Skip to content

AI vendor due diligence: questions to ask

[PUBLISHED_DATE] 2026.11.11
[READ_TIME] 6 MIN
[AUTHOR_ID] TECH INNOVATION

Five groups of questions: what the vendor does with your data, which security controls it can prove, which model and subcontractors sit behind it, what the contract says, and how you exit if the relationship ends. An AI vendor is vetted like any technology vendor, with two differences: there's usually another company behind it running the model, and its terms change faster than any other vendor's.

One warning before you start: the hardest AI vendor to vet is the one you already have, when it switches on an AI feature in an update.

Why is an AI vendor different?

Three reasons:

  1. There's a model behind it. Many AI tools don't have their own model; they use another company's. Your data can pass through a subcontractor that isn't in your contract.
  2. Data can stick around. Training, retention, feedback and safety logs follow different rules, and they're not always in the same document.
  3. It changes fast. Plans, features and terms change within months. What you vetted at signing may not be what you're running a year later.

What should you ask an AI vendor?

Twenty questions in five groups:

Data

  1. Do you use our data, inputs or outputs to train models? Are there exceptions, such as feedback?
  2. How long do you keep conversations, files and logs? Can we set retention?
  3. Where is data stored and processed? Can we restrict the region?
  4. Who on your team can see our data, and when?

Security

  1. Which certifications and reports do you have (SOC 2 Type II, ISO 27001, ISO/IEC 42001)? Can we see the full reports?
  2. Do you support single sign-on, automated provisioning, roles and audit logs?
  3. How, and how fast, do you notify us of a security incident?
  4. How do you protect the tool against malicious instructions hidden in documents or web pages (prompt injection)?

Model and subcontractors

  1. Which model does the tool use, and whose is it?
  2. Which subcontractors process our data, and will you tell us if they change?
  3. What happens if you switch models or model providers?
  4. Do connections to other systems (connectors) carry the same terms as the main product?

Contract

  1. Does the contract let us audit you, and let our regulator do the same?
  2. Do you have a data processing addendum we can review? For PHI, will you sign a business associate agreement?
  3. What service levels do you commit to, and what happens if you miss them?
  4. How much notice do you give before changing your terms or usage policy?

Exit

  1. How do we get our data and documents back if the relationship ends, or if you run into trouble?
  2. In what format, and how quickly?
  3. How do you confirm our data was deleted after we leave?
  4. What happens to what we built on your platform: prompts, workflows, configurations?

What do the major vendors publish?

Serious vendors publish many of these answers before you ask. Two examples, as of October 2026:

  • Anthropic (Claude): its commercial terms bar it from training models on customer content, and it lists ISO 27001:2022, ISO/IEC 42001:2023 and SOC 2 Type I and Type II, with reports in its Trust Portal. For health care it offers a HIPAA-ready configuration with a business associate agreement, which is a configuration, not a certification.
  • Google (Gemini in Workspace): says it doesn't use customer data to train models without permission, that conversations aren't reviewed by people, and that interactions stay within the organization.

A vendor publishing it doesn't replace your review: confirm the terms are in your contract and apply to the plan you'll actually use. We compare what each plan promises in what data employees can put into AI tools.

What do US regulators expect?

  • Banks: the interagency guidance on third-party relationships (OCC Bulletin 2023-17) covers the whole lifecycle: planning, due diligence, contract negotiation, ongoing monitoring and termination. These twenty questions map onto it. In remarks published May 1, 2026, Fed Vice Chair for Supervision Michelle Bowman noted that third-party risk management guidance is being updated, so watch for changes.
  • Credit unions: GAO reported in May 2025 that the NCUA lacks the authority to examine technology service providers, and its recommendation that Congress grant it remains open. Your due diligence file is the only review that vendor gets.
  • Health organizations: HIPAA lets you share PHI with a vendor only with satisfactory assurance, documented in a business associate agreement. No BAA, no PHI.

For more on what governs AI at banks now, see what replaced SR 11-7.

What should worry you?

  • They can't tell you which model they use or who processes it.
  • The certification shows up as a logo, but the report never does.
  • The contract says nothing about audit rights or data return.
  • They only offer individual accounts, with no central administration.
  • They can change their terms without telling you.

An honest caveat: the large AI vendors' standard terms aren't always negotiable. If a clause you need isn't there, document how you cover it, whether through an enterprise agreement, the purchasing route or your own controls.

How do we handle it at Andes?

Our AI governance package includes vendor due diligence templates, along with the AI policy, the use case inventory and the risk assessment. To see what Anthropic keeps and for how long, read whether Claude trains on your company's data.

Did your last AI vendor answer these twenty questions?

If nobody asked, there's still time. See our AI for regulated institutions service and let's build the vendor file together.

Sources

  • OCC Bulletin 2023-17, Third-Party Relationships: Interagency Guidance on Risk Management: https://www.occ.gov/news-issuances/bulletins/2023/bulletin-2023-17.html
  • GAO-25-107197, Artificial Intelligence: Use and Oversight in Financial Services (May 2025): https://www.gao.gov/products/gao-25-107197
  • Federal Reserve, Michelle W. Bowman, Artificial Intelligence in the Financial System (published May 1, 2026): https://www.federalreserve.gov/newsevents/speech/bowman20260501a.htm
  • 45 CFR 164.502, Uses and disclosures of protected health information (Cornell LII): https://www.law.cornell.edu/cfr/text/45/164.502
  • Claude Privacy Center, What Certifications has Anthropic obtained?: https://privacy.claude.com/en/articles/10015870
  • Anthropic, Commercial Terms of Service: https://www.anthropic.com/legal/commercial-terms
  • Google Workspace, Generative AI in Google Workspace Privacy Hub: https://knowledge.workspace.google.com/admin/gemini/generative-ai-in-google-workspace-privacy-hub

Related Articles

View all resources
AI and governance

Agencies, freelancers and branches: how to give them brand access without losing control

2026.11.09
AI and governance

Does Claude train on your company's data? Training, retention and zero data retention

2026.11.06
AI and governance

What an AI acceptable use policy should include, and who approves it, at a bank, credit union, insurer or hospital

2026.11.04
Direct_Answers

Frequently asked questions

Five groups of questions: what the vendor does with your data (training, retention, location), which security controls it can prove (certifications and reports), which model and subcontractors sit behind it, what the contract says (audit, data return, service levels), and how you exit if the relationship ends.

The most common are SOC 2 Type II and ISO 27001 for information security, and increasingly ISO/IEC 42001 for AI management systems. Ask for the reports, not just the logos, and check which products and which period they cover.

Yes. Many platforms you already use, such as your CRM, messaging or document system, add AI features in a routine update. Each one can change where your data goes and who processes it, so it deserves the same review as a new vendor.

Not directly. GAO reported in May 2025 that the NCUA lacks the authority to examine technology service providers, and its recommendation that Congress grant that authority remains open. In practice, credit unions carry the full weight of vendor due diligence themselves.

READY FOR THE NEXT DEPLOYMENT?

Start a Technical Consultation