CAIQ 4.0’s 47 AI Questions: What SaaS Vendors Need to Know

The Cloud Security Alliance quietly made one of the most consequential moves in vendor compliance this year. They added 47 AI-specific questions to the Consensus Assessments Initiative Questionnaire — CAIQ 4.0.

If you’ve been through a SOC 2 audit or filled out a SIG questionnaire, you know CAIQ. It’s the standardized self-assessment that cloud providers use to document their security posture for enterprise buyers. Microsoft, Amazon, and Google submit CAIQ responses. When enterprise procurement teams need to evaluate your AI practices, CAIQ 4.0 is increasingly where they look.

47 new questions sounds overwhelming. It’s not — once you understand the structure.

---

What Changed in CAIQ 4.0

Previous CAIQ versions treated AI as an afterthought. A few questions about automated decision-making buried in the governance section. That was it.

CAIQ 4.0 treats AI as a first-class domain. The 47 new questions span AI lifecycle management, data governance for AI systems, model risk, transparency, and incident response. They align heavily with NIST AI RMF and reference ISO 42001 concepts — which means if you’ve already mapped your controls to those frameworks, you’re ahead of the game.

The questions aren’t theoretical. They’re operational: “Do you maintain an inventory of AI systems?” “How do you validate training data quality?” “What human oversight exists for AI-driven decisions?” Enterprise buyers are adopting CAIQ 4.0 because it gives them a standardized way to evaluate AI vendors without inventing questions from scratch.

---

The 47 Questions, Grouped by What Actually Matters

I’ve organized the 47 questions into five practical categories based on what I see buyers caring about most. This isn’t the CSA’s official taxonomy — it’s a vendor-side prioritization based on what stalls deals and what doesn’t.

Category 1: AI Inventory and Classification (9 questions)

These are the foundation. Buyers want to know what AI you’re running and how you categorize it.

Key questions include:

• Do you maintain a comprehensive inventory of all AI and ML systems?
• How are AI systems classified based on risk level and impact?
• Do you document the purpose, scope, and limitations of each AI system?
• What process exists for approving new AI system deployments?

Why this matters most: Without an inventory, every answer to every other question is theoretical. I’ve reviewed companies that use four different AI providers across eight repositories and can’t name them all. You can’t govern what you haven’t enumerated.

Strong answer pattern: “We maintain a living AI registry documenting 4 providers, 7 model integrations, and 3 risk classifications across 2 repositories. Registry updates are triggered by CI/CD pipeline changes.”

Weak answer pattern: “We track our AI usage as part of our general asset management process.”

Category 2: Data Governance for AI (12 questions)

The largest block — and the one where most vendors have the biggest gaps.

Key questions include:

• How is training data sourced, validated, and documented?
• What data classification applies to AI input/output data?
• Do AI providers have access to customer data?
• What data retention policies apply to AI processing?
• Are Data Processing Agreements in place with all AI providers?

Why this matters: This is where deals die. An enterprise buyer in financial services or healthcare needs specific answers about data flow — not policies, but architecture. What data hits the API? Is PII stripped? Where does the response get stored? For how long?

Strong answer pattern: “Customer data is classified at ingestion. PII is pseudonymized before AI processing. Our Anthropic DPA (signed January 2026) specifies zero training on customer data, 30-day response deletion, and SOC 2 Type II subprocessor compliance.”

Weak answer pattern: “We follow industry best practices for data handling.”

Category 3: Model Risk and Validation (10 questions)

These questions probe whether you actually test your AI systems or just deploy and hope.

Key questions include:

• How do you validate AI model outputs for accuracy and reliability?
• What bias testing or fairness evaluation have you conducted?
• How do you monitor for model drift or degradation over time?
• What rollback procedures exist for AI model failures?

Why this matters: Honest answers win here. Most startups haven’t done formal bias testing — and that’s okay, as long as you have a plan. “We haven’t conducted formal bias evaluation but have scheduled quarterly reviews starting Q2” is infinitely better than “we haven’t identified any bias issues.” One shows maturity. The other shows blindness.

Category 4: Transparency and Explainability (8 questions)

The EU AI Act is pushing transparency requirements into mainstream vendor questionnaires.

Key questions include:

• Are users informed when they interact with AI-generated content?
• Can you explain how AI-driven decisions are made?
• What documentation exists for AI system behavior and limitations?
• How do you handle user requests for AI decision explanations?

Why this matters: Transparency questions used to feel academic. They’re not anymore. The EU AI Act requires transparency disclosures for certain AI categories, and enterprise buyers are building those requirements into procurement. Even if your buyer is US-based, their customers may be in the EU.

Category 5: AI Incident Response and Governance (8 questions)

The operational questions that separate companies with AI governance from companies with AI governance documents.

Key questions include:

• Do you have AI-specific incident response procedures?
• How are AI-related incidents classified and escalated?
• What governance structure oversees AI deployment decisions?
• How do you handle AI system outages or provider failures?

Why this matters: A general incident response plan isn’t enough. When a model hallucinates customer data, when a prompt injection bypasses input filtering, when your AI provider has a 4-hour outage — each of these needs a defined response path with severity levels and notification timelines.

---

How CAIQ 4.0 Maps to Other Standards

CAIQ 4.0 doesn’t exist in isolation. Enterprise buyers often ask you to map your responses across multiple frameworks simultaneously.

CAIQ 4.0 Domain	NIST AI RMF	ISO 42001	EU AI Act
AI Inventory	MAP 1.1–1.6	A.6.2	Art. 6, Annex III
Data Governance	GOVERN 1.3, MAP 2.3	A.7.4, A.8.4	Art. 10
Model Risk	MEASURE 2.1–2.6	A.9.2	Art. 9, Art. 15
Transparency	GOVERN 4.1, MAP 5.1	A.8.3	Art. 13, Art. 52
Incident Response	MANAGE 4.1–4.3	A.10.2	Art. 62

This cross-mapping is why a structured approach matters. Answer CAIQ 4.0 well, and you’ve covered 70-80% of what NIST AI RMF and ISO 42001 questionnaires will ask. Answer in isolation, and you’ll repeat the same work three times.

Our pillar guide covers the full cross-framework mapping.

---

Where Most Vendors Get Stuck

After working with startups navigating enterprise AI questionnaires, I see three consistent failure points with CAIQ 4.0:

Treating it like SOC 2. SOC 2 is a binary — you have it or you don’t. CAIQ is a self-assessment. There’s no auditor giving you a pass/fail. That means the quality of your answers is the product. Buyers can tell the difference between a thoughtful response with evidence and a copy-pasted template.

Answering in isolation. Vendors answer each of the 47 questions independently, which leads to inconsistencies. Question 12 says you retain data for 30 days. Question 34 says your DPA specifies 90-day deletion. Now the buyer is flagging contradictions instead of approving your response.

Missing the evidence layer. CAIQ asks what you do. Enterprise buyers want to see that you do it. An answer that says “we use secrets management for AI API keys” is good. An answer that says “we use secrets management — vault-stored credentials with automatic rotation, no hardcoded keys detected in our most recent codebase scan” is better.

---

How to Prepare Without Drowning

You don’t need to answer all 47 questions from scratch. Here’s the practical approach:

Start with inventory. If you can answer the 9 inventory questions completely — every AI provider, model, version, purpose, and data classification documented — the remaining 38 become significantly easier. Most of those answers flow directly from knowing what AI you use and how.

Map your DPAs. Check whether you have a Data Processing Agreement with every AI provider you use. Not just the major ones. The one developer who added a Groq integration for internal tooling? That needs a DPA too. Our PROVIDER_FACTS database tracks DPA availability across 9 major providers.

Cross-reference frameworks. If you’ve already done work against NIST AI RMF or the SafeBase 11 questions, you’ve already answered a significant portion of CAIQ 4.0. Map what you have before writing new responses.

Be specific, be honest. “We plan to implement bias testing in Q2” is a perfectly acceptable answer. “We have comprehensive bias testing processes” when you don’t is not. Enterprise buyers are experienced enough to distinguish between a company that’s building governance and one that’s making things up.

---

Key Takeaways

• CAIQ 4.0 added 47 AI-specific questions across five domains — this is the new standard for cloud provider AI assessments
• The questions align with NIST AI RMF and ISO 42001, so framework mapping gives you leverage across multiple questionnaires
• Start with AI inventory (9 questions) — everything else flows from knowing what you use and how
• Data governance (12 questions) is the largest block and where most deals stall — DPAs, data flow, retention, and provider training policies
• Evidence-based answers beat policy-based answers every time — buyers want to see what you actually do, not what your policy says you should do
• Our Trust Pack and Response Kit cover all 47 CAIQ AI domains from your actual codebase — download the free AI Due Diligence Question Bank to see how the questions map