Short answer: the SIG AI domain is a block of AI questions inside the Shared Assessments SIG, a licensed vendor questionnaire that a buyer sends you. The CSA AI-CAIQ is a free self-assessment that maps to the Cloud Security Alliance's AI Controls Matrix, and you can complete it yourself and publish it. Both ask how you govern, secure and monitor AI, and both map to ISO/IEC 42001 and the NIST AI Risk Management Framework. Neither is a certificate. The practical move is to build one evidence pack (inventory, data flows, provider terms, tests, logs, policies, incident plan) and answer both forms from it, so you do not rewrite the same facts for every buyer.
This guide explains what each one is, how they differ, what they ask, and how to prepare evidence. It does not repeat the question-by-question sample answers in our post on how to answer the AI section of a security questionnaire. Read that for wording. Read this to understand the two frameworks behind the forms.
What are the SIG AI domain and the AI-CAIQ?
The SIG (Standardized Information Gathering questionnaire) is published by Shared Assessments. Shared Assessments says the SIG is updated every year and lists 21 risk domains, one of which is Artificial Intelligence. It maps to ISO/IEC 42001 and NIST AI 100-1, which is the formal number of the NIST AI RMF 1.0. The SIG is licensed, not free. Third-party risk teams buy it, load it into their tooling and send it to vendors.
The AI-CAIQ (Consensus Assessment Initiative Questionnaire for AI) comes from the Cloud Security Alliance. It sits inside the CSA AI Controls Matrix (AICM), which at version 1.1 (released 22 June 2026) has 247 control objectives across 18 security domains and builds on CSA's Cloud Controls Matrix. CSA describes the AI-CAIQ as a set of questions that map to the AICM and can guide a self-assessment or a review of third-party vendors. The AICM is free to download.
The two tools have different jobs. The SIG is a buyer's instrument: the customer chooses the scope and you answer. The AI-CAIQ can be either: a buyer can send it, or you can complete it proactively and share or publish it.
| SIG (AI domain) | CSA AI-CAIQ | |
|---|---|---|
| Owner | SIG (AI domain): Shared Assessments | CSA AI-CAIQ: Cloud Security Alliance |
| Cost to obtain | SIG (AI domain): Licensed. Buyers license it and send it to you | CSA AI-CAIQ: Free to download from CSA |
| Who starts it | SIG (AI domain): Almost always the buyer | CSA AI-CAIQ: Buyer or vendor |
| Scope | SIG (AI domain): AI is one of 21 risk domains in a wider vendor questionnaire | CSA AI-CAIQ: AI-specific. Questions map to AICM controls (247 objectives, 18 domains) |
| Standards mapping | SIG (AI domain): ISO/IEC 42001, NIST AI 100-1 | CSA AI-CAIQ: ISO/IEC 42001, ISO/IEC 27001, NIST AI RMF and AI 600-1, EU AI Act, BSI AIC4, AIUC-1 |
| Output | SIG (AI domain): A completed workbook for that buyer | CSA AI-CAIQ: A self-assessment you can reuse, or submit to the STAR Registry |
| Certificate? | SIG (AI domain): No | CSA AI-CAIQ: No. Self-assessment only, unless you pursue a separate third-party route |
What does the SIG AI domain ask?

Because the SIG is licensed, we do not reproduce its questions here, and you should check your licence or your buyer's terms before copying them into a public page or a shared answer library. What can be said safely comes from Shared Assessments' own description and from licensees' summaries.
- It is part of a larger form. AI sits alongside the other risk domains. Buyers choose how much of the SIG to send. Licensees describe scoping presets called Lite, Core and Detail, which let a buyer match depth to the vendor's risk. That explains why one buyer sends a short AI section and another sends the full set.
- It follows the AI lifecycle. A licensee summary of the current SIG says AI questions cover the whole lifecycle, from data collection and model training to deployment and bias monitoring, and that content maps to ISO/IEC 42001. That is a secondary source, so treat it as a guide to themes, not as the exact wording.
- It asks for governance as well as security. Expect questions about who owns AI risk, whether you keep an inventory of AI systems, how you assess models and providers, and how you monitor outputs, alongside the usual data protection questions.
- The edition changes. The SIG is updated annually, so ask the buyer which version they are sending and answer that one, not last year's copy.
Do not guess the year the AI domain first appeared or how many questions it has. If a buyer sends you a SIG, the file in front of you is the only reliable source for wording and counts.
What does the CSA AI-CAIQ ask?
The AI-CAIQ asks how your AI service implements each AICM control. CSA's STAR page says the AI-CAIQ captures how your AI service implements each AICM control, and that it is the submitted artifact for STAR for AI. The structure matters more than the question count, which we have not verified.
Three features of the AICM change how you answer:
- Control ownership. The AICM is built around shared responsibility. The AICM names five roles: model provider, orchestrated service provider, application provider, AI customer and cloud service provider. The download includes implementation and auditing guidelines. A SaaS company that calls a third-party model API and adds its own features is usually an application provider, and in some products an orchestrated service provider. Decide your role first, because it decides which controls are yours and which sit with the model vendor or the cloud host.
- Five analysis pillars. The control spreadsheet classes each objective by control type, applicability and ownership, architectural relevance, LLM lifecycle relevance and threat category. This helps you filter out controls that do not apply to your architecture.
- Completion guidance. The download includes CSA's guidance on completing the AI-CAIQ, plus a STAR for AI Level 1 submission guide. Read them before you answer, because they set out what CSA expects a supported answer to look like.
The download also includes mappings to ISO/IEC 42001, NIST AI RMF and AI 600-1, the EU AI Act, BSI AIC4 and AIUC-1, and CSA lists a machine-readable bundle as a related resource. If your buyers cite one of those, you can show how an AICM control lines up with it.
What is STAR for AI, and does the AI-CAIQ lead to a certificate?
STAR is CSA's assurance programme, and STAR for AI is its AI version. On CSA's STAR page, the AI-CAIQ is the submitted artifact for STAR for AI Level 1 and Level 2. The levels differ in who checks the answers.
| Level | What it is | Who checks it |
|---|---|---|
| STAR for AI Level 1 | What it is: Self-assessment. You submit the AI-CAIQ directly, or through Valid-AI-ted for automated scoring and a badge on your Registry listing | Who checks it: You. CSA's Valid-AI-ted tool returns scoring and improvement guidance on the submission |
| STAR for AI Level 2 | What it is: Third-party assurance, paired with an AI-CAIQ passed through Valid-AI-ted. CSA lists routes through ISO/IEC 27001, ISO/IEC 27001 plus 42001, ISO/IEC 42001 and SOC 2 | Who checks it: An independent auditor on that route |
So a Level 1 listing is a public self-declaration. It is useful, because procurement teams can find it and reuse it, but it is not independent assurance. CSA's STAR page announces that Level 2 is now available, while the AICM download page still describes the STAR for AI programme as an upcoming certification. Check CSA's current STAR for AI pages for what each level involves before you plan around it. Mono is not a certification body or an auditor and cannot issue any of these. Independent auditors on the Level 2 routes do that work.
Which one should you complete first?

Use the buyer's request to decide, then use your own situation to decide the rest.
- A buyer sent a SIG or an AI-CAIQ. Answer what was sent, in their format. Do not swap in your own form. Ask which edition and which scope.
- Several buyers are asking, in different formats. Complete the AI-CAIQ once as your master record. It is free, it is AI-specific and it maps to the same standards the SIG uses. Then answer each SIG from it.
- You sell to large enterprises or regulated buyers. Consider publishing an AI-CAIQ at STAR for AI Level 1 once your evidence is solid. A public listing is a statement others will rely on, so have counsel review it before you submit.
- Nobody has asked yet. Start with the evidence pack below. It is useful for any form, including a buyer's own spreadsheet, and costs you little.
If the question is whether you need ISO/IEC 42001 certification as well, read how ISO 42001 compares with SOC 2 and our ISO 42001 readiness page. Both forms map to ISO/IEC 42001, but neither requires you to hold the certificate. Answering a mapped question does not mean you are certified.
What evidence should you prepare for both?
Most AI questions in both frameworks come down to the same set of documents. Build them once, date them and give each an owner.
| Evidence | What it proves | Typical question area |
|---|---|---|
| AI system inventory | What it proves: Every feature, model and provider in use, with owner and purpose | Typical question area: AI governance, inventory, accountability |
| Data flow diagram per AI feature | What it proves: What data goes to which model, where it is processed and stored, and what comes back | Typical question area: Data handling, residency, third parties |
| Model provider terms and DPA | What it proves: Whether your data trains the provider's models, retention, subprocessors, region options | Typical question area: Third-party risk, training on customer data |
| Training and fine-tuning statement | What it proves: What data, if any, you use to train or tune, and the customer opt-out | Typical question area: Data use, model development |
| AI use policy | What it proves: Approved tools and uses, banned uses, who approves a new AI feature | Typical question area: Policy, governance |
| Risk or impact assessment | What it proves: How you assess an AI feature before launch and who signs it off | Typical question area: Risk management, lifecycle |
| Test and evaluation records | What it proves: Prompt injection tests, output quality checks, bias or abuse checks, with dates and results | Typical question area: Security testing, monitoring |
| Human oversight design | What it proves: Where a person reviews or can override AI output, and how that is logged | Typical question area: Oversight, safety |
| Logging and monitoring sample | What it proves: What you log (prompts, outputs, model version, user), retention, access | Typical question area: Monitoring, audit trail |
| Change log for models and prompts | What it proves: Which model version ran when, and how changes are approved | Typical question area: Lifecycle, change management |
| Tenant isolation design | What it proves: How one customer's data and context stay separate in retrieval, caching and fine-tuning | Typical question area: Access control, multi-tenancy |
| Incident runbook with AI scenarios | What it proves: Data leak through a model, harmful output, provider outage, who is called and when | Typical question area: Incident response |
| Existing audit reports | What it proves: Your ISO/IEC 27001 certificate or SOC 2 report where you hold one, with scope | Typical question area: General security, supporting evidence |
Keep these in one folder with an index. When a buyer asks a question, you point to the document and the owner confirms the answer is still true. The nine question groups, with weak and strong sample answers, are in our security questionnaire guide, and our security questionnaire support page shows how we help when a form is already open.
How do you answer so the evidence holds up?
- Answer for your role. Under the AICM, state whether a control is yours, shared or the provider's. Do not claim a provider's control as your own, and do not leave it blank either. Say who owns it and attach the provider's document.
- Use plain answer states. Yes, no, partial or not applicable, each with a one-line explanation and an evidence reference. A bare yes invites a follow-up round.
- Date every answer. Store the date it was last checked against the code and configuration. Re-check when you ship an AI feature or change a model provider.
- Do not paste marketing copy. Reviewers look for specifics: product names, settings, retention periods, test dates.
- Say what is planned as planned. Give the owner and the date. A gap you describe honestly is easier to approve than a claim you cannot prove.
- Keep one source of truth. Store the master answers in one place and generate each buyer's response from it. Several slightly different answers across forms invite questions.
What mistakes cause follow-up rounds?
- Reading "AI" as "generative AI" only. Define AI broadly in your inventory, so classifiers, scoring models and other AI you build or buy are counted too.
- Forgetting AI inside other vendors' products, such as a support tool or a code assistant your staff use. A shadow AI audit finds these before a buyer does.
- Treating the ISO 42001 mapping as proof. A mapping tells a buyer which control a question relates to. It does not show that you meet it.
- Publishing a Level 1 listing and then changing your architecture without updating it. A stale public listing is worse than none.
- Answering from memory. Check configuration, contracts and logs, then answer.
What does this mean for companies in India and the US?
The frameworks are international, and the same forms arrive from US enterprises, European buyers and Indian banks and large corporates. For Indian companies the difference is usually in the surrounding evidence: where data is hosted, which subprocessors are abroad and how personal data is handled under the Digital Personal Data Protection Act, 2023. Keep that material in the same pack. Neither framework replaces legal advice on data protection or sector rules. Take those questions to counsel.
Where does Mono help?
Mono builds the evidence pack with engineering teams: the AI inventory, data flow diagrams, provider-term reviews, test records, logging and a master answer library. We can prepare your AI-CAIQ draft or fill a SIG your buyer sent, and map answers to ISO/IEC 42001 and the NIST AI RMF. Mono is not a certification body, a CPA firm or a law firm. We prepare evidence. CSA, accredited auditors and your counsel do the rest. Start with AI governance if you want the wider programme, or send us the form that is blocking a deal.
Sources: Shared Assessments, SIG overview (sharedassessments.org/sig-excel). Mitratech, "SIG 2026: Key Updates and Considerations", 5 March 2026 (secondary). Cloud Security Alliance, AI Controls Matrix v1.1, released 22 June 2026, and v1.0, 9 July 2025 (cloudsecurityalliance.org/artifacts/ai-controls-matrix-v1-1). Cloud Security Alliance, STAR Program page, accessed 20 November 2026. NIST, AI Risk Management Framework (nist.gov/itl/ai-risk-management-framework). ISO/IEC 42001:2023, Artificial intelligence management system. Last reviewed 20 November 2026.
FAQ
What is the SIG AI domain?
It is the Artificial Intelligence risk domain in the Shared Assessments SIG, a licensed vendor risk questionnaire that buyers send to suppliers. Shared Assessments lists 21 risk domains in the SIG, updates it every year, and maps the content to ISO/IEC 42001 and NIST AI 100-1. Ask your buyer which edition they are sending.
What is the CSA AI-CAIQ?
The AI-CAIQ is the Cloud Security Alliance's Consensus Assessment Initiative Questionnaire for AI. It is a set of questions that map to the CSA AI Controls Matrix, which has 247 control objectives across 18 domains in version 1.1. Vendors use it to self-assess, and buyers use it to review vendors. It is free to download.
Is the AI-CAIQ a certification?
No. The AI-CAIQ is a self-assessment. Submitting it to the CSA STAR Registry gives you a STAR for AI Level 1 listing, which is a public self-declaration. Level 2 adds third-party assurance through routes CSA lists, such as ISO/IEC 42001 or SOC 2. Neither level is issued by Mono or a consultant.
Do I need ISO 42001 to answer the SIG AI domain or the AI-CAIQ?
No. Both frameworks map their questions to ISO/IEC 42001, but they do not require you to hold the certificate. You answer based on what your systems and documents show. Holding ISO/IEC 42001 or ISO/IEC 27001 certificates can support your answers if the scope covers the AI feature in question.
Which should I complete first, the SIG or the AI-CAIQ?
Answer whichever the buyer sent, in their format. If several buyers ask in different formats, complete the free AI-CAIQ once as a master record, then answer each SIG from it. Both ask about the same underlying evidence: inventory, data flows, provider terms, testing, logging, oversight and incident response.
Can I copy SIG questions into my own answer library?
Check first. The SIG is licensed by Shared Assessments, so the licence or your buyer's terms decide what you may copy or publish. A safer approach is to keep answers organised by topic, such as data use or logging, and map them to each form's question numbers privately.
