In April 2024, staff in the FTC's Office of Technology put it plainly: "Quietly changing terms of service agreements could be unfair or deceptive." The same post warned that "any firm that reneges on its user privacy commitments risks running afoul of the law." (FTC Technology Blog, April 2024) In January 2024, the same office told AI model companies to honor their customer commitments "regardless of how or where the commitment was made." (FTC Technology Blog, January 9, 2024)
That's a statement about vendor obligations. It also tells you what your own approval file has to hold. A vendor's promise about your data only protects you if you know which promise applies to your account, as of what date, and whether it has changed since. An approval that records a product name and a yes can't answer any of those questions.
So before your organization sends sensitive data to an AI service, someone should be able to explain what the service will do with that data and produce the evidence behind the answer. The approval should name the service, the account configuration, the permitted data, and the person who made the decision.
Define the use before you review the vendor
Start with the transaction. An employee will submit a particular category of information, through a particular account or application, for a particular purpose. The output will go somewhere, and someone will act on it.
Write that down, including attachments, connected document stores, and any other service involved in the request. A clinic manager pasting appointment notes into a chatbot to draft patient reminders is a different use from the same chatbot summarizing published payer policies, even on the same account. If the proposed use changes during review, update the description before anyone approves it.
Skip this step and procurement can approve one service while a business unit deploys something materially different. Everyone follows the approval they were given, and nobody agrees on what was approved.
Get evidence on six conditions
Retention. What does the provider keep, and for how long? Ask separately about prompts, attachments, outputs, usage logs, backups, and information pulled from connected systems. Record every exception to a deletion setting.
Training and other uses. Can submitted information be used for model training, evaluation, service improvement, or human review? Check whether the answer changes with the subscription tier, a configuration setting, or the feature in use.
Other parties. Which providers handle storage, retrieval, support, and processing? Many AI products run on another company's model. Ask how the vendor tells you about changes to those parties and whether you can object.
Location and access. Where is information processed and stored, and who can reach it for support or administration? A hosting-region setting answers only the question that setting controls.
Changes to the service. What notice do you get when the model, the processing arrangements, the retention terms, or the relevant settings change? Record who receives that notice and who decides whether the approved use can continue. This is the condition the FTC language above speaks to.
Verification and exit. What can you inspect or test, what evidence does the provider supply, and what happens to your information, including copies and backups, when the relationship ends?
Each answer needs a source, a date, and an owner. Good sources include the governing agreement, the service documentation, your own configuration records, and vendor responses that name the exact product and tier under review. A screenshot of a privacy setting counts only if it shows the date and the account identifier.
Note one more thing for each answer: does it come from the contract or from a help-center page? The vendor can edit a help-center page tomorrow. Where the two sources disagree, record the inconsistency.
Check the agreements the law already requires
Some evidence isn't optional. If the data includes protected health information, the vendor is a business associate, and HIPAA requires a business associate agreement before it creates, receives, maintains, or transmits that information for you (45 C.F.R. §§ 164.502(e), 164.504(e)). Confirm the agreement names the exact product and tier your employees will use.
State privacy laws, financial privacy rules, and your own customer contracts may impose their own required terms. List each one and where it appears in the vendor's paper.
Make missing information change the decision
An unresolved fact should carry an operational consequence proportionate to the use. If you can't establish how a service handles sensitive information, the sensitive information waits until the question is resolved. Testing with public or synthetic material can usually continue in the meantime.
The decision record should say what is permitted, what remains unknown, and what would justify reconsidering. A note that "the business accepted the risk" is incomplete unless it names the uncertainty, the person authorized to accept it, and the limits of that acceptance.
Give the evidence a review trigger
Approval is a point-in-time decision. Record the conditions that require another look: a new data category, a new connected application, a changed contract, a material service update, or an expired assurance document such as a SOC 2 report. Assign someone to receive those changes and decide what they mean for the approved use.
Then test the file you already have. Pick one approved AI use and rebuild its approval from the evidence on record. Check the live account settings against that record. Ask the operational owner what employees actually submit today. Every difference between those three answers is a specific issue to resolve before more sensitive information goes in.
Get the full questionnaire
I spent 12 years at Sutter Health, most recently as Assistant General Counsel, working on privacy, security, and AI governance. The AI Tool Diligence Questionnaire, Regulated Data Edition, is the workpaper I built for this review: 51 scored questions covering all six conditions, the agreements the law requires, and what you actually configured, plus the approval record. Eleven gating items cap which data a tool can touch until the gaps close, so a high total can't paper over a missing BAA or an unanswered training question.
It's $79 for the editable Word file and a PDF. Use it for every AI tool you review. Get the questionnaire