> ## Documentation Index > Fetch the complete documentation index at: https://docs.corsa.finance/llms.txt > Use this file to discover all available pages before exploring further. # AI Security - Private Models, PII Guardrails & Data Protection > How Corsa secures AI-powered compliance features - privately hosted models, opt-in controls, PII guardrails, HYOK protection, and no-training guarantees. Corsa embeds AI across the platform - from alert triage and investigation narratives to copilot-powered data exploration. Every AI feature is designed with the same security-first approach that governs the rest of the platform, and evaluated against our responsible AI principles. ## Responsible AI Principles Corsa's AI capabilities are governed by a set of principles that ensure AI is used safely, transparently, and under human control. * **Human in the loop** - AI augments human decision-making but never replaces it. Final compliance decisions always rest with your team. AI recommendations include confidence indicators and the evidence they were based on. * **Explainability** - Every AI action is transparent. Copilot shows its reasoning steps and tool calls. Alert agents explain the signals that drove their recommendation. Investigation narratives cite the specific data points used. * **No black boxes** - Corsa does not use opaque scoring models. When AI produces a risk score, disposition suggestion, or narrative, the inputs and logic are visible and auditable. * **Continuous evaluation** - Models and agents are regularly tested against curated benchmarks, adversarial inputs, and edge cases before deployment to production. ## Privately Hosted AI Models Corsa's default AI configuration uses **privately hosted models** that do not share infrastructure with public AI services. Customer data is never sent to shared or multi-tenant AI endpoints. * No customer data is exposed to public AI APIs during inference * AI models are accessed through secure, private channels with no external data egress * Model versions are pinned and updated through the same change management process as any other production deployment ## Opt-In AI Features AI capabilities in Corsa are **opt-in by default**. Platform administrators have full control over which AI features are enabled for their organization. | Feature | What It Does | Independently Toggleable | | ----------------------- | ---------------------------------------------- | ------------------------ | | **Copilot** | Natural-language data exploration and insights | Yes | | **Alert Agents** | AI-powered alert triage recommendations | Yes | | **Investigation Agent** | Automated investigation narrative generation | Yes | | **DD Agent** | Due diligence data enrichment and profiling | Yes | * AI features can be toggled on or off at the platform level at any time * Changes to AI feature settings take effect immediately * Platforms that choose not to enable AI features continue to operate with full functionality - AI augments workflows but is never a dependency This gives compliance officers and CISOs confidence that AI is only active when explicitly authorized. ## PII & Sensitive Data Guardrails All AI pipelines include guardrails that prevent personally identifiable information (PII) and other sensitive data from being mishandled during inference. ### Input Guardrails * Prompts sent to AI models are scanned and sanitized before inference * Highly sensitive PII fields (government IDs, account numbers, financial credentials) are redacted or tokenized before reaching the model layer * Context windows are scoped to the minimum data required for the task - no unnecessary data is included in prompts ### Output Guardrails * Model responses are validated before being surfaced to users * Outputs are checked for inadvertent PII leakage, hallucinated identifiers, and sensitive content * Structured outputs (risk scores, alert dispositions) are validated against expected schemas to prevent malformed or dangerous responses ### Audit Trail * Meaningful agent operations (alert triage, investigation narratives, due diligence enrichment) are logged with context including timestamp, output, and the user who initiated the request * Guardrail activations (redactions, blocks, sanitizations) are logged separately for security review * AI audit logs are retained according to the platform's data retention policy and can be exported for compliance reporting ## HYOK as an AI Data Protection Layer Corsa's [Hold Your Own Key (HYOK)](/hyok/index) encryption provides an inherent layer of protection for sensitive data in AI workflows. * Fields encrypted with HYOK remain encrypted within Corsa's environment - they can only be decrypted on the customer's side using their own KMS * Because Corsa cannot access the plaintext of HYOK-encrypted fields, these fields are never included in AI model prompts or inference pipelines * This gives customers cryptographic assurance that their most sensitive data (government IDs, financial credentials, account numbers) cannot be exposed to any AI model - regardless of provider or configuration * HYOK protection applies across all AI features: Copilot, Alert Agents, Investigation Agent, and DD Agent For organizations that require the strongest guarantees around sensitive data and AI, HYOK provides a zero-trust boundary that no software-level guardrail alone can match. ## Data Retention & Training Policies Corsa enforces strict policies to ensure customer data is never used to improve AI models. ### No Training on Customer Data Customer data is **never** used to train, fine-tune, or improve any AI model - whether hosted by Corsa or by a third-party provider. This is contractually guaranteed. * This policy applies to all data types: transactions, client records, alerts, cases, investigation narratives, documents, and copilot conversations * Corsa does not aggregate, anonymize, or pseudonymize customer data for model training or benchmarking purposes * When third-party AI providers are involved, Corsa ensures that provider agreements prohibit training on customer data passed through the API ### Data Retention * AI interaction logs (prompts, responses, metadata) follow the same retention policies as other platform data * Customers can configure retention periods based on their regulatory requirements * When data is deleted, it is purged from all systems including AI-related logs, caches, and intermediate processing stores * No AI-generated data persists beyond the configured retention window ### Sub-Processor Commitments * When Corsa's hosted models process data, sub-processor agreements explicitly prohibit the use of customer data for model training, evaluation, or improvement * Corsa maintains a sub-processor list (available on request) that includes any AI infrastructure providers * Changes to AI sub-processors are communicated to customers in advance ### Contractual Guarantees * Data handling commitments are codified in Corsa's contractual agreements * Customers can request amendments for additional AI-specific provisions ## Summary | Capability | How It Works | | --------------------------- | ------------------------------------------------------------------------------------ | | **Responsible AI** | Human-in-the-loop, explainable outputs, no black boxes, continuous evaluation. | | **Privately hosted models** | AI inference uses private models with no shared or multi-tenant endpoints. | | **Opt-in AI** | Every AI feature is off by default. Admins enable them independently per platform. | | **PII guardrails** | Input sanitization, output validation, and scoped context windows. | | **HYOK data protection** | HYOK-encrypted fields are cryptographically excluded from AI pipelines. | | **No training on data** | Customer data is never used to train or improve AI models. Contractually guaranteed. | | **Strict retention** | AI logs follow platform retention policies. Deleted data is purged from all systems. |