Summary: Most vendor questionnaires predate generative AI. They cover security certifications and business continuity, but miss whether the vendor trains on your data, which model provider sits underneath, and what happens when the vendor switches on a new AI feature after you've signed. Add six groups of questions, and monitor for changes between annual reviews.
Why the old questionnaire falls short
A standard vendor due diligence questionnaire asks about SOC 2 reports, encryption, access controls and disaster recovery. Those still matter. But an AI vendor raises questions the old format doesn't reach:
- Your data may be used to train or improve a model.
- The vendor is often a wrapper around a model provider you have no contract with.
- The product changes far more often, and a new AI feature can arrive in a routine update.
Regulators are moving in the same direction. FINRA's 2026 Annual Regulatory Oversight Report expects firms to maintain inventories of vendor services, systems and the firm data they access, including AI tools and embedded features. Amended Reg S-P requires advisers to oversee service providers that handle customer information.
Wilson Sonsini's analysis of the first shadow AI 8-K makes a related point: vendor certifications do not guarantee that AI features added through product updates meet your data handling requirements.
Six groups of questions to add
1. Training and data use
- Is customer data used to train, fine-tune or improve any model? Is that the default, and how do we opt out?
- Does the opt-out apply to all plans, or only enterprise tiers?
- Are prompts, outputs or uploaded files reviewed by humans at the vendor?
2. Retention and deletion
- How long are prompts, outputs and files retained?
- Can we set our own retention period, including zero retention?
- How do we confirm deletion when the contract ends?
3. Model providers and sub-processors
- Which model providers does the product use, and for which features?
- Do those providers have their own retention or training terms? Do they apply to our data?
- Where is data processed and stored? Can we restrict it to the US or EU?
- How will we be notified when a sub-processor is added or changed?
4. Security and access
- Does the product support SSO and role-based access?
- Are there audit logs of who used the tool and what they accessed?
- Can admins restrict which features and connectors users can turn on?
5. Embedded and future features
- Which features use generative AI today? Can each be disabled separately?
- How much notice do we get before a new AI feature is enabled?
- Does the product connect to other systems (email, files, data rooms) through integrations or agents? What permissions do those require?
6. Contract terms
- Does the DPA cover AI processing explicitly, including model providers?
- Is there a commitment to notify us of material changes to data use or sub-processors?
- What are the breach notification terms, and do they meet our Reg S-P obligations?
Answer once, then watch for changes
The bigger change is timing. AI vendors update their terms, plans and features constantly. An answer you got in January may be wrong by June: a training opt-out becomes enterprise-only, a new sub-processor appears, or a meeting assistant starts summarizing calls by default.
Annual reviews miss this. A better pattern is to collect the answers once at approval, then monitor the vendor's public terms, trust center and product announcements for changes that affect what you approved. When something changes, the approval goes back for review.
Match diligence depth to the use case
Not every AI tool needs the full set of questions. A tool used for public-information research with no firm data needs less scrutiny than one that processes LP records or deal terms. Tie the depth of review to the data classifications the tool will touch, and record that classification in the approval. That keeps reviews fast for low-risk requests and thorough where it matters.
Where Clarier fits
Clarier researches each AI vendor against these questions and produces a trust report covering training terms, retention, sub-processors, security posture and AI-specific risks. After approval it keeps monitoring, so you hear when a term, a setting or a feature changes. Our own answers are on the Clarier Trust page.
To see how your current vendor process compares, take the AI program maturity assessment.
