What it means
- Before buying or switching on AI from a vendor (an LLM API, a credit scoring service, an AI feature in a CRM), check it is safe, lawful and fit for purpose.
- Ask AI-specific questions on top of normal vendor security checks: what data trained the model, is your data used to train it, where is data processed, how is it tested for bias, what happens when the model is updated, can you get logs and explanations.
- Put the answers into contracts: data use limits, audit rights, change notice, incident notice, liability split.
- Also covers staff use of free or off-the-shelf GenAI tools, which must go through the same approval.
- If you sell AI, be ready to answer the same questions from your customers.
Why it matters
- RBI FREE-AI says staff use of third-party AI tools must fall under the internal AI policy, and REs stay liable for outsourced AI.
- Most mid-size companies buy more AI than they build; vendor risk is often the biggest AI risk.
- Government and enterprise buyers (and Tamil Nadu's policy) expect vendors to hand over an evidence pack.
Checklist
- Add an AI questionnaire to vendor onboarding (training data, data reuse, hosting location, testing, bias, explainability, model change process, sub-processors).
- Tier vendors using SC-AI-004 AI risk classification and use-case tiering; deeper review for high risk uses.
- Add AI clauses to contracts and DPAs: no training on your data without consent, notice of material model changes, incident notification timelines, audit or assurance rights, exit support.
- Maintain an approved GenAI tools list; block or monitor unapproved tools where feasible.
- Review high risk AI vendors at least annually (refresh questionnaire, check SOC 2 or ISO 42001 reports).
- Prepare your own customer-facing AI evidence pack if you sell AI.
Evidence examples
- Completed AI vendor questionnaire and risk rating for the LLM API provider.
- Signed contract or addendum showing AI-specific clauses (data use, change notice, audit rights).
- Approved GenAI tools list with approval dates and owner.
- Annual vendor review record, including the vendor's latest SOC 2 or ISO 42001 certificate.
- Proxy or CASB report showing blocked unapproved AI sites.
Common pitfalls
- Relying on the vendor's SOC 2 report alone; it rarely covers model behaviour or bias.
- Accepting default terms that let the vendor train on your customer data.
- No notice clause, so a silent model update changes outputs overnight.
- Ignoring AI features quietly added to existing SaaS tools.
Satisfies
- VoluntaryISO/IEC 42001
- BindingEU AI Act
- SupervisoryPRA SS1/23
- VoluntaryNIST AI RMF and GenAI Profile
- BindingNYC Local Law 144 (AEDT)
- SupervisoryNYDFS AI Cybersecurity Letter 2024
- BindingTexas TRAIGA
- BindingKorea AI Basic Act
- SupervisoryOSFI E-23
- VoluntaryRBI FREE-AI
Also relevant to
- ICMR AI in Healthcare Guidelinesno numbered obligations; mapped in its practical notes (liability split in contracts with hospitals)
- NITI Aayog Responsible AI for Allno numbered obligations; mapped in its practical notes (evidence for government buyers)
- Tamil Nadu Safe and Ethical AI Policyno numbered obligations; mapped in its practical notes (vendor evidence pack for the buyer)
Standards crosswalk
Clause references only. Check the standard itself for the exact wording.
ISO/IEC 42001
- A.10.2 Allocating responsibilities
- A.10.3 Suppliers
- A.10.4 Customers
ISO/IEC 27001
- A.5.19 Information security in supplier relationships
- A.5.20 Addressing information security within supplier agreements
- A.5.21 Managing information security in the ICT supply chain
- A.5.22 Monitoring, review and change management of supplier services
- A.5.23 Information security for use of cloud services
SOC 2
- CC9.2
Key takeaways
- Most AI risk arrives through vendors; extend vendor management with AI questions and contract clauses.
- Strong reuse from ISO 27001 supplier controls (A.5.19 to A.5.23) and SOC 2 CC9.2; the AI questionnaire and model-change clauses are the new parts.
- Cover free GenAI tools used by staff, not only formal purchases.