These emails land in SME inboxes every week. The old supplier questionnaire, with SOC 2 reports, GDPR clauses and uptime pages, does not cover what the software does.
If the vendor adds a model to the product you rely on, you may need a new part. Here’s what to ask, according to the scenario, before you sign up.
Vendors That Build Chatbots into Existing Products
This is the most common case and the easiest to study. The products you already use have spawned chat assistants, “draft with AI” buttons, or summary panels. The vendor calls the feature. To register your risk as a new data stream.
Ask who created the model, who hosted it, and what the tenant left behind when the user clicked the button. If a vendor directs to a third-party API, that provider is now a subprocessor that you inherit without signaling anything. A software partner you can trust, whether it’s an internal or external team software development companywill sketch the request path on the whiteboard without flinching.
- Data in directions. Which fields from your records are sent to the model, and customer identifiers are removed before leaving?
- Practice reuse. Specify that your inputs and outputs are not used to train models of vendors or upstream providers.
- Opting out at the tenant level. You want to switch your control, not a support ticket you need to raise.
Vendor whose product is a model
It’s a different story altogether. Here the software is AI: document classifiers, forecasting tools, voice agents. The questionnaire should go deeper than security and how it works.
Ask for a model card or similar documentation: what is being trained, what to do and not to design, how often to retrain, and how changes are communicated to you. An Atlas Systems walkthrough recommend treating the vendor’s inability to explain how its model works as a red flag rather than a technical one. The point is not to make a thesis writing supplier. This is to see if anyone can answer the question.
Push in evaluation as well. What is the vendor’s published accuracy or error rate, measured against what benchmarks, and when was it last reviewed? Headline accuracy numbers without definition are marketing, not evidence.
Vendors Sell You Agents That Take Action
Agent features are the fastest growing part of vendor pitch 2026. The software is not just a suggestion. It’s books, emails, returns, posts and files. That changes the shape of risk entirely, because the wrong agent can now do something before people see it.
- Blast radius. What actions can these agents take without human approval, and what is the maximum financial, reputational or data impact of a bad call?
- Human-in-the-loop control. Can you require a review step for a defined category, such as payment over the threshold, customer order or record deletion, and is it enforced by the product or policy?
- Audit log. Each agent’s actions, with the resulting prompts and inputs, are retained long enough to investigate complaints months later.
- Switch off. A single documented way to pause agents in your tenant without removing the entire product.
You Inherit Your Vendor Compliance Posture
Buy AI software and you buy part of the vendor’s regulatory story. When customers or auditors ask whether the framework manages the model, “the vendor handles it” is not the answer that keeps coming up. Ask your recognized framework for a supplier map, and ask to see the mapping.
At NIST The AI Risk Management Framework is a common reference point for a reason: it gives both sides a shared vocabulary for governance, measurement and oversight without locking you into one jurisdiction. If you or your customers operate in Europe, you also want to know how the vendor positions itself in the EU AI Act provider-versus-deployer split, because some obligations flow to you as a buyer regardless of the said contract.
Certification is the starting point, not the finish line. SOC 2 and ISO 27001 cover pipelines. They say nothing about how the model performs.
Read the Evidence, Not the Answer
A questionnaire is only as good as you make the responses. If the vendor says yes to a control, ask for artefacts: policy, sample log, architecture diagram, subprocessor list with date. A senior engineer on either side can tell within ten minutes whether the evidence is real or someone has pasted the right word in the right box.
The point is that this new section does not turn every SME into an AI auditor. This is to ensure that the person behind the sales phone can answer the questions that buyers will need to ask in 2026. If they can’t, then the findings.