Know what’s on your site before a plaintiff’s firm does
A free privacy audit shows which cookies, pixels and trackers are running — and which ones commonly appear in privacy claims.
Our AI Vendor Due Diligence Guide gives you a repeatable process for reviewing AI tools before you buy them: risk tiers, a seven-step review, a 64-question vendor questionnaire you can send today, a scoring rubric with red flags, the evidence to request, AI contract terms and ongoing monitoring. Read all 18 pages below.
Compliance Guide · Edition 01The AI Vendor
Due Diligence
Guide
A practical process for reviewing every AI tool your business buys, with a 64-question vendor questionnaire you can send today, a scoring rubric, red flags and the contract terms that protect you.
The guide
Most AI risk arrives through a purchase order.
Very few companies build their own models. They buy chatbots, note-takers, hiring tools, writing assistants and AI features inside software they already use. Each one can send customer or employee data to a model provider you’ve never reviewed, under terms you’ve never read.
Regulators and courts have noticed. California’s rules require AI providers to support customers’ risk assessments. Colorado requires developers to hand deployers documentation. And recent rulings turn on whether a vendor used conversations for its own purposes. This guide gives you a repeatable way to ask the right questions before you sign.
The 60-second answer
Before any AI tool touches personal data, get written answers to three questions: Do you train on our data? Who are your model providers? How long do you keep it? Then put the answers in the contract.
This guide is general information, not legal advice.
Why AI vendors are different
A normal SaaS vendor stores and processes your data. An AI vendor may also learn from it, pass it to a model provider, generate new data about people, and make or shape decisions.
The legal hooks
| Rule | What it means for vendor review |
|---|---|
| CCPA service provider terms | Vendors must be contractually limited to your purposes, or their processing may count as a sale or sharing |
| CCPA risk assessments & ADMT | AI providers must give business customers the facts needed for their risk assessments; ADMT duties start Jan 1, 2027 |
| Colorado SB 26-189 | Developers must give deployers documentation on uses, limitations, training data and human review |
| GDPR Article 28 | Processor contracts, subprocessor approval and assistance with DPIAs |
| EU AI Act | Deployers of high-risk systems must follow provider instructions and keep logs; Article 50 chatbot disclosures apply now |
| Wiretap litigation | Courts ask whether a vendor used conversations for its own purposes, such as training |
Tier your AI vendors
Not every AI tool needs the full review. Tier first so the deep work goes where the risk is.
Full review
- Makes or shapes decisions about people
- Handles customer conversations
- Processes sensitive or children’s data
- Customer-facing generated content
Do: all 64 questions, evidence, legal review, AI addendum, annual re-review.
Core review
- Internal tools that see personal data
- Employee productivity and note-taking
- Marketing tools using customer lists
Do: sections A–E, contract terms on training and retention, review every two years.
Light review
- No personal or confidential data
- Public content only
- Sandboxed experiments
Do: the three key questions, an acceptable-use rule, and re-tier if use changes.
Five tiering questions
| Question | If yes |
|---|---|
| Will it influence hiring, lending, housing, insurance, health or education decisions? | Tier 1 |
| Will customers or the public type, speak or upload to it? | Tier 1 |
| Will it process health, financial, biometric, children’s or other sensitive data? | Tier 1 |
| Will employees put customer or employee personal data into it? | Tier 2 at least |
| Will it touch confidential business data or source code? | Tier 2 at least |
Watch for feature creep
Many vendors are adding AI features to tools you approved years ago. A new AI feature in an existing tool is a new review, especially if it’s on by default.
The review process
Seven steps, owned by named people, with a record at each step. Most reviews should take two to three weeks for Tier 1.
Intake form: ask the requester
- What problem does the tool solve, and for whom?
- What data will go in, and whose is it?
- Will its output affect a decision about a person?
- Who will own it day to day?
- Is there an existing approved tool that does the same thing?
Who does what
AI vendor questionnaire (1 of 4)
Answer every question in writing. Attach evidence where noted. Tags show the main risk area: Data, Legal, Ops.
Describe the product and each AI feature we’d use.Include whether features are on by default.
OpsWhich AI models power each feature? Are they your own, open-source, or from a third-party provider?
DataWhere is the service hosted, and in which countries is our data processed or accessed?
LegalWhat is the intended use of the product? What uses do you prohibit or advise against?
LegalHow do you notify customers of new AI features or material model changes?
OpsCan AI features be turned off or limited by admin, user or workspace?
OpsWho is the contact for privacy, security and AI governance questions?
OpsDo free, trial or self-serve plans have different data terms than enterprise plans? Which applies to us?
LegalDo you use our data, inputs, outputs or usage data to train, fine-tune or improve any model?If yes, which models, and are they used for other customers?
DataIf training is optional, is it off by default? How is the setting documented?
DataDo any humans (employees, contractors, labelers) review our inputs or outputs? When and why?
DataWhat data do you collect beyond what we submit (telemetry, metadata, identifiers)?
DataDo you use our data for advertising, profiling, analytics products or resale?
LegalHow long do you keep inputs, outputs, logs and backups? Can we set shorter retention?
DataHow do you delete our data at contract end, and how do you confirm deletion?
DataDo you de-identify or aggregate our data? What standard do you use, and what do you do with it?
DataAI vendor questionnaire (2 of 4)
List every subprocessor that can access our data, including model providers, hosting and support vendors.
DataWhat terms apply to your model providers? Do they train on, retain or human-review our data?
DataDo you use zero-retention or no-training settings with your model providers? Can you show it?
DataHow much notice do you give before adding or changing a subprocessor? Can we object?
LegalDo you flow down equivalent privacy and security terms to subprocessors?
LegalWhich cross-border transfer mechanisms do you rely on for EU, UK or other data?
LegalIf a model provider changes or is replaced, how will you tell us and re-test?
OpsDo any subprocessors run in countries subject to data-localization or export restrictions?
LegalWhich security certifications or reports do you hold (SOC 2 Type II, ISO 27001, ISO 42001)? Attach current copies.
OpsIs our data encrypted in transit and at rest? Who manages the keys?
DataHow is our data isolated from other customers’ data, including in prompts, caches and embeddings?
DataHow do you protect against prompt injection, data leakage and model abuse?
OpsWho at your company can access our data, and how is access logged and reviewed?
DataWhen did you last run a penetration test, including the AI features? Share a summary.
OpsHow quickly will you notify us of a security incident or data exposure? Through what channel?
LegalDo you support SSO, role-based access and audit logs for our admins?
OpsAI vendor questionnaire (3 of 4)
Will you sign our data processing agreement with CCPA service provider and GDPR processor terms?
LegalHow do you help us answer access, deletion and correction requests, including data inside models or logs?
DataCan you honor opt-outs, including stopping processing for a person within 15 business days?
LegalDo you process sensitive data, children’s data or biometric data? With what extra safeguards?
DataWill you provide the facts we need for our CCPA risk assessment or GDPR DPIA?
LegalIf the tool records or transcribes conversations, how are participants notified?
LegalDo you place cookies, pixels or trackers on our sites through your product?
DataHow do you handle legal requests from governments for our data?
LegalDo you have a documented AI governance program? Who is accountable for it?
OpsProvide model or system documentation: intended uses, known limitations and training data categories.
LegalHow do you test accuracy and reliability for our use case? Share results.
OpsHow do you test for bias and unlawful discrimination? How often, and with what results?
LegalWhat human oversight do you recommend, and how does the product support it (overrides, explanations)?
OpsCan the product explain an individual output in plain language (inputs, factors, result)?
LegalHow do you monitor for drift, hallucination or harmful output after release?
OpsHow do you handle AI incidents, and will you notify us of ones affecting our use?
OpsAI vendor questionnaire (4 of 4)
Is any feature used to make or substantially shape decisions about jobs, credit, housing, insurance, health care or education?
LegalHow will you support California ADMT notices, opt-outs and access requests from January 1, 2027?
LegalWill you provide Colorado SB 26-189 developer documentation and notice of material updates?
LegalUnder the EU AI Act, are you a provider of any high-risk or general-purpose AI system we’d use?
LegalDo chatbot or generated-content features support AI disclosure and labeling requirements?
LegalHave you received regulatory inquiries, enforcement actions or lawsuits about your AI or privacy practices?
LegalDo you align with a framework such as NIST AI RMF or ISO/IEC 42001?
OpsWhich sector rules do you support (HIPAA, GLBA, FERPA, COPPA)? Will you sign related agreements?
LegalWill you commit in the contract to no training on our data and to purpose limitation?
LegalWho owns outputs? Do you claim any rights to our inputs or outputs?
LegalDo you indemnify us for IP infringement claims about outputs and for your privacy or security breaches?
LegalWhat are your liability caps, and do they carve out data breaches and confidentiality?
LegalWill you provide technical declarations and records if we face a claim about your product?
LegalWhat insurance do you carry (cyber, technology E&O, media)? Share certificates.
OpsCan we audit or receive independent assessments of compliance with these terms?
LegalHow do we export our data and configurations if we leave, and in what format?
OpsScore the answers
Rate each section 0–3, then check for deal-breakers. A high score never overrides a red flag.
| Score | Meaning | What it looks like |
|---|---|---|
| 3 | Strong | Clear written answer, matching evidence, and a willingness to put it in the contract |
| 2 | Acceptable | Clear answer, partial evidence, or a reasonable commitment with a date |
| 1 | Weak | Vague answer, marketing language, or “available on request” with nothing provided |
| 0 | Fail | No answer, contradictory answers, or refusal |
Deal-breakers for Tier 1
- Trains on your data by default and won’t contract out of it
- Won’t name model providers or other subprocessors
- Uses your data for ads, profiling or resale
- No security report or independent assessment
- Can’t delete data or stop processing for a person
- Can’t explain an individual decision when the tool affects people
Decision bands
Eight sections × 0–3 points = 24 maximum.
Evidence to request
Answers are claims. Evidence is what you’ll show a regulator, auditor or court.
| Evidence | What to check | Tier |
|---|---|---|
| SOC 2 Type II report | Current period, scope includes the AI features, exceptions noted | 1, 2 |
| ISO 27001 / ISO 42001 | Certificate scope and expiry; 42001 covers AI management | 1 |
| Subprocessor list | Names model providers; dated; change-notice process | 1, 2 |
| Model provider terms | No training, retention limits, abuse monitoring settings | 1 |
| Model or system card | Intended uses, limitations, training data categories | 1 |
| Accuracy and bias testing | Method, dates, results relevant to your use case | 1 |
| Penetration test summary | Recent, includes AI features, critical findings closed | 1 |
| Data processing agreement | Service provider and processor terms; deletion; audits | 1, 2 |
| Data-flow description | Where inputs go, who can access them, encryption | 1 |
| Insurance certificates | Cyber and tech E&O limits that fit your exposure | 1 |
Test it yourself
Check admin settings for training toggles, retention and default AI features. What the product does should match what the questionnaire says.
Store answers, evidence, scores and approvals in one vendor file. It doubles as the input to your risk assessments.
The AI contract addendum
Turn the good answers into obligations. These are the terms we see matter most.
| Term | What it should say |
|---|---|
| No training | Vendor and its subprocessors won’t use customer data, inputs or outputs to train or improve models for anyone else |
| Purpose limitation | Processing only to provide the service to you; no ads, profiling, sale or sharing |
| Subprocessors | Current list including model providers; advance notice of changes; right to object |
| Retention and deletion | Defined periods for inputs, outputs and logs; deletion on request and at exit, with confirmation |
| Human review | No human access to content except for support or security, with logging |
| Model changes | Notice of material model or feature changes; right to re-test or disable |
| Regulatory cooperation | Documentation and help for risk assessments, DPIAs, ADMT, Colorado and EU AI Act duties |
| Rights requests | Help with access, deletion, correction and opt-outs within set timelines |
| Incidents | Notice of security and AI incidents within a defined number of hours |
| Outputs and IP | You own outputs; vendor indemnifies for IP claims about outputs |
| Litigation support | Technical declarations and records if you face a claim tied to the product |
Colorado note
Under SB 26-189, contract terms that indemnify a party against its own anti-discrimination violations involving covered ADMT are void as against public policy. Draft indemnities accordingly.
Monitor after you sign
AI products change faster than contracts. Set triggers that send a vendor back for review.
Re-review triggers
- New AI feature, especially one on by default
- New or changed model provider
- Changes to terms of service or privacy policy
- New use case or new data type inside your company
- Security incident, lawsuit or regulatory action
- Acquisition of the vendor
- Contract renewal
Find shadow AI
- Review expense reports and card spend for AI tools
- Check SSO and browser extension logs
- Scan websites for chat and AI scripts
- Ask teams quarterly what AI they use
- Offer approved tools so people don’t go around you
Where each vendor type goes wrong
Watch: training on conversations, model provider retention, missing chat notices. Ask first: B1, C2, E6.
Watch: auto-joining calls, recording without notice, training on recordings. Ask first: B1, E6, A6.
Watch: scoring candidates without explanation, bias testing gaps. Ask first: F4, F6, G1–G3.
Watch: explainability for adverse actions, data sources. Ask first: F2, F6, G1.
Watch: confidential data and source code in prompts, output IP. Ask first: B1, D3, H2.
Watch: using customer lists for vendor products, pixels, opt-outs. Ask first: B5, E3, E7.
Watch: broad system permissions, actions without approval, weak logs. Ask first: D4, D5, F5.
Watch: features switched on by default, new subprocessors. Ask first: A1, A5, C4.
The common thread
In every category the same three questions catch most problems: training, subprocessors and retention. Ask them before the demo, not after the signature.
Thirty days to a working program
You don’t need new software to start. You need a list, a form and a habit.
Is your AI vendor program ready?
Tick each item only if you can show evidence for it.
Visibility
Review
Contract
Settings
Ongoing
Questions we hear most
Start with sections A–E and the deal-breakers. Large vendors often publish trust centers that answer many questions; map their documents to your list.
Yes, if employees put company or personal data into them. Free tiers often have broader training rights than paid plans.
It covers part of it. Security reviews rarely ask about training, model providers, explanations or bias, which drive AI risk.
Someone with authority to accept the risk for the business, documented in writing, with privacy and legal input.
Yes. Vendor answers are a main input to CCPA risk assessments, DPIAs and Colorado records. Collect them once, use them everywhere.
Sources
Current as of November 2026. General information, not legal advice. The questionnaire may be copied and adapted for your own vendor reviews.
Know every AI tool
touching your
customers’ data.
Captain Compliance helps you find the AI and tracking vendors on your sites, manage vendor reviews and disclosures, and keep the records regulators ask for.
Discover
Scan sites for AI, chat and tracking vendors.
Disclose
Vendor disclosures that stay current.
Document
Records ready for audits and assessments.
captaincompliance.com
By Richart Ruddie, Founder & CEO
Not legal advice. Current as of November 2026.
Put this into practice
Turn privacy guidance into working controls.
Captain Compliance helps teams discover tracking technologies, enforce consent, manage privacy requests and document the evidence behind every decision.
