AI security
Private AI for compliance: privately hosted models, no training on your data
Corsa can run AI on privately hosted models that don't share infrastructure with public AI services, and customer data is contractually never used to train any model. Each AI feature is opt-in, PII guardrails sit on every pipeline, agent actions and guardrail activations are logged, HYOK-encrypted fields never reach AI, and analysts make every final decision.
In short: Corsa can run its AI features on privately hosted models that don't share infrastructure with public AI services, and customer data is never used to train, fine-tune or improve any model. That guarantee is contractual. Every AI feature is opt-in, prompts pass through PII guardrails, every meaningful agent action is logged, and an analyst always makes the final compliance decision.
Privately hosted models
Corsa's AI can run on privately hosted models. When AI runs on privately hosted models, customer data isn't sent to shared or multi-tenant AI endpoints or exposed to public AI APIs during inference. Models are reached through private channels with no external data egress, and model versions are pinned and updated through the same change management process as any other production deployment.
Opt-in, one feature at a time
AI in Corsa is off until a platform administrator turns it on, and each feature is toggled independently:
Feature | What it does |
|---|---|
Copilot | Natural-language data exploration and insights, with its reasoning steps and tool calls shown |
Alert Agents | Alert triage recommendations with the signals behind them |
Investigation Agent | Investigation narrative drafts that cite the data points used |
DD Agent | Due diligence enrichment and profiling |
Changes take effect immediately. A platform that leaves AI off keeps full functionality, because AI augments workflows and is never a dependency.
PII guardrails on every AI pipeline
Input guardrails. Prompts are scanned and sanitized before inference. Highly sensitive fields, such as government IDs, account numbers and financial credentials, are redacted or tokenized before they reach the model layer.
Data minimization. Context windows are scoped to the minimum data the task needs.
Output guardrails. Responses are checked for PII leakage, hallucinated identifiers and sensitive content, and structured outputs such as risk scores and dispositions are validated against expected schemas.
No training on your data, in writing
Customer data is never used to train, fine-tune or improve any AI model, whether Corsa hosts it or a third-party provider does. The policy covers transactions, client records, alerts, cases, investigation narratives, documents and Copilot conversations. Corsa doesn't aggregate, anonymize or pseudonymize customer data for training or benchmarking. Sub-processor agreements prohibit using customer data for model training, evaluation or improvement, the sub-processor list includes AI infrastructure providers, and changes to AI sub-processors are announced in advance. Customers can request additional AI-specific contract provisions.
AI audit logging and guardrail logging
Meaningful agent operations, such as alert triage, investigation narratives and due diligence enrichment, are written to the audit log with the timestamp, the output and the user who started the request. Guardrail activations, such as redactions, blocks and sanitizations, are logged separately for security review. AI logs follow the platform's retention policy, which you configure to your regulatory requirements, and they can be exported for compliance reporting or to your SIEM. When data is deleted, it's purged from AI logs, caches and intermediate processing stores too.
HYOK fields stay out of AI entirely
For the most sensitive identifiers, customers can add a cryptographic boundary. Fields encrypted with Hold Your Own Key (HYOK) are stored in Corsa only as ciphertext and decrypted only in authorized users' browsers, using the KMS key in the customer's AWS account. Corsa's AI features have no plaintext for those fields, so the values can't appear in prompts or inference pipelines for Copilot, Alert Agents, the Investigation Agent or the DD Agent. The trade-off is deliberate and chosen per field: AI answers and summaries can't use HYOK-encrypted values, while every other field stays available to AI. With BYOK, AI keeps working on protected fields under the same guardrails.
The analyst makes the final call
Corsa's AI follows four principles. A person stays in the loop, and final compliance decisions always rest with your team. Every AI action is explainable: the Triage Agent recommends Resolve, Escalate or Review with its rationale and the data it examined, and investigation narratives are editable before they're finalized. There are no opaque scores, because the inputs and logic behind a risk score, disposition or narrative are visible and auditable. Models and agents are tested against curated benchmarks, adversarial inputs and edge cases before they reach production.
Why this matters for banks
Model risk and third-party risk teams ask the same questions of every AI vendor: where the models run, whether data leaves the environment, whether it trains anything, and whether each AI output can be traced and reviewed. US model risk guidance, SR 26-2 and OCC Bulletin 2026-13, leaves generative and agentic AI outside its scope pending a request for information, so banks apply their own governance today. Private hosting, opt-in controls, logged agent actions and human sign-off give that review concrete evidence.
Learn more: AI security documentation · Corsa AI agents · HYOK for AML compliance · AI model governance · Security and data privacy · Corsa for banks · Most secure AML compliance platforms for banks
Frequently asked questions
Does Corsa train AI models on customer data?
No. Customer data is never used to train, fine-tune or improve any AI model, whether hosted by Corsa or a third-party provider, and this is contractually guaranteed. Sub-processor agreements also prohibit using customer data for training, evaluation or improvement.
Where do Corsa's AI models run?
Corsa can run AI on privately hosted models that don't share infrastructure with public AI services. With privately hosted models, customer data isn't sent to shared or multi-tenant AI endpoints, and models are reached through private channels with no external data egress.
Can we turn AI off in Corsa?
Yes. AI is opt-in, and administrators enable Copilot, Alert Agents, the Investigation Agent and the DD Agent independently. Changes take effect immediately, and a platform without AI keeps full functionality.
How are AI actions audited?
Meaningful agent operations are logged with the timestamp, output and initiating user, and guardrail activations such as redactions and blocks are logged separately. AI logs follow your configured retention and can be exported for compliance reporting.
Can sensitive identifiers be kept out of AI completely?
Yes. Fields encrypted with HYOK are stored in Corsa only as ciphertext and decrypted only in authorized users' browsers, so no AI feature receives their plaintext. You choose which eligible fields to encrypt, and other fields stay available to AI.
