Skip to content

Data Handling and Compliance

Access control decides who can reach data; this page covers what happens to the data itself: what reaches a language model, where conversation data lives, how long it is kept, and the compliance posture behind it.

These rules hold across all Alation AI services:

  • Customer data is never used to train foundation models.
  • Customer content is processed only to provide the service and is never shared with model providers beyond that.
  • Customer data is encrypted in transit and at rest.
  • AI features inherit the security and compliance posture of the Alation Cloud Service.
  • AI features are opt-in, and customers can disable them at any time.

Agent Studio reaches language models through three paths, all under no-training commitments:

PathModelsData commitments
Amazon BedrockAnthropic and other foundation modelsBedrock does not store or log prompts or responses, and customer data is not used for training
OpenAI Enterprise APIOpenAI modelsContractually barred from training on customer content; content deleted within 30 days of termination
Bring your own modelYour endpoints (Bedrock, Azure OpenAI, OpenAI, Anthropic, Google Vertex AI, or OpenAI-compatible)Governed by your own provider agreement

Bedrock’s automated abuse detection also screens for prompt injection, jailbreaking, and other adversarial inputs, with no human review of inputs or outputs.

During an Agent Studio chat, the model receives:

  • The chat title and the user’s messages
  • Agent responses, which can include earlier messages and tool outputs
  • Tool outputs, which can contain data retrieved from your connected sources (for example, rows returned by the SQL Execution tool the user is permitted to run)

Catalog metadata sent as context (names, descriptions, tags, lineage, linked SQL) is scoped to what the requesting user is allowed to see, as described in the request lifecycle.

WhatWhereRetention
Chat data (messages, responses, tool outputs)Postgres, encrypted with the customer’s encryption keyUntil deleted
Application logs (LLM requests/responses for debugging)Alation-hosted observability tooling30 days

Logs can contain user inputs and data that appeared in tool outputs. Admins can review activity through interaction logs and usage.

Agent Studio inherits the Alation Cloud Service compliance posture:

  • SOC 2 Type II
  • ISO 27001 / 27701
  • HIPAA

Regional availability of AI features follows Amazon Bedrock and OpenAI model availability in your region.