Checklist · AI & Privacy · Vendor Diligence

AI Vendor Due-Diligence Questionnaire

Questions to put to an AI vendor before you buy, renew, or switch on a new feature, and a way to tell the answers you can rely on from the ones you cannot.

By Lincoln Square Law · 12-minute read

🖨 Print this checklist

Most AI products reach an organization through a purchase order, a renewal, or a feature switched on inside software the business already licenses. By the time Legal sees the tool, the contract is often signed and data is already flowing to the vendor. This questionnaire is meant to run before that point. It asks the vendor for the facts that decide which laws apply, what the contract must say, and what evidence you will need if anyone later asks how the tool was chosen.

The questions follow the order in which a review usually unfolds: the use case, the data, the evidence behind the vendor’s claims, the legal triggers, oversight and security, the contract, and the decision record. They track the NIST AI Risk Management Framework, a voluntary framework that calls for policies on AI risks from third-party entities and contingency processes for failures in high-risk third-party AI systems (NIST AI 100-1, GOVERN 6.1 and 6.2). As of this writing, NIST is revising AI RMF 1.0, so check the framework page for the current version before mapping your controls to it.

1. Pin down the use case and the decision it affects

• Ask the vendor to describe each AI-enabled feature you plan to use and the business process in which it will run, in terms more specific than “assistant” or “copilot.”
• Ask whether each feature makes, recommends, ranks, scores, summarizes, or predicts, or only drafts content for a person to review, and record the answer feature by feature.
• List every decision about a person that the output can influence, including hiring, promotion, discipline, eligibility, access, pricing, or care.
• Obtain a current inventory of models and features, and identify which ones the vendor or an administrator can switch on by configuration after launch.

2. Map inputs, outputs, and data flows

• List every category of data that will be sent to the service, whether it concerns customers, employees, applicants, patients, consumers, or the public.
• Ask what the service creates from that data, including embeddings, scores, inferences, transcripts, prompts, logs, and outputs, and where each one is stored.
• Identify hosting regions, subprocessors, support access, transfers outside the United States, and every category of recipient.
• Ask whether your data, prompts, feedback, or outputs train, fine-tune, or improve models offered to other customers, and whether the contract, not only a settings toggle, lets you turn that off.
• When the vendor says customer data is not used for training, confirm that the promise also covers prompts, logs, feedback, derived metadata, and support tickets.
• Ask for the retention period of each data type, including backups, and require that decision logs remain available long enough to reconstruct an individual outcome.

3. Test the evidence behind performance claims

• Request validation materials for the intended use, including known failure modes and the conditions under which accuracy degrades.
• Ask what population and setting the testing represented and where they differ from your users, your workforce, or your patients.
• Ask how a confidence score, ranking, or rating should be read, and what the vendor says the system should not be used to decide.
• Request the documentation behind every accuracy, bias, or capability claim in the proposal, the marketing materials, or the demonstration, and keep dated copies.
• Before repeating a vendor claim to your customers or employees, confirm you can support it. The FTC Act declares deceptive acts or practices in commerce unlawful, and a claim you publish becomes your statement (15 U.S.C. 45(a)(1)).

4. Identify employment, biometric, health, and accessibility triggers

• If the output feeds recruitment, hiring, promotion, discipline, discharge, or other terms of employment, review the use under the Illinois Human Rights Act provision in effect since January 1, 2026, which bars AI use that has the effect of discriminating on the basis of protected classes or that uses ZIP codes as a proxy, and requires notice to employees. The statutory definition of AI includes generative AI (775 ILCS 5/2-101(M), 5/2-102(L), P.A. 103-0804). See our AI in Hiring white paper.
• Ask whether the vendor can supply the facts your employee notice will need, and check whether the Illinois Department of Human Rights has adopted the notice rules the statute directs it to issue before you finalize that notice.
• If the tool analyzes recorded video interviews for Illinois-based positions, confirm the vendor supports advance notice, an explanation of how the AI works and the general characteristics it evaluates, consent, limited sharing, and deletion within 30 days of an applicant’s request, including backup copies. If you would rely solely on the AI analysis to decide who receives an in-person interview, also ask whether the vendor can supply the race and ethnicity data the employer must report each year (820 ILCS 42/5, 42/10, 42/15, 42/20).
• Ask whether the system captures or creates retina or iris scans, fingerprints, voiceprints, or scans of hand or face geometry, or information based on them that is used to identify a person (740 ILCS 14/10). A statement that the product does not “store biometrics” does not answer whether it creates face geometry or voiceprints during processing.
• If covered biometrics are in the flow, confirm the vendor will support your written notice, written release, public retention and destruction schedule, and disclosure limits, and that it will not sell or otherwise profit from the data (740 ILCS 14/15).
• If the vendor will create, receive, maintain, or transmit protected health information on behalf of your organization as a HIPAA covered entity, the vendor is a business associate (45 C.F.R. 160.103). Require a business associate agreement with the contract elements in 45 C.F.R. 164.504(e) and the security terms in 45 C.F.R. 164.314(a), including reporting of security incidents.
• Confirm that any use of protected health information to train or improve the vendor’s models is a use the business associate agreement actually permits, because the agreement must bar uses and disclosures beyond those it allows or the law requires (45 C.F.R. 164.504(e)(2)(ii)(A)).
• If licensed therapy professionals in Illinois will use the tool, confirm that each feature is limited to administrative or supplementary support for a licensed professional and that none provides therapy or makes therapeutic decisions, as the Wellness and Oversight for Psychological Resources Act requires (225 ILCS 155). Before AI is used on recorded or transcribed sessions, review the Act’s notice and consent conditions. See our guide to the Illinois AI therapy law.
• Ask how a person with a disability can be moved to an alternative process, since an employer covered by the ADA must make reasonable accommodations absent undue hardship and must administer employment tests so the results reflect what the test measures rather than an impaired sensory, manual, or speaking skill (42 U.S.C. 12112(b)(5)(A), 12112(b)(7)).

5. Review oversight and human control

• Ask the vendor to name the role responsible for model risk, privacy, security, and incident escalation.
• Ask what a human reviewer sees, whether the reviewer can override or reject an output, and whether overrides are logged.
• Ask how the vendor monitors for drift, harmful bias, unsafe outputs, abuse, and material changes in performance, and what it reports to customers.
• Obtain the release and change-management process, and require advance notice of material changes to the model, training data, features, or subprocessors.

6. Review security, incidents, and continuity

• Request the security architecture, encryption, access controls, logging, tenant separation, and recent independent testing or assurance reports, and record any exceptions they disclose.
• Ask how the vendor defines an AI or data incident, and require notice of materially wrong outputs, cross-customer data exposure, and unauthorized use as well as traditional breaches.
• Where the product can act through connectors or tools, ask for its controls against prompt injection, data exfiltration, excessive permissions, and unauthorized autonomous actions.
• Ask for recovery objectives, backups, continuity arrangements, the external models or APIs the product depends on, and the plan if one of them becomes unavailable.

7. Contract for audit, deletion, and exit

• Write the approved use, prohibited uses, confidentiality, security, retention, deletion, subprocessor, and incident-notice commitments into the agreement or an incorporated exhibit.
• For personal information of Illinois residents, include a provision requiring the vendor to implement and maintain reasonable security measures (815 ILCS 530/45(b)). Our privacy vendor contract checklist covers the general data terms.
• Require cooperation with regulatory inquiries, discrimination reviews, access and deletion requests, and litigation holds, and obtain audit or evidence rights matched to the claims and controls you rely on.
• Define the export format, transition support, deletion certification, and treatment of derived data, including embeddings and any model fine-tuned on your data, at termination.

8. Separate verified answers from sales assertions

• Mark each material answer by its support: contract term, technical documentation, independent attestation, your own testing, or sales representation.
• In a controlled environment, test the controls you can observe, such as deletion, export, permissions, and human review, and keep screenshots or logs of the configuration you approved.
• Record exceptions and open questions with an owner, a due date, and whether each one must close before go-live.

9. Record the go-live decision

• Record the decision (approve, approve with conditions, pilot, remediate, or reject), who made it, and what evidence was available at the time.
• Record the approved use cases, prohibited uses, system owner, legal owner, data owner, and renewal date.
• Attach the completed questionnaire, contract, security review, notices, any risk or impact assessment, and any risk the organization accepted.
• Set monitoring measures and list the events that reopen diligence: a change in model, use case, data, jurisdiction, subprocessor, or material feature.
• Schedule the renewal review early enough before any automatic renewal to renegotiate or exit.

Scaling the questionnaire to the tool

Not every AI purchase needs every question. Two facts set the depth of the review: how significant the decision is that the output touches, and how sensitive the data is that the tool receives. A drafting assistant used on nonconfidential internal material can usually clear groups 1, 2, 6, and 7 with short answers. We recommend the full questionnaire whenever the output influences a decision about a person, the tool receives protected health information, biometric data, or applicant video, or the product can take actions through connectors without a person approving each step.

AI features that arrive inside software you already license

Many AI features are never purchased. They arrive in an update to a human resources platform, an electronic health record, or an office suite, sometimes enabled by default, under a contract that may let the vendor change its terms by posting new ones online. When a new AI option appears in an administrator console, run groups 1, 2, and 4 before turning it on, and check whether your negotiated data-use terms reach the new feature. Save a dated copy of any incorporated online terms as they stood when the feature was enabled.

When the vendor will not answer

A refusal is information. Vendors decline because the material is confidential, because it sits behind a nondisclosure agreement, or because no one there knows the answer. The response depends on the reason: sign a reasonable confidentiality agreement to see the documentation, accept a summary backed by a contractual representation, narrow the use case, pilot with synthetic or nonsensitive data, or decline the product.

Some gaps cannot be accepted. If the vendor will not say whether it creates voiceprints or face geometry, the organization cannot give the BIPA notice or obtain the release the statute requires. If a vendor handling protected health information will not sign a business associate agreement, a covered entity should not send it that information. Log those gaps as exceptions and hold go-live until they close.

What the contract cannot move to the vendor

An agreement allocates cost and cooperation. It does not transfer duties the statute places on the buyer. An Illinois employer still owes the Human Rights Act notice, and a covered entity still answers for its own HIPAA compliance. The contract’s job is to make the vendor provide the information, controls, assistance, and evidence the buyer needs to meet those duties.