AI Vendor Due Diligence: 7 Questions Before You Sign 

AI Security & Regulation Brief · By Nabiha Sofia Herradi · November 2026

The vendor demonstration is impressive. The AI can draft, summarize and analyze information in seconds, and the business case looks promising. Leadership is enthusiastic, procurement is ready to move forward, and everyone seems satisfied with what they’ve seen. But before approving the purchase, someone raises a question that should have been addressed much earlier: can the vendor demonstrate how our data is protected, how the AI system is governed, and who will be accountable if something goes wrong?

The response is familiar: a security certificate, a privacy policy and a presentation describing the vendor’s commitment to responsible AI. These documents may be useful, but they don’t necessarily answer the question. A security certification, for example, can provide assurance about specific controls within its scope without showing that the AI system has been assessed for accuracy, bias, transparency or operational risk.

After more than 25 years working in risk, security and compliance, I’ve learned to pay close attention to the difference between what a vendor claims and what it can substantiate. A well-written policy is a starting point, not proof that controls are working. The same applies to certifications, assessment reports and contractual assurances. What matters is whether the evidence is relevant to the system being purchased and to the way the organization intends to use it.

This distinction matters even more with AI. Organizations increasingly rely on external vendors for models, infrastructure and capabilities they don’t fully control, yet they remain responsible for the risks of their own use of those systems.

You can outsource AI capabilities. You can’t outsource accountability.

Effective AI vendor due diligence therefore requires more than a standard security questionnaire. It means understanding the technology, evaluating the vendor’s controls, assessing regulatory obligations and establishing clear responsibilities before deployment. Before approving an AI vendor, these are the seven areas I would examine, along with the evidence I would expect to see.

The 7 areas at a glance

#AreaEvidence to request
1What AI you are buying and who controls itAI system inventory, data-flow diagram, subprocessor list
2Data use and protectionData processing agreement, retention schedules, written data-use limits
3EU AI Act role and classificationClassification assessment, intended-use documentation, responsibility split
4Governance in practiceCertification scopes, risk assessments, independent assessment results
5Testing for accuracy, security and fairnessEvaluation methods and results, red-team summaries, known limitations
6Incidents and human oversightAI incident procedure, notification terms, logging, tested escalation
7Contract protectionsAI addendum, audit and information rights, exit and deletion terms
1

What AI are we actually buying, and who controls it?

Many products marketed as “AI-powered” depend on a combination of the vendor’s own software, third-party foundation models, cloud infrastructure and external APIs. These dependencies are not always visible during procurement, but they have real implications for data protection, security, resilience and regulatory compliance.

I would start by asking which models the vendor uses, which components it develops internally and which come from third parties. I would also want to understand where prompts, outputs and logs are processed and stored, who can access them, and what happens when an underlying model or service provider changes.

At a minimum, I would expect an AI system inventory, a data-flow diagram and an up-to-date subprocessor list. For systems supporting critical business processes, I would also examine the vendor’s dependency management and its arrangements for service disruption or changes in underlying models.

Red flag: A vendor can legitimately protect its intellectual property. But “it’s proprietary” is not an acceptable answer to “where does our data go?”
2

What happens to our data, and can the vendor prove it?

“Will you train your models on our data?” has become the standard question in AI procurement. It is a good question, but it covers only part of the risk. Data can be retained, logged, transferred across borders or used for model evaluation, service improvement or other purposes without ever being used for training.

A meaningful review looks at the full lifecycle: the default data-use settings, how long prompts and outputs are kept, how deletion works (including backups), whether personal data is transferred internationally and under what safeguards, and how the vendor will support obligations such as data subject requests.

The answers should be reflected in a data processing agreement and documented retention schedules, not just in a sales conversation. Where personal data is involved, organizations must also assess their own obligations under the GDPR, including whether a data protection impact assessment is required.

Red flag: Broad language such as “we use customer data to improve our services,” without clear limits, should prompt a follow-up rather than an approval.
3

What is the vendor’s role under the EU AI Act, and has it done the assessment?

The EU AI Act assigns obligations according to an organization’s role, the type of AI system or model involved and its intended use. Providers, deployers, importers, distributors and providers of general-purpose AI models carry different responsibilities.

Buyers should also understand that, under the conditions set out in Article 25, certain modifications to a high-risk AI system, or changes to an AI system’s intended purpose, can result in an organization assuming provider obligations.

I would ask the vendor to show its classification of the system and the reasoning behind it, to explain which obligations it is meeting and which remain with the customer, and to confirm what documentation it will provide over the system’s lifecycle.

Timeline to watch: under the EU AI Act’s transitional provisions, providers of AI systems generating synthetic text, images, audio or video that were placed on the market before 2 August 2026 have until 2 December 2026 to meet the marking and detectability requirements of Article 50(2). Organizations assessing affected vendors should request evidence of how these requirements are being addressed, rather than relying on a general statement of compliance.
Red flag: A claim of being “fully EU AI Act compliant,” without reference to specific provisions or evidence, tells you very little.
4

Can the vendor show that its governance works in practice?

Responsible AI principles are easy to publish; operational governance is harder to demonstrate. Frameworks and assurance mechanisms such as ISO/IEC 42001, the NIST AI Risk Management Framework, ISO/IEC 27001 and SOC 2 all have value, but they are not interchangeable.

A SOC 2 report does not establish that an AI system is reliable or fair, and an ISO/IEC 42001 certificate does not make every use of AI legally compliant. The scope of any certification is as important as the certificate itself.

I would want to know who is accountable for AI risk at executive and operational levels, how risks are assessed before deployment and monitored afterward, and which independent assessments have been completed and what they covered. Just as important, I would check whether that evidence actually supports the commitments the vendor is making, including those in the contract. An organization may have a mature information security management system while still lacking effective processes for assessing model limitations, monitoring AI performance or escalating AI-specific risks.

Red flag: Polished responsible AI principles with no evidence of how they operate in practice.
5

How does the vendor test for accuracy, security and fairness?

AI systems fail in ways that conventional software testing may not capture. They can produce confident but incorrect answers, behave inconsistently, generate discriminatory outcomes, or be manipulated through techniques such as prompt injection and data poisoning.

A mature vendor should be able to explain the metrics it uses, how it measures error and hallucination rates, whether it conducts adversarial testing, and what triggers re-testing when a model or configuration changes. Testing should also reflect the environment in which the system will operate: results obtained under controlled conditions may not represent performance once the AI is connected to organizational data, external tools or operational workflows.

The level of assurance should match the stakes. A tool that drafts internal meeting notes does not need the same scrutiny as one that screens job candidates or supports lending decisions.

Red flag: Claims of high accuracy without metrics, test conditions or known limitations should not be taken at face value.
6

What happens when the system fails, and who is accountable?

Models change, new vulnerabilities emerge, and performance can deteriorate over time. What matters is whether the vendor has the controls to identify failures early, contain their impact and respond effectively.

I would ask how the vendor defines an AI incident, how quickly it will notify customers, whether AI features can be suspended without disrupting the wider service, what logs will be available for investigation, and who can intervene. For higher-impact applications, I would also expect documented escalation procedures, clear decision-making authority and evidence that incident response arrangements have been tested.

Human oversight deserves particular attention. An approval button in a workflow is not, on its own, effective oversight; it works only when the people involved have the information, authority and practical ability to step in.

Red flag: A general cybersecurity incident policy that says nothing about harmful outputs, model manipulation or the failure of a third-party model.
7

What will the vendor commit to in writing?

Assurances given during the sales process are only as reliable as the contract that follows. Depending on the risk of the use case, an AI-specific addendum is worth considering. It would typically address restrictions on training and secondary use of customer data, notice of material changes to models or subprocessors, incident notification timelines, access to evidence and proportionate audit rights, data deletion and exit support, and the right to suspend use if risks cannot be adequately controlled.

Contractual responsibilities should also account for the vendor’s reliance on third-party models and infrastructure, particularly where those dependencies could affect service availability, data protection or compliance. Contracts do not eliminate AI risk, but they establish accountability, information rights and remedies when controls fail.

Red flag: A vendor that makes significant commitments in meetings but declines to document them is telling you something important.

How to evaluate AI vendor due diligence findings

I would not rely on a lengthy questionnaire or a single aggregate score. Start with the use case: how sensitive the data is, how much autonomy the system has, and who could be affected if it fails. Then assess each area according to the strength of the evidence.

RatingMeaningAction
GreenControls demonstrated with sufficient evidenceApprove with appropriate monitoring
AmberControls only partially demonstratedApprove conditionally, with remediation owners and deadlines
RedMaterial controls missing or unsupportedEscalate and hold approval until the risk is addressed

Not every gap is a deal-breaker; many can be corrected. But a favorable overall result should never hide a critical issue. In my experience, five issues should stop approval until they are resolved:

  • Unlawful processing of personal data
  • A prohibited AI practice
  • Unacceptable exposure of sensitive or confidential information
  • Missing essential safeguards on a high-impact system
  • A vendor’s refusal to put critical commitments in writing

The assessment should also document residual risks, the rationale for accepting them and the person authorized to make that decision. This creates an accountable record of why a vendor was approved and under what conditions.

AI vendor due diligence does not end at contract signature

An AI vendor assessment should not be treated as a one-time procurement exercise. The system approved today may not be identical to the one operating six months later. Underlying models may be replaced, new capabilities introduced, subprocessors changed and vulnerabilities discovered, and each of these can affect the original risk assessment, contractual assumptions and compliance obligations.

For this reason, organizations should establish reassessment triggers and ongoing monitoring proportionate to the system’s risk. Material changes, significant incidents, deteriorating performance and changes in intended use should all prompt additional review. The objective is to maintain confidence in the controls throughout the vendor relationship, not simply at the point of purchase.

The bottom line

The question to ask is not whether a vendor says it is compliant, but whether it can show, with relevant evidence, that its AI risks are understood, controlled and governed for the way your organization will actually use the system. A defensible AI procurement decision requires a clear understanding of the technology, reliable evidence of controls, appropriate contractual protections and continuing oversight.

That is the difference between buying AI and making a procurement decision you can defend.

Free: AI Vendor Due Diligence Checklist

A one-page checklist covering the seven assessment areas, the evidence to request, common warning signs, the five approval blockers and a decision record.

Request the checklist

Frequently asked questions

What is AI vendor due diligence?

It is the process of assessing an AI vendor’s technology, data practices, governance, testing, incident response and contractual commitments before approval, and monitoring them throughout the relationship, with the depth of review scaled to the risk of the use case.

Is a SOC 2 report or ISO 27001 certificate enough for an AI vendor?

Usually not on its own. These provide assurance about information security controls within their scope, but they do not necessarily cover AI-specific risks such as model accuracy, bias, hallucinations or changes to underlying models. Always check the scope.

Does the EU AI Act apply to organizations that buy AI rather than build it?

It can. Deployers have their own obligations, and under the conditions in Article 25, an organization that substantially modifies a high-risk system or changes a system’s intended purpose can take on provider obligations. Applicability depends on role, jurisdiction and intended use.

How often should AI vendors be reassessed?

At a frequency proportionate to risk, and whenever a trigger occurs: a model change, new capabilities, a subprocessor change, a significant incident, deteriorating performance or a change in how your organization uses the system.

NH
Nabiha Sofia HerradiFounder & Principal Advisor, AI GRC Advisory · 25+ years in risk, security and compliance · CISM · CISA · CIPP/E · CIPP/US

How AI GRC Advisory can help

We support organizations in assessing third-party AI risks, establishing AI governance processes and preparing for relevant regulatory and assurance requirements. Our work includes AI vendor risk assessments, AI system inventories, AI governance frameworks and third-party risk management aligned with ISO/IEC 42001, the NIST AI Risk Management Framework and applicable EU AI Act obligations.

Whether you are evaluating your first AI vendor or reviewing an existing portfolio of AI-enabled services, the objective is the same: informed decisions supported by evidence, clear responsibilities and appropriate risk controls.

Book a consultationWatch on YouTube

Disclaimer: General education, not legal advice. Regulatory obligations depend on the applicable jurisdiction, organizational role, intended use and relevant legal provisions. Regulatory context: November 2026.