AI Features and Data Privacy

Catalio is designed with privacy at its core. This guide explains how your data is protected when using AI-powered features, including requirement analysis, semantic search, and the AI chat assistant.

Core Privacy Principles

Catalio’s AI features are built on four fundamental privacy principles:

1. Your Data Stays Yours

2. Organization Isolation

All AI processing is isolated per organization. Data from one organization is never accessible to another, and AI features cannot cross organizational boundaries.

3. Explicit Control

You control which data is accessible to AI. Every requirement has an “AI Accessible” toggle, allowing you to exclude sensitive content from AI processing.

4. Transparency

We’re transparent about how data flows, what providers receive, and how your content is processed. This document provides complete visibility into AI data handling.

Supported AI Providers

By default your organization uses Catalio AI, our managed provider. Organizations with Bring Your Own LLM (BYOLLM) enabled — granted per organization on request — can instead connect one of these providers directly:

Provider Data Policies

Each provider has their own data handling policies (summaries as of September 2026 — verify with the provider before relying on them for compliance decisions):

OpenAI: API requests through the enterprise API are not used for model training. Data may be retained for 30 days for abuse monitoring.

Anthropic: As of September 28, 2025, Anthropic’s data policy varies by plan type. Commercial offerings (Claude for Work, Claude Gov, Claude for Education, and API usage) retain full protections—data is never used for model training. Consumer plans (Free, Pro, Max) operate under an opt-out model where data may be used for training unless users disable this in settings. Since Catalio integrates via the API, your data is protected under commercial terms and is not used for training.

Azure OpenAI: Data remains in your Azure tenant with full enterprise controls. You control data residency, retention, and encryption.

AWS Bedrock: Requests run through your own AWS account and are not used for model training. Data residency is not guaranteed by “runs in your AWS account” alone: the inference profile you select determines routing. An in-region profile keeps a request within one region, while a geographic cross-region or global inference profile can route prompts and outputs to other regions, so content may leave the configured source region. Choose the inference profile that matches your residency requirements.

Ollama: Data never leaves your infrastructure. This is the highest privacy option for organizations with strict data handling requirements.

API Key Security

When you configure AI providers in Catalio, your API keys are protected with enterprise-grade security:

AES-256-GCM Encryption

All API keys are encrypted at rest using AES-256-GCM, the same encryption standard used by financial institutions and government agencies:

• 256-bit encryption keys
• Galois/Counter Mode for authenticated encryption
• Unique initialization vectors for each encryption
• Separate key management from database storage

Encryption Implementation

API Key Entry
     |
     v
AES-256-GCM Encryption
     |
     v
Encrypted Storage in Database
     |
     v
(When needed for API call)
     |
     v
Decryption in Memory
     |
     v
API Request to Provider
     |
     v
Immediate Memory Cleanup

The encryption key is stored separately from the encrypted data, typically in environment variables or a secrets management system, ensuring database access alone cannot reveal API keys.

Keeping Credentials Out of Logs

Catalio applies two separate redaction mechanisms. Neither is a single blanket filter over everything the system writes, so it’s worth knowing what each one covers.

Request parameter filtering. Incoming request parameters whose names match a fixed list are replaced with [FILTERED] before the request is logged. The list covers password, password_confirmation, token, api_key, secret, client_secret, secret_key_base, access_token, refresh_token, id_token, auth_code, and the PII fields ssn, social_security, credit_card, cvv, and pin. Matching is on the parameter name, so a credential sent under a name outside this list is not caught by it.

Error string sanitization. Errors routed through Catalio’s log-sanitization helper have Bearer tokens, Authorization header values, URLs, and long key-shaped strings redacted before the message is written. This applies to error text explicitly passed through that helper — it is not an interceptor on every log line in the system.

Error reporting. Only an explicit allowlist of diagnostic metadata keys is forwarded to our error tracker; keys outside that list are dropped rather than sent.

Stored integration credentials are separately protected by the encryption described above, so a credential at rest is not readable even where a log line references the record that holds it.

Data Flow for AI Features

Understanding how your data flows helps you make informed decisions about AI feature usage.

Requirement Analysis

When you analyze a requirement for quality, sentiment, or categories:

Requirement Content
     |
     v
Catalio Application
     |
     | (Selected fields only)
     v
Your AI Provider
     |
     v
Analysis Results
     |
     v
Stored with Requirement

What’s sent: Title (quality and category suggestions only), and your user story text — user want for all three, plus user benefit for sentiment analysis specifically. Acceptance criteria and linked use case content are not currently sent, even though quality assessment and category suggestions accept an acceptance-criteria field in principle.

What’s NOT sent: Acceptance criteria, linked use cases, user identifiers, organization metadata, audit trail information, internal IDs

Semantic Search (Embeddings)

For semantic search, Catalio generates vector embeddings:

Requirement Text
     |
     v
Embedding Model API
     |
     v
Vector (1536 numbers)
     |
     v
Stored in Catalio Database
     |
     v
Used for Similarity Search

What’s sent: Combined text from title, user want, and user benefit fields

What’s stored: Only the numeric vector representation, not the original text sent to the API

Important: Embeddings are mathematical representations that cannot be reversed back into original text. They enable semantic search without storing your content at the provider.

AI Chat Assistant

When using the AI chat assistant:

User Message
     |
     v
Catalio Chat System
     |
     | + Context from tools
     v
Your AI Provider
     |
     v
AI Response
     |
     v
Displayed to User
     |
     v
Stored in Conversation

What’s sent: Your message, conversation history, and context from tool calls (requirement summaries, search results)

What’s stored in Catalio: Complete conversation history for your reference and continuity

Data Control and Deletion

You maintain full control over your AI-related data.

AI Accessible Toggle

Every requirement has an “AI Accessible” toggle:

• Enabled (default): Requirement is included in AI analysis and semantic search
• Disabled: Requirement is excluded from all AI processing

Use this for:

• Sensitive or confidential requirements
• PII-containing content
• Internal notes not suitable for AI processing
• Compliance-restricted information

Conversation Deletion

Users can delete AI chat conversations:

1. Open the conversation
2. Click the options menu
3. Select Delete Conversation
4. Confirm deletion

Deleted conversations are permanently removed and cannot be recovered.

Organization Data Deletion

When an organization is deleted from Catalio:

• All requirements and associated AI data are deleted
• All embeddings are removed
• All chat conversations are deleted
• All provider configurations and encrypted API keys are deleted

Right to Deletion (GDPR)

For GDPR compliance, Catalio supports data subject requests:

• Individual user data can be anonymized or deleted
• Organization data can be exported or deleted
• Audit trails maintain minimal necessary information

Contact support@catalio.ai for data deletion requests.

Compliance Considerations

GDPR (European Union)

For EU organizations or those handling EU citizen data:

Recommended: Use Azure OpenAI with EU data residency

• Deploy Azure OpenAI resource in West Europe or Sweden Central
• Data never leaves EU boundaries
• Full GDPR compliance controls

Considerations:

• Standard OpenAI API processes data in the US
• Document your legal basis for AI processing
• Include AI processing in your privacy policy
• Enable data subject access and deletion

HIPAA (Healthcare)

For organizations handling protected health information (PHI):

Recommended: Use Azure OpenAI with BAA

• Microsoft offers Business Associate Agreements for Azure OpenAI
• Configure with HIPAA-compliant settings
• Enable audit logging

Considerations:

• Exclude PHI from AI-accessible requirements
• Document AI processing in your HIPAA policies

SOC 2

For organizations requiring SOC 2 compliance:

Recommended: Use enterprise provider tiers

• Azure OpenAI includes SOC 2 compliance
• OpenAI Enterprise provides compliance documentation
• Anthropic offers enterprise agreements

Considerations:

• Document AI provider security controls
• Include in your third-party risk assessment
• Monitor for security advisories

Financial Services

For banks, insurance, and financial services:

Recommended: Self-hosted or Azure OpenAI

• Ollama for complete on-premise control
• Azure OpenAI with financial services certifications
• Document AI processing in regulatory filings

Best Practices

Minimize Sensitive Data in Requirements

Write requirements to minimize PII and sensitive information:

Instead of:

“John Smith (john.smith@company.com) needs to export his social security number for tax filing”

Write:

“Users need to export personal tax identifiers for compliance reporting”

Use AI Accessible Toggle Appropriately

Mark requirements as non-AI-accessible when they contain:

• Personal identifiable information (PII)
• Financial account numbers
• Healthcare information
• Trade secrets or proprietary formulas
• Classified or restricted information

Review Provider Policies

Before configuring a provider:

1. Review their data handling policies
2. Understand their data retention periods
3. Verify compliance certifications
4. Consider data residency requirements

Audit AI Usage

Regularly review AI feature usage:

1. Check which providers are configured
2. Review feature assignments
3. Audit who has configuration access
4. Monitor for unusual usage patterns

Document Your Policies

Create internal documentation covering:

• Which AI providers are approved
• What data can be processed by AI
• Who can configure AI features
• How to handle AI-related incidents

Summary

Resources

• Supported AI Providers - Provider comparison and capabilities
• Bring Your Own LLM - Configure your AI provider
• Catalio Privacy Policy - Legal privacy documentation

Questions?

For privacy-related questions about AI features:

• Email: privacy@catalio.ai
• Security: security@catalio.ai
• General Support: support@catalio.ai

Last Updated: December 2025