Data privacy questions to ask any AI vendor

Why vendor questions are your job

When you connect an AI tool to your phones, website chat or customer records, its data practices become part of yours. Customers will not distinguish between your company and a vendor if a recording leaks or a transcript is misused. They are already wary: nine in ten US adults think people often or sometimes pretend to be someone else to steal personal information by phone or online^[2]. A careful company asks hard questions before signing, not after a problem.

You do not need a security team to do this well. You need a consistent list of questions, a habit of getting the answers in writing and a few clear red flags that end the conversation. The rest of this guide gives you all three.

1. What data do you collect?

1. Which data does the product collect from each call, chat or text: audio, transcripts, summaries, caller ID, location, metadata?
2. Does it collect anything beyond what I configure, such as device information from website visitors?
3. Can I switch off recording or transcription for certain numbers, call types or hours?
4. How does the system handle sensitive details a caller volunteers, such as card numbers, health information or alarm codes? Can it redact them?

A good answer is specific and comes with documentation. A vague answer such as we collect what we need to provide the service is a reason to keep asking.

2. Where is it stored, and for how long?

1. In which countries and with which hosting providers is my data stored?
2. What is the default retention period for recordings, transcripts and summaries, and can I set my own?
3. Is data deleted automatically at the end of the retention period, including backups, and on what schedule?
4. Can I delete an individual customer's records on request, and how long does that take?

3. Is my data used to train AI models?

1. Are my recordings, transcripts or customer data used to train or improve your models or any third party's models?
2. If yes, is that the default, can I opt out and is the opt-out in the contract?
3. Do the AI model providers you use retain my data, and under what terms?
4. Is my data ever combined with other customers' data, even in anonymized form?

This is the question most specific to AI vendors and the one where answers vary most. Ask for the answer in the contract or the data processing terms, not only in a sales email. A vendor does not need to name its technology suppliers in public, but it should be willing to describe their data terms to a customer, under a confidentiality agreement if necessary.

4. Who can access it?

1. Which of your staff can access my recordings and transcripts, for what reasons, and is that access logged?
2. Which subprocessors handle my data, and will you notify me before adding new ones?
3. Can I control access in my own account by role, so a technician sees summaries but not every recording?
4. Do you support multifactor authentication and single sign-on for my users?

5. How is it secured, and what happens when something goes wrong?

1. Is data encrypted in transit and at rest?
2. Do you have an independent security report or certification, such as a SOC 2 Type II report or ISO 27001, and can I see it under NDA?
3. When was your last independent penetration test, and were findings fixed?
4. How quickly will you notify me of a security incident affecting my data, and is that commitment in the contract?
5. When I cancel, how do I export my data, and when is everything deleted?

6. Questions specific to AI behavior

Privacy is not only about storage. AI tools also make choices in real time about what to say and what to record, and those choices affect your customers. Customers care about transparency: 76% of US adults say it is extremely or very important to be able to tell whether content was made by AI or by people^[3].

1. Does the system tell callers it is AI, and can I control the wording?
2. Does it ever clone or imitate a real person's voice, including my staff, and can that be switched off?
3. Are the instructions I give the system, such as approved answers and service-area lists, stored separately from customer data?
4. Can the system be prompted by a caller to reveal other customers' information, and how have you tested for that?
5. Are AI-generated summaries labeled as such in my records, so staff know to check them?
6. If the system makes a mistake, such as recording the wrong address, can I correct the record and see the original transcript?

The fourth question is worth testing yourself. On a test call, ask the system for the address of the last customer who called or for another customer's appointment time. It should refuse every time.

Recording consent and customer notices

Privacy also covers how calls are recorded in the first place. Federal law permits recording when one party consents^[4], but some states require all parties to consent; Florida permits intercepting a call only when all parties have given prior consent^[5]. Ask the vendor where the recording notice plays, whether you can edit it and whether recording starts only after the notice. Our guide to call recording consent by state covers the state picture.

State privacy laws may add obligations. The California Consumer Privacy Act applies to for-profit businesses with gross annual revenue over $25 million, or that buy, sell or share the personal information of 100,000 or more California residents or households, among other tests^[1]. Businesses that sell personal information must give consumers a clear Do Not Sell or Share My Personal Information link^[1]. Many small contractors fall below thresholds like these, but your privacy notice should still describe accurately how call data is collected and used. See state privacy laws for small businesses for more.

Running the review

1. 01

   Send the questions in writing

   Copy the five groups above into an email or a short questionnaire and send the same list to every vendor.

2. 02

   Score the answers

   For each question, mark clear and documented, vague, or missing. Vague answers on training use, retention and incident notification count as missing.

3. 03

   Read the contract

   Check that key answers appear in the terms of service or data processing agreement, not only in marketing pages.

4. 04

   Configure for minimum data

   After signing, set the shortest retention that works, restrict access by role, turn on multifactor authentication and switch off features you do not need.

5. 05

   Recheck yearly

   Ask for updated security reports and subprocessor lists each year, and whenever the vendor announces major new AI features.

Example: two vendor answers compared

Example: a fictional cleaning company sends the questionnaire to two AI answering vendors. Vendor A replies within two days with a retention setting the customer controls, a contractual statement that call data is not used to train models, a subprocessor list, role-based access and a 72-hour incident notification commitment. Vendor B replies that data is handled according to industry best practices and that training use helps improve the product for everyone, with no opt-out in the contract. The owner prefers Vendor B's voice on test calls, but chooses Vendor A. Voice quality can be tuned later; data terms are hard to change after signing.