Ask an AI vendor where your data goes, how long it stays, whether it trains any model, who can access it, which other companies process it, how they'd tell you about a breach, and what independent evidence backs their answers. A good AI vendor security questionnaire covers those seven areas in plain language and asks for documents, not reassurances. The questions that matter most for AI, as opposed to ordinary software, are about model training and the chain of providers behind the product.

This post gives you the questions grouped by topic, explains why each one matters and what a good answer sounds like, and ends with a checklist table you can send to vendors as-is.

Why do AI vendors need different questions from other software?

Most of a standard software security review still applies: encryption, access control, backups, certifications. But AI tools add a few wrinkles.

  • Your data may be used to improve models. Some tools use customer inputs to train or fine-tune models by default, unless you're on a business plan or opt out.
  • There's usually a chain. Many AI products are built on top of a foundation model from another company, hosted on a third company's cloud, with extra services for search, transcription or monitoring. Your data may pass through all of them.
  • Prompts and outputs are data too. What staff type in, the documents they upload, and what the AI writes back may all be logged and stored.
  • Agents act. If the tool can send emails, update records or browse the web on your behalf, its access to your systems is a security question in itself.

For the fundamentals of what to keep out of AI tools in the first place, start with our guide to AI data privacy and security for business. This post assumes you've decided to use a vendor and need to check them.

What should you ask about data retention and model training?

These are the questions that most often reveal a problem.

  • Is our data (inputs, files, outputs) used to train or improve any model, yours or a third party's? You want a clear "no" for business data, in the contract, not just on a marketing page. If the answer is "only with your opt-in," check what the default is.
  • How long do you keep prompts, uploaded files, outputs and logs? Ask for each separately. "We don't store anything" is rarely fully true once logs and abuse monitoring are counted.
  • Can we set or shorten retention, and delete data on request? Including from backups, within a stated time.
  • Where is data stored and processed geographically? This matters for data protection laws such as GDPR, India's DPDP Act and sector rules, which may restrict transfers. Obligations depend on your jurisdiction; check with counsel if you handle sensitive personal data.
  • Is our data isolated from other customers? Particularly for tools that build a knowledge base or memory from your documents.

A good answer is specific and written down. A weak one points you to a general privacy policy that covers consumers, not business customers.

Who else touches your data? The model provider chain

Ask the vendor to name everyone in the chain.

  • Which foundation model providers do you use, and under what terms? Does the model provider retain or train on data sent through the vendor's account?
  • Which cloud and hosting providers? And in which regions?
  • What is your full list of sub-processors (companies that process customer data on your behalf), and how do you notify customers of changes?
  • Do you switch model providers dynamically? Some products route requests to different models depending on load or task. That's fine, but each route should meet the same data terms.
  • Do any features send data to additional services? Web search, transcription, translation, analytics, error tracking.

The point isn't to reject anyone with sub-processors; every modern product has them. It's to know the chain exists and that the vendor's commitments flow down to each link.

How do they control access to your data?

  • Which of the vendor's staff can see customer data, and under what circumstances? Look for role-based access, approvals and logging of staff access, with access granted only for support or security reasons.
  • Do you support single sign-on and enforce MFA for our users? And can we control who in our organization has access?
  • Do you offer admin controls? User management, permission levels, data export, audit logs of who did what.
  • For AI agents and integrations: what permissions does the tool request in our systems? Read-only versus write access, scoped to specific mailboxes, folders or records, and revocable. Ask how the tool defends against prompt injection when it reads external content like emails or web pages.
  • How are API keys and connection credentials stored?

What certifications and evidence should you ask for?

Certifications don't make a vendor secure, but they show that a third party has examined their controls. Common ones to ask about:

  • SOC 2 Type II report, an independent auditor's report on how security controls operated over a period of time.
  • ISO/IEC 27001 certification, the international standard for information security management systems.
  • ISO/IEC 42001, a newer standard for AI management systems, which some AI vendors pursue to show they govern AI development and use systematically. See our explainer on what ISO/IEC 42001 is.
  • Penetration test summaries from an independent tester, with a date.
  • Sector-specific attestations where relevant to you, such as healthcare or payment card requirements.

Ask for the report or certificate itself, check the scope (does it cover the product you're buying?) and the date. A small or early-stage vendor may not have formal certifications yet; that's not automatically disqualifying, but you should then expect clearer written answers and possibly a call with their technical lead.

How will they handle incidents and changes?

  • How and how quickly will you notify us of a security incident affecting our data? The contract should commit to a timeframe and the information you'll receive.
  • Do you have a documented incident response process, and when was it last tested?
  • How do you handle harmful or wrong AI outputs reported by customers? A vendor that takes this seriously will have a process for that too.
  • How do you notify us of material changes, such as a new model provider, new data uses or changes to retention?
  • What happens to our data when we leave? Export format, deletion timeline, confirmation of deletion.

If something does go wrong, you'll need your own playbook as well; see our post on what to do when an AI system goes wrong.

The AI vendor security questionnaire checklist

Here's a version you can copy into an email or spreadsheet. Scale it to the risk: a tool used for drafting social posts needs a lighter review than one that reads customer records.

AreaQuestionWhat a good answer includes
TrainingIs our data used to train or improve any model?Contractual "no" for business data, including by sub-processors
RetentionHow long are prompts, files, outputs and logs kept?Specific periods per data type; configurable; deletion on request
LocationWhere is data stored and processed?Named regions; transfer mechanisms where required
ChainWhich model, cloud and other sub-processors handle our data?Named list; change notification; same terms flow down
IsolationHow is our data separated from other customers?Logical or physical separation explained, including for knowledge bases
Vendor accessWhich staff can access our data, and how is that controlled?Least-privilege roles, approvals, access logging
Customer controlsSSO, MFA, roles, audit logs, export?Available on the plan you're buying
IntegrationsWhat permissions does the tool need in our systems?Minimal, scoped, revocable; defenses against prompt injection
EvidenceWhat independent assurance do you have?Current reports or certificates with scope covering the product
IncidentsHow and when will you notify us?Defined timeframe in contract; tested process
ChangesHow do you tell us about new data uses or providers?Advance notice and the right to object or leave
ExitWhat happens to our data when we leave?Export, deletion timeline, written confirmation

Keep the answers. They become part of your inventory of AI tools, which is the backbone of right-sized AI governance.

What are the red flags in a vendor's answers?

  • Vague answers to the training question, or a different answer in the contract than on the website.
  • No list of sub-processors, or "we can't disclose our model provider."
  • Consumer-grade terms for a product sold to businesses.
  • No way to delete data or set retention.
  • Agent integrations that request broad write access "to make setup easier."
  • Pressure to sign before you've seen the data processing agreement.

None of these alone has to end the conversation, but each needs a clear explanation before your data goes in.

Frequently Asked Questions

Do small businesses really need to send a security questionnaire?

For tools that will see customer, employee or financial data, yes, though it can be short. A dozen focused questions by email, plus the vendor's data processing agreement, is proportionate for most SMBs. Low-risk tools that never see sensitive data need much less.

Is a SOC 2 report enough on its own?

It's strong evidence about general security controls, but it may not answer AI-specific questions such as model training, the model provider chain or agent permissions. Ask those separately, and check that the report's scope covers the product you're buying.

What if the vendor uses a well-known model provider?

That's common and usually fine, but the vendor's own handling of your data still matters: their logs, storage, staff access and other sub-processors. Ask what terms apply between the vendor and the model provider, especially on retention and training.

Should we avoid AI vendors without certifications?

Not necessarily. Smaller and newer vendors may not have formal certifications. Weigh the sensitivity of the data against the strength of their written answers, contract terms and technical setup, and consider limiting what data the tool receives.

Who in a small company should run vendor reviews?

Whoever owns your AI tools inventory, often an operations lead or IT manager, with input from whoever handles legal and data protection. The important thing is that one person owns it and that answers are filed, not lost in email.

Need help evaluating or replacing an AI vendor?

When a vendor's answers don't hold up, the alternative is often a system you control, with your data staying in your own environment. We build that kind of private, integrated software through our enterprise software work. Tell us which tool you're weighing up and we'll help you judge the risk honestly.