A Vendor AI Questionnaire a Small Company Can Get Answered

A forty-page vendor questionnaire sent by a twelve-person company does not get answered; it gets a link to the vendor’s trust page and a signed order form. The tool is bought anyway, and the company has created a paper record that it asked questions it never got answers to.

The long form fails at small scale because neither side can afford to own it

Enterprise vendor questionnaires are written by procurement teams for vendor-response teams. Both sides have people whose job is the form. A small company has neither: the person sending it is also the person who needs the tool, and the vendor’s sales engineer triages inbound questionnaires by deal size. In our experience the result is predictable. The long form is returned late, partially, or not at all, and the buyer, who has already decided, treats silence as acceptable.

The questions still need asking, and the reason is structural rather than bureaucratic. The National Institute of Standards and Technology (NIST) explains in its AI Risk Management Framework (AI RMF 1.0), published in January 2023, that measuring AI risk is hard when the organization that built the system “may not be transparent about the risk metrics or methodologies it used.” A buyer cannot inspect a hosted model. The questionnaire is the only instrument it has, which is exactly why it must be short enough to be answered and specific enough that the answers mean something.

NIST’s own supply chain guidance models the right shape. Special Publication 800-161 Revision 1, updated in November 2024, is its guidance for federal agencies on cybersecurity supply chain risk management (C-SCRM); it addresses products and services that “may potentially contain malicious functionality, are counterfeit, or are vulnerable due to poor manufacturing,” and its supplemental material includes a scoping questionnaire, published as a fillable form. Scope first, then depth where the scope justifies it, is the discipline a small buyer should copy from it.

Seven questions with published standards behind them and a contract clause in front of each

The questionnaire below is the one we send. The first six questions trace to a published control, so a vendor that declines one is declining a recognized standard rather than a buyer’s preference; the seventh is our own check on the other six. Every question maps to a clause that the answer becomes.

1. Which third-party models, APIs, and datasets does your product depend on, and who is accountable for each layer? AI RMF subcategory MAP 4.2 calls for “internal risk controls for components of the AI system, including third-party AI technologies” to be identified and documented. The Cloud Security Alliance (CSA) AI Controls Matrix, released in July 2025 with 243 control objectives across 18 domains, maps shared responsibility across cloud service providers, model providers, orchestrated service providers, and application providers. A vendor that cannot name its layers cannot tell you who is responsible when one fails. The answer becomes the sub-processor and subcontractor disclosure clause.
2. Is our data used to train or improve any model, where is it stored, and when is it deleted? AI RMF GOVERN 6.1 expects “policies and procedures are in place that address AI risks associated with third-party entities,” including infringement of intellectual property and other rights. The CSA matrix carries a dedicated data security and privacy lifecycle domain. The answer becomes the data-use clause.
3. What does the product actually do with AI, and what evidence supports the performance claims? The Federal Trade Commission (FTC) put four questions to AI sellers in its February 2023 business guidance, reproduced by Kelley Drye: “Are you exaggerating what your AI product can do?”, “Are you promising that your AI product does something better than a non-AI product?”, “Are you aware of the risks?”, and “Does the product actually use AI at all?”. Ask the vendor the regulator’s questions, and the answer becomes a warranty.
4. What are the known failure modes, and what is your process when the model or a supplier behind it fails? AI RMF GOVERN 6.2 calls for contingency processes to handle failures or incidents in high-risk third-party data or AI systems. A vendor with no answer has no incident process, and your contract’s notification clause has nothing to attach to.
5. Which pre-trained models are in use, how are they monitored, and will you tell us before the underlying model changes? AI RMF MANAGE 3.2 states that “pre-trained models which are used for development are monitored” as part of regular maintenance. A silent model swap changes the product you tested; the answer becomes a change-notification clause.
6. Which attestations do you hold, and will you share the report rather than the badge? The CSA’s August 2025 mapping gives a control-by-control view of how its matrix aligns to ISO/IEC 42001:2023, the certifiable AI management system standard, and sets out a path for vendors to publish an AI self-assessment to the STAR Registry, CSA’s public registry of security and AI self-assessments. A vendor on that path has already answered most of this form once; ask for the artifact. The answer becomes the audit-and-attestation clause, with a right to the current report.
7. Who at your company owns AI risk, and may we see the policy? No single control sits behind this one; it is our own accountability check. A named owner and a dated policy are the simplest evidence that the other six answers were written by someone accountable for them. The answer becomes a named-contact and policy representation.

A questionnaire a vendor will not answer is not a control. It is a record that you asked.

The teeth are in the contract and the follow-up, not in the length of the form

A short questionnaire has force only if the answers go somewhere. Three practices give it teeth. First, every answer becomes a representation in the order form or master agreement, so an inaccurate answer is a contract issue rather than a disappointment. Second, questions four and five carry an obligation to notify, so the vendor owes you an update when the model, a supplier, or the data-use terms change. Third, the form is resent at renewal, and a changed answer reopens the decision.

The regulatory backdrop makes the first practice easier than it was. In September 2024 the FTC announced Operation AI Comply, five enforcement actions against companies that used AI claims to deceive, with Chair Lina Khan stating that “there is no AI exemption from the laws on the books.” One respondent had marketed a “robot lawyer” while, the FTC alleged, never testing whether its output matched a human lawyer’s; it agreed to a proposed order, open for public comment, requiring it to pay $193,000. A vendor asked in writing what its AI does, and what evidence supports the claim, is being asked the same question the regulator asks, and a careful vendor knows it.

The owner of this process in a small company is whoever signs the vendor agreements, usually the operations lead or the founder, and the timeline is the purchase itself: the form goes out when the trial starts, and the agreement is not signed until it comes back. Seven questions take a vendor an afternoon. A vendor that will not spend the afternoon has told you how it will behave when something goes wrong.

The enterprise questionnaire failed small companies not because it asked too much but because it asked for nothing a vendor could be held to. Seven questions, six traced to a standard the vendor already claims to follow and all seven written into the contract, produce something the long form never did: answers. Send the short one, and do not sign until it returns.