Understand exactly what reaches the model

• Privian runs on managed serverless infrastructure. There is no customer-selectable region pinning or data-residency tier today.
• Privian is a privacy layer in front of managed model providers. It does not provide self-hosted inference.
• BYOK is supported. The upstream provider call uses your credential, and the provider's own region settings apply on their side.
• If strict regional residency is a hard requirement, self-hosted inference is the right tool — see Privian vs self-hosted LLMs.

• Your prompt text with detected supported entities replaced by tokens like PERSON_1, EMAIL_1, PHONE_1.
• Model and routing parameters you set (model id, temperature, max tokens, etc.).
• Any free-text in your prompt that does not match a supported entity pattern — Privian does not invent classifications for it.

Placeholders are deterministic within a request — the same value mentioned three times becomes the same token three times, so the model can still reason about co-reference. On the way back, tokens are replaced with the original values inside the gateway before the response returns to your application.

• The original surface form of any entity Privian successfully masked before the outbound call.
• Your BYOK provider credentials themselves — they are used to authenticate the request, not embedded in prompt content.
• Internal request metadata (your Privian API key, account id, billing context).

Privian does not overclaim here. If a value falls outside the supported entity set, it is not masked and will reach the provider unchanged. The supported set is listed on PII masking.

• Your application

  Sees everything

  It is the source of the prompt and the consumer of the rehydrated response. Privian does not change that.

• Privian (data plane)

  Sees the request in memory only

  Detection, masking, the BYOK provider call and rehydration happen in a single in-memory pass. Bodies are not written to storage.

• Upstream provider

  Sees the masked prompt

  OpenAI, Anthropic, Google etc. receive your prompt with supported entities replaced by tokens. Their terms apply from there.

• Privian staff

  Cannot browse raw prompts

  Raw prompt and response bodies are not retained, so there is no operator console for them. Support work uses request ids and sanitized events.

Your application  →  Privian  →  Managed model provider

Privian is:                       Privian is not:
  prompt-level data protection      governance / policy engine
  privacy-first routing             self-hosted inference
  BYOK + no raw-prompt retention    prompt-injection / jailbreak defense
  small JSON contract               audit-log product

• Customer support AI where messages contain names, emails, account ids and free-text from end users.
• Internal copilots that draft from tickets, HR records, support transcripts or CRM notes.
• SaaS AI features built on top of managed providers (GPT, Claude, Gemini) where prompts reflect customer data.
• Privacy-sensitive workflows where you want one place to enforce masking before a managed model sees a prompt.

• Workloads where regulation or policy requires that no data ever leaves a controlled environment — that is a self-hosted inference problem, not a gateway problem.
• Use cases where the primary requirement is prompt-injection or jailbreak defense — that is a separate product category from prompt privacy.
• Teams that need full governance tooling (per-tenant AI off-switches, fine-grained policy engines, end-to-end audit logging) — Privian is a focused privacy layer, not a governance platform.
• Drop-in replacement for the OpenAI SDK — Privian exposes a small JSON contract, not SDK-level compatibility today.

• Self-hosted models

  Strong isolation, higher operating cost

  Best when policy or regulation requires that data never leaves a controlled environment. You take on inference operations, capacity planning and model maintenance.

• Managed models + Privian

  Managed-model quality with prompt-level masking

  Use GPT, Claude or Gemini, but mask supported sensitive values before the prompt reaches them, keep BYOK, and avoid raw-prompt retention at the gateway.

• Hybrid

  Route by sensitivity

  Highly sensitive workflows go to self-hosted inference; lower-risk workflows use managed providers through Privian. Common pattern for teams that want both.

Full side-by-side: Privian vs self-hosted LLMs.

• Detects and masks supported entity types only; custom user-defined entity types are not supported yet.
• No tool / function calling in the current beta.
• No native provider token streaming yet — stream: true returns artificially chunked text.
• No claim to detect or block prompt injection or jailbreaks.
• No HIPAA / SOC 2 / PCI certifications at this time.
• No tenant-level governance controls, audit log product or per-tenant AI off-switch.

What reaches the model when I call Privian?

The prompt you sent with supported personal and sensitive values replaced by deterministic placeholders (PERSON_1, EMAIL_1, …). Anything that does not match a supported entity type is passed through unchanged.

Does Privian store my prompts?

No. Raw prompt bodies and rehydrated response bodies are not written to storage. Privian retains structural counters and sanitized observability events, plus your account, billing, hashed API keys and AES-GCM-encrypted BYOK credentials.

Can Privian staff read my prompts?

Operators cannot browse raw prompts because raw prompts are not retained. Support work uses request ids and sanitized event metadata.

What does OpenAI, Anthropic or Google see?

The masked prompt. They never receive the original values that Privian successfully detected and masked. Anything outside the supported entity set is passed through unchanged, and the provider's own terms apply from there.

What does BYOK change?

With BYOK, the call to the upstream provider is authenticated with your own provider key, decrypted only in process for the request. Billing, provider-side rate limits and provider-side terms are between your account and the provider.

When should I choose self-hosted inference instead?

When policy or regulation requires that data never leaves a controlled environment, when you cannot rely on any third-party model provider, or when you need a level of isolation that any hosted gateway by definition cannot provide. See Privian vs self-hosted LLMs.

Does Privian support regional data residency?

Not today. Privian does not currently offer customer-selectable data residency or region pinning. Requests are processed on managed serverless infrastructure and routed to the managed model provider you configure via BYOK; the provider's own region settings apply on their side. If strict residency is a hard requirement, self-hosted inference is the right tool — see Privian vs self-hosted LLMs.

• GDPR and LLMs

  How prompt-level controls fit EU privacy expectations.

• Security

  Privian's security model, retention and credential handling.

• Sub-processors

  Providers Privian uses and what each one sees.

• Privacy

  Privacy notice for the Privian application and gateway.

• Security

  Operational security posture and trust boundaries.

• Architecture

  How the gateway is built end-to-end.

• GDPR and LLMs

  Reducing prompt-level sensitive-data exposure under GDPR.

• PII Masking

  Supported entity types and how masking works.

• Prompt Security

  Prompt-level data protection in practice.

• LLM Gateway

  The hosted gateway in front of managed providers.

• Zero retention

  What no raw-prompt retention means in practice.

• Compare

  Privian vs other AI gateways and self-hosted models.

• Subprocessors

  Provider relationships and the BYOK boundary.