Login Get a Demo
01 Checklist

AI Vendor Due Diligence Checklist

Twenty-five checks across five phases, from the first demo to the day you leave. Each one names the action, the owner and the artifact it produces, alongside what the EU AI Act, DORA and US banking guidance already expect of the buyer.

The short answer

What is AI vendor due diligence?

Last updated
Read time
10 min read
From
Cranium AI
Share on LinkedIn

AI vendor due diligence is the process a buyer runs to verify how a supplier’s AI is built, tested, governed and changed, before signing a contract and for as long as the contract runs.

It has five phases: before you buy, contract, onboarding, in-life monitoring, and renewal or exit. Each phase leaves records an auditor can inspect.

Key takeaways
  • AI vendor due diligence is a lifecycle with five phases, not a form sent once at procurement.
  • Each of the 25 checks names an owner and the artifact it leaves behind.
  • DORA has required pre-contract due diligence and a register of ICT contracts from EU financial entities since January 17, 2025.
  • EU AI Act duties for deployers of Annex III high-risk AI apply from December 2, 2027, with fines of up to EUR 15 million or 3% of turnover.
By the numbers Each figure links to its sourceAll sources 
02 The five phases

Due diligence in five phases

Most programs stop after the first phase. The risk does not: a vendor’s AI keeps changing after you approve it, so the checklist keeps going until the contract ends.

  1. Before you buy

    Find out what AI is in the product, tier the risk and collect evidence. Six checks.

  2. Contract

    Turn what you learned into terms: disclosure, notice, data use, evidence and exit. Six checks.

  3. Onboarding

    Register the vendor, configure the AI, assign oversight and test your integration. Four checks.

  4. In life

    Watch for change events, re-request evidence and report upward. Five checks.

  5. Renewal and exit

    Decide again with fresh evidence, or leave without losing data or access control. Four checks.

03 Before you buy

Before you buy: what should you check before signing?

The aim is a decision you can defend: what AI is in the product, how exposed you are and what evidence you hold. Start with the vendor AI risk questionnaire, then ask for the artifacts behind the answers.

CheckWhat to doOwnerArtifact it produces
Check 01
Inventory the vendor’s AI
Ask which models, providers and AI subprocessors process your data, and for which features. The vendor AI risk questionnaire covers this.Vendor riskCompleted questionnaire with a model and subprocessor list
Check 02
Tier the relationship
Rate the vendor by the data its AI touches and how much the AI can do on its own, not by contract value.Vendor risk with the business ownerRisk tier and rationale on the vendor record
Check 03
Classify the use case
Decide whether your intended use is high-risk under the EU AI Act, and whether you are a deployer or, by rebranding or modifying the system, a provider.Legal with AI governanceClassification memo
Check 04
Collect evidence, not assurances
Request an AI bill of materials, the latest adversarial test summary and the framework mapping.SecurityEvidence pack: AI‑BOM, test summary, framework mapping
Check 05
Check concentration
Record which model provider sits behind the vendor, and whether you already depend on that provider elsewhere.Vendor riskConcentration note in the risk assessment
Check 06
Decide and record
Approve, approve with conditions or decline, and write down the residual risk someone accepts.Business owner with the risk committeeSigned approval with conditions and risk acceptance
04 Contract

Contract: which AI terms belong in the agreement?

A promise in a sales call is not a control. Six terms make the later phases possible. The AI Supply Chain Risk Report 2026 includes a clause starter to take to counsel. This checklist is a starting point, not legal advice.

CheckWhat to doOwnerArtifact it produces
Check 07
Require disclosure
Oblige the vendor to disclose every AI model, agent and AI subprocessor used to deliver the service.LegalDisclosure clause and schedule
Check 08
Require notice of change
Oblige the vendor to give written notice before it adds or materially changes AI, with a defined notice period.LegalNotice clause with a named contact
Check 09
Restrict training on your data
Prohibit use of your data to train or fine-tune models without written consent, and bind upstream providers to the same.Legal and privacyData-use clause and data processing agreement
Check 10
Secure evidence on request
Reserve the right to an AI‑BOM in CycloneDX or SPDX and to test evidence, on request and after a material change.Security with legalEvidence and audit-rights clause
Check 11
Set incident notification
Agree a notification window for AI-related security incidents, including incidents at the vendor’s model provider.Security with legalIncident clause aligned to your response plan
Check 12
Agree exit terms
Fix termination rights, data return and deletion, and a transition period.Legal with the business ownerTermination and transition clause
05 Onboarding

Onboarding: what has to be true at go-live?

Approval on paper is not the same as a safe configuration. Before the first user logs in, the vendor’s AI belongs in your inventory, with its components recorded in an AI‑BOM where the vendor provides one.

CheckWhat to doOwnerArtifact it produces
Check 13
Register the vendor and its AI
Add the vendor, its AI features, models and subprocessors to your inventory and, where DORA applies, to your register of information.Vendor riskInventory entry and register line
Check 14
Configure before go-live
Set the tenant options: training opt-outs, retention, regions and which AI features are switched on.Business owner with securityConfiguration record with exports
Check 15
Assign human oversight
Name the people who review, override and stop the AI, and train them on its instructions for use.Business ownerNamed oversight roles and training record
Check 16
Test your integration
Run adversarial tests on the application you build on the vendor’s model before it handles sensitive work.SecurityTest report and record of fixes
06 In life

In life: how do you monitor a vendor’s AI after approval?

What you approved is a snapshot. Monitoring means noticing when the snapshot is out of date and asking again. Our article on how platforms keep vendor assessments current goes deeper.

CheckWhat to doOwnerArtifact it produces
Check 17
Watch for change events
Track model or provider swaps, new subprocessors, new data uses and new AI features, from vendor notices and your own telemetry.Vendor risk with securityChange log for each vendor
Check 18
Re-request evidence on change
When a change event fires, re-send the affected questionnaire groups and ask for an updated AI‑BOM.Vendor riskUpdated evidence pack with dates
Check 19
Review on a cadence by tier
Review high-tier vendors on a fixed schedule even when nothing has been reported.Vendor riskPeriodic review record
Check 20
Keep logs and incident records
Retain the AI logs under your control, and record incidents involving the vendor’s AI with what you reported and to whom.Security with complianceLog retention evidence and incident register
Check 21
Report upward
Give the risk committee the number of AI-using vendors with current evidence, open findings and overdue responses.AI governanceQuarterly report to the risk committee
07 Renewal and exit

Renewal and exit: how do you decide again, or leave cleanly?

Renewal is the one moment the buyer has leverage again. Use it. And plan the exit while the relationship is healthy, because that is when a vendor will help you test it.

CheckWhat to doOwnerArtifact it produces
Check 22
Re-run due diligence at renewal
Treat renewal as a new decision: refresh the questionnaire, the evidence and the tier.Vendor riskRenewal assessment
Check 23
Renegotiate the gaps
Add any contract clause the first agreement missed, using what monitoring found.LegalContract amendment
Check 24
Test the exit plan
For vendors that support critical functions, confirm you can move to another provider or in-house without disruption.Business owner with ITExit plan and test result
Check 25
Close out
On exit, have data returned or deleted, revoke access tokens and integrations, and archive the evidence trail.Security with legalDeletion certificate, revocation record and archived file
08 Regulation

What does regulation already expect of the buyer?

Three sets of rules put duties on the organization that buys or deploys, not only on the vendor. The rows below state what each text says, with links to it. For the frameworks side by side, see AI compliance frameworks, compared. This is a practitioner’s summary, not legal advice.

01

DORA does not mention AI

It covers ICT services, defined as digital and data services provided through ICT systems on an ongoing basis. That is how most vendor AI reaches a financial entity, so checks 05, 12, 13 and 24 mirror its text.

Regulation (EU) 2022/2554
02

The EU AI Act duties are for high-risk use

Articles 25 and 26 apply to high-risk AI systems. Check 03 exists because your use of a vendor’s system, not the vendor’s marketing, decides whether they apply to you.

Regulation (EU) 2024/1689
03

US bank guidance leaves GenAI to you

The April 2026 model risk guidance places generative and agentic AI outside its scope. Until the agencies say more, a bank’s own third-party program carries the weight for vendor GenAI.

OCC Bulletin 2026-13
RuleWho it applies toWhat the text expects of the buyerApplies
EU AI Act, Article 26Deployers of high-risk AI systemsUse the system according to its instructions for use, assign human oversight to people with the competence, training and authority to do it, monitor the system in operation, report risks and serious incidents, and keep its logs for at least six months. Fines reach up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher.From December 2, 2027 for Annex III systems and August 2, 2028 for Annex I
EU AI Act, Article 25A buyer that rebrands, substantially modifies or repurposes an AI system so that it is high-riskYou are treated as the provider and take on the provider’s obligations. Providers of high-risk systems must also agree in writing with third-party suppliers on the information and technical access needed to comply.Same dates as Article 26
DORA, Article 28EU financial entities that use ICT servicesYou remain fully responsible for compliance. Keep a register of information on every ICT contract. Before contracting, assess whether the service supports a critical or important function, assess the risks including concentration, and run due diligence on the provider. Hold exit strategies for critical or important functions.Since January 17, 2025
DORA, Article 30The same entities, in each ICT contractThe contract must set out the services and any subcontracting conditions, the locations where data is processed, data protection provisions, service levels, incident assistance, cooperation with authorities and termination rights. Critical or important functions add notice and reporting duties, audit and access rights, and an exit strategy with a transition period.Since January 17, 2025
US Interagency Guidance on Third-Party RelationshipsBanking organizations supervised by the Federal Reserve, FDIC and OCCUsing a third party does not diminish the bank’s responsibility. The guidance covers any business arrangement with another entity, by contract or otherwise, and describes a life cycle of planning, due diligence and selection, contract negotiation, ongoing monitoring and termination.Final June 6, 2023. A replacement was proposed in September 2026
US Revised Guidance on Model Risk ManagementBanking organizations, most relevant above $30 billion in total assetsReplaces SR 11‑7 and includes considerations for vendor and other third-party products. It states that generative AI and agentic AI models are not within its scope, and the agencies plan a request for information on banks’ use of AI.Issued April 17, 2026
09 Who does what

Who does what, and when?

Due diligence fails when everyone assumes someone else owns it. Five roles share the 25 checks. Small teams combine roles; the work stays the same.

RoleLeadsChecks
Vendor risk and procurementThe inventory, tiering, evidence requests, the change log and renewal. Owns the vendor record from first contact to exit.01, 02, 05, 13, 17, 18, 19, 22
SecurityEvidence review, integration testing, incident handling and access revocation.04, 10, 11, 16, 20, 25
Legal and privacyUse-case classification, every contract term and later amendments.03, 07, 08, 09, 12, 23
AI governance and complianceFramework mapping and reporting to the risk committee. Supports classification and log retention.21, with 03 and 20
Business ownerThe decision to buy, the configuration, human oversight and the exit plan.06, 14, 15, 24
10 Where Cranium AI fits

Where Cranium AI fits

Cranium AI supplies the AI evidence for this checklist through the five moves of the AI Trust Loop. It does not replace your TPRM suite: intake, contracts and issue workflow stay where they are. It does not carry banking regulators’ guidance as built-in frameworks; its evidence supports a bank’s own program. See Cranium AI for third-party AI risk and AI Cards.

MoveWhat Cranium AI doesChecks it supports
DiscoverThe Shadow AI view shows AI-enabled SaaS in your connected SIEM activity, and AI‑BOMs record the models and AI technologies in your own code.01, 13, 17
ObserveAI Card requests can repeat on a refresh schedule: monthly, quarterly or a custom interval. The vendor is asked for an updated card each cycle, and you can re-request sooner when a change event fires.17, 18
GovernAI Card requests name the compliance frameworks you need: the EU AI Act, NIST AI RMF and ISO 42001.04, 21
SecureArena tests the AI applications you build, including those you build on vendor models.16
ProveYou request an AI Card with frameworks, a bill of materials and documentation, and a due date. The vendor responds, and you accept it, request changes or decline it. My Vendors aggregates the cards you receive, and AI‑BOMs export as JSON, CycloneDX, SPDX and XLSX.04, 10, 18, 22
11 Keep going

Related reading.

12 Questions

Frequently asked questions.

Straight answers, kept current. Last updated .

Ask us anything →
What is AI vendor due diligence?

It is the process a buyer runs to verify how a supplier’s AI is built, tested, governed and changed, before signing a contract and for as long as it runs. It covers the models and providers behind the product, what happens to your data, the contract terms that keep you informed and the evidence you keep.

What should an AI vendor due diligence checklist include?

It should cover five phases: before you buy, contract, onboarding, in-life monitoring, and renewal or exit. Each check needs an action, an owner and an artifact, such as a completed questionnaire, an AI bill of materials, a notice-of-change clause, named human oversight roles, a change log and an exit plan. This checklist has 25 checks.

What does the EU AI Act require of companies that buy AI?

For high-risk AI systems, Article 26 requires deployers to follow the instructions for use, assign competent human oversight, monitor the system, report risks and serious incidents and keep logs for at least six months. After the 2026 Digital Omnibus these duties apply from December 2, 2027 for Annex III systems. A buyer that rebrands or substantially modifies a system can become its provider under Article 25.

What does DORA require for AI vendors?

DORA does not mention AI by name. It applies to ICT services used by EU financial entities, which is how most vendor AI is delivered. Since January 17, 2025, Article 28 has required a register of ICT contracts, due diligence before contracting and exit strategies for critical or important functions, and Article 30 lists the provisions every ICT contract must contain.

Does US banking guidance cover AI vendors?

The 2023 Interagency Guidance on Third-Party Relationships covers any business arrangement between a bank and another entity, which takes in AI vendors, and describes a life cycle from planning to termination. In September 2026 the agencies proposed guidance to replace it. The April 2026 Revised Guidance on Model Risk Management covers vendor products but states that generative AI and agentic AI models are not within its scope.

How is a due diligence checklist different from a vendor AI questionnaire?

A questionnaire is what you ask the vendor. A checklist is what your own organization does: who sends the questions, who reviews the evidence, which terms go into the contract, who watches for change and what record each step leaves. The questionnaire is one input to the first phase of the checklist.

How often should AI vendor due diligence be repeated?

Repeat the affected checks whenever a change event occurs, such as a new model, provider, AI subprocessor, data use or AI feature. Review high-tier vendors on a fixed schedule even when nothing is reported, and treat every renewal as a new decision with refreshed evidence. Calendar reviews alone leave a gap between the change and the day you notice it.

Sources

Primary sources for the figures and frameworks cited on this page. Last verified .

  1. Verizon, 2026 DBIR
  2. EU AI Act, Article 99(4)
  3. European Supervisory Authorities
13 See it for yourself

Run the checklist once.
Keep the evidence current.

See how Cranium AI collects vendor AI evidence with AI Cards and keeps it beside the AI you build.