Bottom Line
Vertiva's core compliance thesis is simple: compliance is a continuous, cross-cutting program that starts on day one, not a feature bolted on at the end of the roadmap. Every phase of the platform's architecture is required to extend the same control baseline — encryption, access control, audit logging, data lineage, and residency — so certification evidence is accumulating from the first chat query, not assembled retroactively before a sale.
A critical key for data sovereignty: moving AI inference in-house (self-hosted models) means customer content — including PHI/PII — never has to leave the customer's own environment. That one architectural decision removes an entire subprocessor from the compliance surface. The platform is hosted in the customer's enterprise cloud platform, so all existing data controls remain intact and secure.
Certification & Framework Coverage
| Framework | Applies To | What It Requires |
|---|---|---|
| SOC 2 Type II | All enterprise customers | Continuous evidence across Security, Availability, Processing Integrity, Confidentiality, and Privacy over a 6–12 month observation window |
| HIPAA | Healthcare customers handling PHI | Administrative, physical, and technical safeguards, plus signed Business Associate Agreements (BAAs) with the platform and every subprocessor — before any PHI is processed |
| ISO 27001 | Optional certification | Information security management controls; substantially overlaps the SOC 2 control set |
| ISO 42001 | AI-specific governance | AI management system controls — model governance, risk management, lifecycle oversight |
| EU AI Act | AI risk governance | Risk classification alignment for AI systems in scope |
| NIST AI RMF | AI risk governance | Structured risk-management framework for AI system behavior and change |
| GDPR / UK-GDPR | EU/UK personal data | Data-subject rights (including right-to-be-forgotten), lawful basis, DPIAs, residency |
Forward-looking, scope-dependent: FedRAMP, PCI-DSS, and HITRUST are tracked for future public-sector, payment, or healthcare-composite scope.
Data Sovereignty Is the Differentiator
Self-hosted inference is the platform's single biggest compliance lever. When generation and embedding models run inside the customer's own boundary rather than calling a third-party AI API, customer content — including PHI/PII — never leaves that boundary. This removes the AI vendor as a subprocessor entirely, collapsing the HIPAA and GDPR surface area in one architectural move rather than through incremental controls.
- Off by default: Under HIPAA scope, any bring-your-own-model option that would send customer data off-boundary (a customer's own hosted endpoint, provider API key, or private weights) is disabled by default.
- Gated, not blocked: it is only enabled after a signed BAA and a formal data-egress review — so a regulated customer never has an unreviewed path for PHI/PII to leave its environment.
- Change control on model selection: this is treated as governed, capacity-planned change, not a self-service toggle, because it touches customer trust boundaries directly.
Compliance Starts on Day One
Compliance is explicitly not scheduled as a late-stage "governance phase." A control baseline is established as the platform's foundational build closes, so the SOC 2 evidence clock — which requires controls to be operating over a 6–12 month window — starts as early as possible rather than at the point of sale. Every subsequent phase of platform work is required to extend that baseline without regression: encryption, access control, audit logging, tenant isolation, data residency, and cost/compliance evidence are non-negotiable acceptance criteria on every phase, not optional hardening.
Each of the platform's regulated capabilities is delivered and then deployed into a live customer environment as a distinct rollout step — provisioning, identity federation, data onboarding, and a dedicated security & compliance sign-off (BAA execution, SOC 2 report sharing, penetration test, audit support) before that customer goes live. Compliance sign-off is a gating, repeatable step in every customer rollout, not a one-time project milestone.
Governance Capabilities Built on Top of the Controls
Beyond certification, the platform is building the operational governance layer that lets a regulated enterprise trust — and prove — what its AI is doing:
- Lineage & grounding: every AI response can be traced end-to-end from source document through retrieved context to the final answer, and no answer is served without a citation back to its source.
- Guardrails: input/output policy checks for prompt injection, toxicity, and PII egress, plus per-tenant topic controls, implemented as enforced checkpoints rather than best-effort filters.
- Right-to-be-forgotten: a verifiable deletion cascade removes a document from every store and cache it touched — supporting HIPAA and GDPR data-subject rights.
- Model governance: model cards, approval workflows, version pinning, and deprecation policy govern which models are in production and who approved them.
- Audit & evidence automation: an immutable audit trail and full request tracing generate the evidence packs that satisfy SOC 2 and HIPAA audit controls on demand, rather than requiring a manual evidence-gathering exercise.
- Responsible-AI testing: bias/fairness evaluation and an adversarial red-team harness run continuously against the production model set.
Supporting working-notes priorities reinforce the same posture: data masking to strip PII/PHI/IP before it ever reaches an embedding model or the LLM, human-in-the-loop review for the most regulated workflows, and a standing commitment to zero data retention with any third-party model provider so customer prompts are never used for third-party training.
Shared Responsibility, Clearly Divided
- Vertiva delivers the platform-level controls and the evidence that proves they operate — encryption, audit trail, tracing, guardrails, lineage.
- Subprocessors (cloud infrastructure and compute providers) are covered by BAAs and DPAs; self-hosting inference removes the AI subprocessor from this list entirely.
- The customer owns its own data classification, user provisioning, acceptable-use policy, and its broader regulatory posture.