How to Vet an AI Vendor's Data Handling Before You Sign

A sales team finds an AI note-taking tool that joins every client call, writes the summary and drops it into the CRM. The trial is free, the demo is good, and the rep wants a signature by Friday. Nobody has asked where the recordings go, who can read them, or whether they help train the vendor's next model.

Those questions take an hour to ask before you sign. They take months to unwind after. This guide covers what to ask, which answers are acceptable, and what to get in writing.

Why the standard vendor review falls short for AI

A normal software review asks about uptime, access controls and certifications. Those still matter. AI tools add a few risks a standard questionnaire misses:

Your data may be used to improve the product. Some vendors train models on customer inputs unless you opt out or sit on a business tier.
Prompts and files are often stored. Many tools keep conversation history, uploaded files and outputs for days, months or until you delete them.
The vendor is often a reseller. Many AI products are built on a model from another company, so your data may pass through two or three parties.
Terms change. The FTC warned in 2024 that quietly rewriting terms of service or privacy policies to allow new uses of customer data, including AI training, could be an unfair or deceptive practice. The warning exists because it happens.

The cost of getting this wrong is measurable. IBM's 2025 Cost of a Data Breach report found that organizations with high levels of unsanctioned "shadow AI" paid about $670,000 more per breach than those with little or none, and that only 37% of organizations had policies to manage AI or detect shadow AI. The global average breach cost was $4.44 million.

Step 1: Classify the data before you read any terms

The right level of scrutiny depends on what the tool will touch. Sort the use into one of three tiers first.

Tier	Data the tool touches	Level of review
Low	Public content, generic drafting	Read the terms, confirm business tier
Medium	Internal documents, meeting notes, code	Full question list below
High	Client data, health, financial, HR records	Full list plus a signed data processing agreement

If staff will paste client contracts, patient information or payroll data into a tool, treat it as high. Your AI acceptable use policy should already say which tier of data may go into which kind of tool. If it does not, write that first.

Step 2: Ask the six questions that matter

Send these in writing and keep the answers. A vendor that cannot answer them clearly has told you something.

1. Is customer data used to train models?

You want a flat "no" for your plan tier, stated in the contract and not only on a marketing page. Check the plan itself. Free and consumer tiers often carry different terms than business or API tiers, and staff tend to sign up with the free one.

2. How long is data retained, and can we control it?

Ask for retention periods on prompts, uploaded files, outputs and logs. Ask whether you can set a shorter period, and whether "deleted" means removed from backups and logs or only hidden from your view. Many tools also keep a short window of data for abuse monitoring, which is reasonable if it is disclosed and limited.

3. Who can see our data inside the vendor?

Ask whether vendor employees can read prompts and files, under what conditions, and whether that access is logged. Human review for quality or safety is common. It should be documented, restricted and covered by confidentiality terms.

4. Which subprocessors and model providers touch the data?

Ask for the full list, including the underlying model provider and the cloud host. Ask how you get notified when it changes. If the vendor resells another company's model, you need the terms of both.

5. How is data protected and where does it live?

Ask for encryption in transit and at rest, single sign-on and multi-factor support, role-based access, audit logs, and the region where data is stored. Ask for a current independent audit report, such as a SOC 2 Type II, and read the scope. A report that covers the vendor's corporate network but not the AI product tells you little.

6. What happens when we leave, or when something goes wrong?

Ask how you export your data, how quickly it is deleted after termination, and whether deletion is certified in writing. Ask for the breach notification commitment: how fast they notify you and what they tell you.

Step 3: Get the answers into the contract

A sales call is not a commitment. Terms of service often let the vendor change the rules with notice, or none. Before signing, push for these in a data processing agreement or order form:

A statement that your data is not used for training, tuning or product improvement
Named retention and deletion periods
A subprocessor list with advance notice of changes
A breach notification window
The right to delete data and receive written confirmation at termination
Notice and a right to terminate if the data terms change materially

Smaller vendors will sometimes refuse custom terms. That is a decision point. If the tool only touches low-tier data, accepting standard business-tier terms may be fine. If it touches client or regulated data and the vendor will not commit in writing, pick a different tool.

Red flags

The privacy answer lives only in a blog post or FAQ, not the contract
"We anonymize your data" with no detail on how
No list of subprocessors, or "we use leading providers"
Training opt-out requires each user to switch it off individually
No independent audit report, or one that excludes the AI product
The vendor will not say which model provider sits underneath

NIST's AI Risk Management Framework and its generative AI profile treat third-party and data-handling risk as something an organization governs, not something it assumes the vendor handled. It is voluntary guidance, but it gives you a neutral reference when a vendor pushes back.

Who should run the review

At a company with 25 to 200 people, this review usually lands on whoever found the tool, which means it often does not happen. Three habits help:

Route every new AI tool through one owner. One person or one shared inbox, not each department.
Keep a list of approved tools and the data tier each may touch. Update it when terms change.
Re-check annually, and at renewal. The tool you approved last year may not have the same terms today.

Example engagement: a 70-person accounting firm asks about adopting an AI tool that summarizes client documents. Elevaire would classify the data as high tier, send the six questions to the vendor, review the audit report and subprocessor list, and mark up the data processing agreement. If the vendor declines to commit to no training use in writing, Elevaire would recommend an alternative and add the tool decision to the firm's approved-tools list.

This is part of what a virtual CISO handles: vendor assessments and written risk decisions that a board or client can read later. Elevaire Managed Security then enforces the outcome, with controls that restrict unapproved AI tools and monitoring for sensitive data leaving the company. It works alongside your existing IT provider, which usually owns the purchasing and setup side.

Frequently Asked Questions

What should I ask an AI vendor about data privacy?

Ask whether customer data is used for training, how long prompts and files are kept, who inside the vendor can read them, which subprocessors and model providers are involved, how data is protected, and how deletion works when you leave. Get the answers in writing and tie the important ones to the contract.

Does an AI vendor train on my company's data?

It depends on the vendor and the plan. Business and API tiers often exclude customer data from training by default, while free and consumer tiers may not. Do not rely on a website statement. Confirm it for your specific plan and have it written into the agreement.

Is a SOC 2 report enough to trust an AI vendor?

No. A SOC 2 Type II report shows the vendor runs security controls consistently, which is useful. It does not tell you whether your data trains models, how long prompts are retained, or which model provider sees them. Check that the report covers the AI product you are buying, and still ask the data handling questions.

How much does an AI vendor review cost?

Done in house, a review of one tool takes a few hours of someone's time. Elevaire includes short vendor questionnaires and security vendor management in Elevaire Managed Security, which carries predictable monthly pricing with no hourly billing. For context on the stakes, IBM's 2025 research put the extra cost of breaches involving high levels of shadow AI at about $670,000.

How does this work alongside our existing IT provider?

Your IT provider typically handles purchasing, licensing and setup. Elevaire adds the security review, the written risk decision and the controls that enforce it. Neither replaces the other.

How do I get started?

List the AI tools your teams already use, including free ones, and sort them by the data they touch. Start the full review with the highest-tier tools. The free Security & AI Risk Scorecard is a quick way to see where your gaps are.