User data egress
Personal identifiers leave your perimeter on every AI call.
Solution
Add AI features without exposing sensitive user data
One masking layer for every AI feature you ship.
Route user prompts through Privian so personal and sensitive values are replaced with placeholders before any LLM provider call. New AI features inherit the same protections automatically.
Problem
Every AI feature is a new data path
Shipping AI features in a SaaS product means sending user data to a third-party model. End users paste customer names, emails, account identifiers and sometimes credentials into your search bars, chat boxes and smart forms. Each new feature is another path for that data to reach an LLM provider that doesn't need to see it.
Why this matters
What's at risk
Repeated review work
Each new AI feature triggers a fresh data-handling review.
Provider coupling
Without a privacy layer, switching models means re-auditing data flow.
Accidental secrets
API keys and tokens pasted by users end up in provider context.
How Privian fits
One masking layer for every feature
Provider-routed gateway
Pick the model per request: OpenAI, Anthropic, Google, DeepSeek.
BYOK
Your provider keys, encrypted at rest. Usage shows up on your account.
Simple JSON contract
model + prompt + stream. No SDK rewrites once you're wired up.
Stable error model
Predictable error categories instead of raw provider errors.
Architecture
Where Privian sits
User input → Your backend → Privian gateway → LLM provider
│ mask
│
(placeholders)
│
▼
rehydrate ← provider response
│
▼
Your app UIFrontends do not need to change. Your backend already composes the prompt — Privian just changes where it goes.
Example flow
Example request
curl https://api.privian.io/v1/gateway \
-H "Authorization: Bearer $PRIVIAN_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "openai/gpt-5.5",
"prompt": "Summarize this doc for user jane@acme.com (account #4821).",
"stream": false
}'Use cases
AI features that benefit
Search assistants
Natural-language search over user content without sending raw PII.
Smart forms
AI-completed fields where the model never sees the underlying record.
Content generation
Generate copy referencing user context with identifiers masked.
In-product chat
Embed an assistant in your product without changing its data posture.
Summarization
Summarize user-generated content without forwarding it verbatim.
Embedded copilots
Ship vertical copilots in days, not weeks of review.
Implementation
Developer integration path
Step 1
Sign up
Create a Privian account and add a provider key (OpenAI / Anthropic / Google / DeepSeek).
Step 2
Generate a gateway key
Use a test key in staging and a live key in production.
Step 3
Send your first request
POST /v1/gateway with model, prompt and optional stream.
Step 4
Migrate features
Repoint each AI feature at the gateway. No SDK changes required.
Step 5
Observe
Track request and entity counts in the dashboard. Content is never logged.
Step 6
Ship the next feature
New AI surfaces inherit the same privacy layer by default.
Transparency
Beta limitations
- No OpenAI SDK drop-in — small proprietary JSON contract
- No native provider streaming (artificial chunking in beta)
- No tool / function calling support yet
- No custom user-defined entities yet
- No HIPAA / SOC 2 / PCI compliance claims today
- Prompt size capped at 32 KiB
FAQ
Frequently asked questions
- How do I secure AI features in SaaS?
- Route every model call through Privian's gateway. User prompts are scanned for the supported entity set; matched values are replaced with deterministic placeholders before egress and restored in the response. The provider only ever sees the masked text.
- Does Privian work with GPT and Claude?
- Yes. Privian is provider-agnostic. You add your own OpenAI, Anthropic, Google or DeepSeek key and select the model per request using a provider/id format such as openai/gpt-5.5 or anthropic/claude-sonnet-4.5.
- Does Privian support custom entity masking?
- Not yet. The current beta ships a fixed catalog of personal identifiers, financial and national IDs, and developer secrets. Custom user-defined entities are tracked as future work.
- What does the request look like?
- A small JSON body: model (provider/id), prompt (string up to 32 KiB) and an optional stream flag. There is no messages[] array and no OpenAI SDK drop-in today.
- Will adding Privian slow my AI feature down?
- Masking is an in-memory step. The dominant cost in any AI feature is still the provider's own latency.
- Do my users' prompts get stored?
- No. Raw prompts and responses are never persisted. Only structural counters used for billing and observability are kept.
Learn more