What Should We Ask an AI Vendor About Privacy, Retention, Training and Indemnification Before We Sign?
Five areas settle most of it: data handling, vendor security, compliance, access scope, and the rollback plan. That list is the 5-Point Vetting Checklist inside AI Transformation Framework's governance module, and it covers the questions that get expensive once a contract is signed.
Sequence matters here. Classify your data first, then ask the vendor questions, because the answers you need depend on which data the tool will touch.
What does the evidence say about where this goes wrong?
Companies name data as their biggest AI risk and then track very little of what touches it.
Morgan Stanley's report on how boards and executives are governing the rise of AI, published August 27, 2026 and drawn from 200 executives at global companies above $100 million in revenue, found that 56 percent name data risks as the top risk from their company's use of AI, while 41 percent have a formal inventory of AI models in use with assigned ownership.
ISACA's 2026 AI Pulse Poll, published in May 2026 and drawn from more than 3,400 digital trust professionals, found privacy violations cited among the top AI risks by 74 percent and loss of intellectual property by 58 percent.
A vetting checklist is what closes the distance between the risk leaders name and the inventory they keep.
What do we ask about data handling?
Four questions, in this order.
Does the vendor train models on our inputs, and what changes that answer? Consumer tiers and business tiers of the same product often carry different terms, so confirm which tier the contract covers.
How long is our data retained after we submit it, and how long after we cancel?
Which sub-processors touch our data, and where is it processed?
What happens to our data when we leave, and how do we verify deletion?
Ask these with a specific data bucket in mind. Under AITF's first governance pillar, a company defines three to five classification tiers, and the Restricted tier, covering trade secrets, PII and regulated data, permits no AI tools at all. A vendor answer that works for Internal data can fail for Confidential.
What do we ask about security and compliance?
Two questions carry the most weight. Ask which certifications and audit reports the vendor will hand over, and ask what the contract says about liability when the tool causes harm.
ISACA's European research, published March 23, 2026 and drawn from 681 digital trust professionals surveyed in February, found that 20 percent of organizations do not know who bears ultimate accountability for harm caused by AI, and 38 percent place that accountability at board or executive level. Settle the question inside the contract while you still have leverage, and have counsel review the indemnification and limitation of liability terms before you sign.
What do we ask about access scope?
Ask what the tool can reach and what it can do once it gets there. A tool with read access to a shared drive and a tool with write access to your CRM belong in separate risk conversations.
AITF's second governance pillar is where that gets recorded. The Tool Permission Matrix maps approved AI tools to each data classification bucket, so the answer to "can we use this one here" is already written down when someone asks.
What do we ask about shutting it off?
Ask how fast you can turn it off, who has the authority to do it, and what stops working when you do.
ISACA's 2026 poll found that 39 percent of organizations have no documented process for shutting down an AI system. A rollback plan covers three things: the revocation steps, the person who can execute them without waiting for a meeting, and the work that pauses while the tool is off.
Where does a checklist like this live?
Behind an intake platform, with a published service level for answering requests.
AITF's fourth governance pillar puts the vetting checklist behind an intake process that runs on Jira, Microsoft Teams, or a simple web form, with an SLA commitment for reviewing and approving or denying each request. Speed is the point. An approval path that takes weeks sends people back to their own accounts.
In the AITF engagement, the Governance gate closes at the end of month three, and one of its conditions is a live process for answering new tool requests. Morgan Stanley found that 71 percent of executives cite a requirement to use approved AI tools, and that requirement holds up when the approved list stays current.
Frequently asked questions
-
Cover five areas: how the vendor handles your data including training and retention, the vendor's own security posture, compliance and liability terms, the scope of systems the tool can access, and how quickly you can shut it off. Have counsel review the indemnification and limitation of liability language.
-
Yes. Classify your data into three to five tiers first, then evaluate the vendor for the specific tier the tool will touch. A trade secret, PII or regulated data tier can be closed to AI tools entirely, while an internal tier carries lighter restrictions.
-
Fast enough that employees use it. Publish a service level for reviewing and approving or denying requests, and run the intake through a tool people already open, such as Jira, Microsoft Teams, or a web form.
Twenty minutes with an AITF coach will show you where your tool approval process stands today.
