Skip to content

Privacy & data handling

This page is a product-specific supplement to InPolicy’s main Privacy Policy. Everything here is within the commitments made there; where there’s any apparent conflict, the Privacy Policy governs.

Short answer: InPolicy does not store Content. The messages, tool results and agent replies you send, and the policy analysis we return, are processed transiently and are not written to persistent storage. The Service does not ask for identifiers of your end users. For each conversation, InPolicy keeps a conversation state (a Policy State Summary), which contains no Content, along with the limited records described below.

The main Privacy Policy defines a handful of terms that apply directly to this product. Using them here for consistency:

  • Content: Input (user messages, tool returns, agent drafts processed through the Service) and Output (the policy analysis, flagged passages and suggested rewrites returned). InPolicy does not store Content. It is processed transiently and is not written to persistent storage.
  • Customer Data: account and organizational data, including the policies you author.
  • Usage Data: information about how the API is used, such as call counts and violation metadata. No Content, no end-user identifiers.
  • Policy State Summary: a system-generated record of which policies are currently applicable to a conversation. Contains the applicable policies (which are Customer Data), a turn counter and a log of when each policy started or stopped applying. Contains no message text and no personal information. (Defined in Privacy Policy §1.)

Kept for the life of the conversation:

Data Purpose
Caller-chosen sessionId Identifies the conversation. Opaque to us.
Turn count To number binding events by turn.
The policies that currently apply To report what changed on the next turn.
A log of when each policy began or stopped applying Returned when you request the conversation state.

Session-level retention. Conversation state is kept while a conversation is active. It expires once the conversation has had no new turn for a set period: 4 hours by default, or a period of 5 minutes to 7 days that you choose. You can end a conversation at any time through the API, which deletes its conversation state. Related temporary records used to process the conversation expire automatically.

Kept for the customer relationship:

Data Purpose
API key metadata (hashed key, settings) Authentication, rate limits and feature settings. The key itself is never stored.
Usage Events Usage-based billing and rate-limit accounting. See below.

Usage Events record which API key made a call (or that a signed-in user did), when, its result and its billable volume. They contain no Content and no end-user identifiers. They are Usage Data under the Privacy Policy (§2.2).

  • Content. The text your users, tools and agents send, and the policy analysis, flagged passages and suggested rewrites we return, are processed transiently and are not written to persistent storage (Privacy Policy §2.2).
  • recentContext and priorHistory you pass on record_turn. These are Content and are handled the same way.
  • endUserContext and country. Processed only for the request in which they are sent, and not written to persistent storage.
  • Violation text from check_output / check_tool_call. Flagged passages and suggested rewrites are Content: they are returned to you in the API response and are not written to persistent storage. A record that a check found a violation, without any text, may be kept as Usage Data.
  • End-user identifiers, unless you send them. The Service does not ask for them. A conversation is identified by the sessionId you choose, within your organization’s account.

Apart from Usage Data, our systems also generate technical values from Content, such as one-way fingerprints. These values do not contain the text of the Content, are used solely to provide, operate and secure the Service, and are kept only as long as needed for that purpose.

If you want durable audit trails of binding events or violations:

  • Capture diff.bindingEvents on every recordTurn response.
  • Capture violations on every postInference / checkToolCall response.
  • Write these to your own logging system (Datadog, Splunk, S3, a dedicated Postgres table, whatever fits).

InPolicy is not a system of record for binding events or violation details. If you need an audit trail, keep your own copy.

Two fields we treat as opaque. Don’t put personally identifiable information in them:

  • sessionId: kept with the conversation state and related temporary records, and may appear in our operational logs.
  • endUserContext.attributes: processed for the request only and not written to persistent storage, but it does pass through our servers. Keep it non-identifying (tier: enterprise, not email: bob@example.com).

All Personal Data is hosted on GCP in the United States. InPolicy does not transfer Personal Data outside the US. Per Privacy Policy §7. For GDPR customers transferring from the EEA, transfers are governed by the SCCs in our Data Processing Agreement.

Full list is in Privacy Policy §4.1. For the governance API specifically:

  • Google Cloud Platform (hosting and data storage): holds Customer Data and Usage Data. United States.
  • Google Cloud Vertex AI (LLM inference for policy retrieval and violation detection): handles transient Content only. Vertex does not store prompts or responses for InPolicy’s account.
  • TypeSafe AI (decision-model scoring of candidate policy matches during violation detection): receives transient Content (the text being checked, its message header, and its conversation context: earlier messages in the email thread or chat conversation, with their authors’ names) and the Customer Data needed to judge it (the applicable policy rules with their examples and exceptions, and relevant Tenant Knowledge Base facts). Zero data retention is enabled for InPolicy’s account, so TypeSafe keeps no copies beyond what is needed to return the response and meet its legal obligations. United States. Added October 2026; not in use before the notice period has run.

We give 30 days’ notice before adding or replacing any subprocessor.

Per Privacy Policy §3.3, InPolicy does not train any AI models using Content or Customer Data. This extends to all subprocessors. Your policies, your users’ messages, and your agents’ outputs are never used to train anyone’s model.

InPolicy is preparing for a SOC 2 Type II examination (Privacy Policy §6). Current security posture documentation available upon request at security@inpolicy.ai. Infrastructure summary:

  • TLS in transit, everywhere.
  • Access to production limited to authorized personnel with MFA.
  • All internal access is logged and audited.
  • Hosted on GCP, United States only.

Full rights and request procedures are in Privacy Policy §8. For the governance API specifically, access/deletion/portability requests should go to privacy@inpolicy.ai. We respond within 30 days (EU/UK) or 45 days (California), with one possible extension.

InPolicy maintains a per-tenant knowledge base (product documentation, public filings, regulatory registers) used to disambiguate policy enforcement decisions. It contains no Content and no employee communications. Your organization can review and correct it at any time by contacting security@inpolicy.ai. See Privacy Policy §2.1.

For anything not covered here, the main Privacy Policy is authoritative. Direct questions to privacy@inpolicy.ai.