Security and Privacy Questions to Ask AI Software Vendors

Before you buy or switch AI productivity software, ask the vendor how your data is stored, who can access it, whether it is used to train models, how long it is kept and what independent audits exist. Good vendors answer in writing, point to documentation and are specific about differences between plans. Vague answers are a warning. This guide gives you a practical question list organized by topic, explains what strong answers look like and tells you what documents to request. It is general information, not legal or security advice; for regulated data or large contracts, involve a qualified professional.

Why the questions matter more with AI tools

AI tools often process the most sensitive material in a business: client emails, contracts, financial figures, meeting recordings and employee information. Some tools send content to third-party model providers or keep it for a period to provide features. The details differ widely between products and between plans of the same product, so you need answers for the exact plan you will buy.

Data use and training

Retention and deletion

• How long is content retained after use, and can we set retention periods?
• Can users and administrators delete content, and is deletion permanent?
• What happens to our data after we cancel, and how long until it is deleted?
• Are backups included in deletion, and on what schedule?

Look for specific timeframes rather than phrases such as "as long as necessary."

Access control and administration

• Does the plan support single sign-on and multi-factor authentication?
• Can we assign roles and restrict who sees which workspace or integration?
• Are audit logs available, and how long are they kept?
• Can administrators remove a departed employee's access immediately?
• Does the tool respect the permissions of connected systems, so an assistant cannot surface a document someone should not see?

That last point deserves attention. AI search tools that index your files can expose poorly protected folders if permissions are already messy. Clean up permissions before connecting a knowledge tool.

Certifications and assurance

Ask for independent evidence rather than self-description. Common items include:

1. A recent SOC 2 Type II report, usually shared under a confidentiality agreement.
2. ISO/IEC 27001 certification, if the vendor holds it.
3. A summary of penetration testing and vulnerability management.
4. A security whitepaper or trust page describing encryption in transit and at rest.
5. An incident response and breach notification policy with timeframes.

A certification does not guarantee your data is safe, but its absence in a vendor that handles sensitive information is worth probing. The National Institute of Standards and Technology publishes frameworks that many security programs reference, including its Cybersecurity Framework and AI Risk Management Framework.

Contract terms

• A data processing agreement where the vendor processes personal data for you.
• Clear ownership of your inputs and outputs.
• Breach notification duties and timelines.
• Subprocessor change notices so you know when new providers are added.
• Limitation of liability that is reasonable for the risk.
• Termination and export rights.

Have an attorney review contracts with significant exposure. Some industries have specific rules: healthcare organizations may need a business associate agreement under federal health privacy rules, and financial firms have their own obligations. If you handle such data, do not rely on a general-purpose plan without professional guidance.

Special topics: recordings and call transcription

If you use meeting or call tools, recording and transcription raise consent issues. U.S. states differ on whether one party or all parties must consent to recording, so tell participants and check the rules in the states where your participants are located. Also ask where recordings are stored, who can download them and how long they are retained.

A simple scoring approach

Assign each topic a pass, conditional or fail rating. For example, a vendor that provides a clear no-training commitment, SOC 2 report, SSO and defined retention would pass. One that offers the commitment only on higher tiers might be conditional, which means you would need that tier. A vendor that will not answer in writing fails. Record the answers in your comparison notes so you can defend the choice later.

Internal measures that matter just as much

Vendor security is only half the picture. Your own practices decide what data reaches the tool:

• Write a short policy listing what may and may not be entered.
• Use business accounts, not personal ones.
• Turn on multi-factor authentication.
• Review integrations and remove ones you no longer use.
• Train staff on common risks, including pasting confidential content into the wrong tool.

See the guide on writing an AI use policy for a small business.

Frequently asked questions

Do small businesses really need a security review?

Yes, a lightweight one. Read the data terms, ask for the audit report and turn on basic controls.

Is a free plan safe for confidential data?

Not by default. Read that plan's terms, and use a business plan when handling client or employee information.

What if the vendor refuses to share a SOC 2 report?

Many share it under a nondisclosure agreement. If they refuse any independent assurance, consider whether the risk is acceptable.

Who should review the answers?

Someone with security or legal knowledge where possible; otherwise a senior operations lead with outside advice for high-risk data.

Sources and further reading

• National Institute of Standards and Technology — Cybersecurity Framework and AI Risk Management Framework.
• Federal Trade Commission — guidance for businesses on data security and privacy.
• Cybersecurity and Infrastructure Security Agency — practical cybersecurity resources for organizations.