Certifications are necessary, never sufficient. The frontier APIs and platform context engines have closed the certification gap — but every project still requires its own legal, architectural, and contractual review before regulated data touches the system.
EU regulation (2018) governing the processing of personal data of EU residents. Establishes lawful basis, data subject rights, breach notification, and cross-border transfer rules. Fines up to 4% of global annual turnover.
US law (1996) protecting Protected Health Information (PHI). The Security Rule mandates technical, administrative, and physical safeguards. Enforced by HHS Office for Civil Rights.
AICPA framework auditing controls across five Trust Services Criteria: Security, Availability, Processing Integrity, Confidentiality, Privacy. Type I = controls exist on a date. Type II = controls operated effectively over 3–12 months.
27001: Information Security Management Systems. 27701: Privacy Information Management (extends 27001 for GDPR alignment). 42001: AI Management Systems — the world’s first AI-specific standard (2023).
US federal authorization for cloud services. Three impact levels: Low, Moderate, High. High covers data whose loss could cause severe damage to national security, life, or finances. Required for US federal agency deployments.
The two contracts that turn certification eligibility into actual compliance. BAA: the HIPAA contract between you and your vendor. DPA: the GDPR contract defining controller / processor responsibilities. Both are non-negotiable prerequisites, not optional add-ons.
| Provider / Tier | SOC 2 Type II | HIPAA (BAA) | GDPR (DPA) | ISO 27001 | FedRAMP |
|---|---|---|---|---|---|
| Anthropic Claude API (commercial / enterprise tier) | ✓all systems & API | ✓BAA on enterprise | ✓DPA + SCCs | ✓ | ✓High — via AWS / GovCloud |
| OpenAI API (API platform / enterprise) | ✓API + business products | ✓BAA via Healthcare / API | ✓DPA + data residency | ✓27001, 27017, 27018, 27701 | ●via Azure OpenAI |
| Google Gemini / Vertex AI | ✓ | ✓Google Cloud BAA + project flag | ✓EU residency available | ✓+ ISO 42001 for AI | ✓High |
| AWS Bedrock / Bedrock AgentCore | ✓SOC 1, 2, 3 Type II | ✓HIPAA eligible + BAA | ✓ | ✓27001/17/18/701, CSA STAR L2 | ✓High in GovCloud |
| Azure AI Foundry / Azure OpenAI | ✓ | ✓Microsoft Online Services BAA | ✓100+ Azure compliance offerings | ✓ | ✓High |
| Google Vertex AI Agents | ✓ | ✓ | ✓ | ✓ | ✓ |
| Salesforce Agentforce, IBM watsonx, ServiceNow, UiPath, Kore.ai (enterprise tiers) | ✓ | ✓BAA on enterprise | ✓ | ✓ | ●varies by vendor |
| n8n (self-hosted open source) | ●inherits your infra | ●customer responsibility | ●customer responsibility | ●inherits your infra | — |
| Free / Pro / Plus tiers of any provider (ChatGPT Pro, Claude Pro, Gemini Advanced) | — | — | ●limited DPA | — | — |
Certifications apply to commercial / enterprise / business tiers. Consumer plans (Free, Pro, Plus, Advanced) are out of scope — never use them with regulated data, even for prototyping.
An eligible service is not a compliant service until the contract is signed. HHS has fined covered entities for missing BAAs even when no breach occurred. The “no-view” defense does not work.
VPC endpoints, regional residency, encryption keys, retention policy, audit logging, prompt logging settings — all customer-side. The platform gives you the levers; you must pull them, document them, and prove them in audit.
The certification covers the platform, not your application on top of it. Your planner, your agents, your retrieval pipeline, your tracer, your memory store — your responsibility, audited separately. The compliance perimeter has moved up the stack.
Every project starts from scratch with the legal department. Data flow, threat model, contractual chain, regulatory surface, configuration — all project-specific, all reviewed before deployment. The certifications above are the floor that lets the conversation start. They are not the conversation.