 State of Third Party Risk Management surveyVenminder
  5. Crowe's 2025 TPRM Benchmark StudyCrowe
  6. AI RMF Core - AIRCNIST SI Resource Center
  7. NIST's work on third-party AI riskNIST
  8. NIST's AI Risk Management FrameworkNIST

Sources: EU AI Act dates are attributed to the EDPS timeline and the European Commission's regulatory framework and GPAI guidelines; breach-cost figures to IBM's 2024 Cost of a Data Breach research; TPRM practice to Venminder and Crowe surveys and NIST's AI RMF and third-party AI risk work.

← All articles in this topic ← All articles