Meta tracking pixel
Aloan

Security AI Data Security

How Aloan uses AI on your loan data

Your loan files never train a model, each AI call carries only your institution's data, and every AI-drafted statement cites its source for your analysts to check.

Reviewed October 2026

On this page

Providers

Two enterprise AI services, both in the United States.

Aloan uses AI to classify and extract data from loan documents, spread financials, answer analyst questions and draft credit memo narrative for analyst review. The application, the document pipelines, the extraction logic and the AI orchestration are Aloan's own. The foundation models underneath are commercial models reached through enterprise APIs.

Those models run on two providers: Google, through the Gemini API and Gemini Enterprise Agent Platform, and AWS Bedrock, which runs private model copies within AWS. Both process data in the United States, and the platform fails over between them automatically. Aloan does not train, fine-tune or host a proprietary model, and customer data never goes into consumer or public AI tools.

This page covers the AI detail. Encryption, access control, the SOC 2 report and the vendor-review documents are on the security overview.

Training and retention

Your loan files never train a model.

Aloan does not train, fine-tune or improve any model on customer data, and the paid API and enterprise terms with Google and AWS prohibit either provider from using inputs or outputs for training. Those terms flow down to your data. No customer-specific model exists.

Retention is a separate question, and the answer differs by provider:

  • AWS Bedrock stores no inputs or outputs, and Aloan holds it at that setting.
  • Google may cache inputs in memory for up to 24 hours and keeps prompts and outputs only for abuse monitoring. On the Gemini Enterprise Agent Platform, a prompt is logged only if Google's automated safety classifiers flag it, and the log is kept up to 90 days in the US. On the Gemini API, prompts and outputs are kept 55 days for the same purpose. Logged data trains no model other than those Google uses for policy enforcement.

A zero data retention configuration is available under our Google Cloud Master Agreement: the Gemini Enterprise Agent Platform project runs with caching disabled and Google's abuse-monitoring logging removed, so neither provider writes a prompt or output to durable storage. Google issues no standalone certificate for this; the executed agreement and Google's zero data retention documentation are what we show your reviewer. Your configuration is named in your agreement and confirmed in writing before any of your data is processed.

Access and isolation

Each AI call carries one institution's data.

Aloan is multi-tenant with logical isolation per institution. Every record is scoped to its institution, and permission checks enforce that scoping on every request. AI processing works the same way: a prompt contains only the requesting institution's data, and nothing derived from one institution's documents can influence outputs for another. Our January 2026 penetration test tested that boundary directly.

Gemini and Bedrock are managed inference services, and the providers' engineering teams have no access to your prompts or outputs. The one exception is Google's abuse monitoring: content its classifiers flag may be reviewed by authorized Google employees, solely to check for a Prohibited Use Policy violation.

Every AI surface that accepts free text separates user content from system instructions, sanitizes input and runs behind prompt-injection guardrails covered by automated tests. The annual independent penetration test exercises those surfaces, and the January 2026 test included prompt-injection testing.

Human oversight

Your analysts make every credit decision.

Aloan makes no credit decision. It produces no credit score and no approval or denial recommendation, and protected-class attributes are not model inputs. Low-confidence extractions route to human review, every extracted value is editable with the original preserved, and a credit memo advances only through analyst review and a named approval.

Every AI-drafted statement cites its source, extracted values carry the page range they came from, calculated metrics show their formulas, and overrides record who, when and why. An approved memo freezes the analysis behind it.

In production, an automated pass checks each document's extracted figures against the document's own printed arithmetic. A fixed set of sample loan packages with known answers also runs regularly, and always before a model change or a major change to the code that runs our AI processing; every extracted value and downstream calculation is checked against those answers, so a regression surfaces before release.

Governance

Written notice before your AI terms change.

AI use is governed by Aloan's AI Governance Policy and AI and LLM Data Isolation Policy. The CTO owns the AI provider inventory, AI change management and AI incident response; material changes to AI use cases or providers need the CEO's approval as Chief Compliance Officer. Any provider that would touch customer data first needs a vendor risk assessment, confirmed no-training terms and CTO approval.

We notify you in writing before a provider that receives customer data is added or removed, before your retention configuration changes, and before any change to a provider's retention or training terms takes effect. A model version change within a provider under unchanged terms runs under our change management process and is not separately notified.

For model risk, SR 26-2 and OCC Bulletin 2026-13 replaced SR 11-7 on April 17, 2026. The new guidance is risk-based and leaves generative and agentic AI out of scope, so your own program sets the controls, and the citations and override records above are the evidence it draws on. For OCC-supervised banks, OCC Bulletin 2026-54 (November 2026) gives examiners a framework for generative and agentic AI. It sets no expectations for banks, so the classification stays with your program. For third-party risk, the OCC 2023-17 guide for AI underwriting vendors walks the lifecycle.

Documents

What your AI review can draw on.

Through the Trust Center or on request:

  • The AI Governance Policy and the AI and LLM Data Isolation Policy
  • A NIST AI RMF mapping
  • The subprocessor list and a data flow diagram
  • The penetration test report, under NDA
  • Our SOC 2 Type II report, under NDA

The full document list is on the security overview.

FAQ

AI data security questions

Is our loan data used to train AI models?

No. Aloan does not train, fine-tune or improve any AI model on customer data, and its AI providers, Google and AWS, are contractually prohibited from doing so. No customer-specific model exists, and nothing from your documents influences outputs for any other institution.

Which AI providers does Aloan use?

Google (the Gemini API and Gemini Enterprise Agent Platform) and AWS Bedrock, which runs private model copies within AWS. Both are commercially available enterprise services processing data in the United States. Aloan built the application, document pipelines and AI orchestration; it does not train, fine-tune or host a proprietary model.

Do the AI providers keep our data?

Neither provider trains on it. AWS Bedrock stores no inputs or outputs. On the zero data retention configuration, available under our Google Cloud Master Agreement, Google writes no prompt or output to durable storage either. On the standard configuration Google may cache inputs in memory for up to 24 hours and retains prompts and outputs only for abuse monitoring under its published terms. Your configuration is named in your agreement and confirmed in writing before any data is processed.

Can the model providers see our data?

Gemini and Bedrock are managed inference services, and the providers' engineering teams have no access to your prompts or outputs. The one exception is Google's abuse monitoring: content its automated classifiers flag may be reviewed by authorized Google employees, solely to check for a Prohibited Use Policy violation.

How is the AI protected against prompt injection?

Every AI surface that accepts free text separates user content from system instructions, sanitizes input and runs behind prompt-injection guardrails covered by automated tests. The annual independent penetration test exercises these surfaces directly; the January 2026 test included prompt-injection testing.

Does Aloan's AI make credit decisions?

No. The platform produces no credit score and no approval or denial recommendation, and protected-class attributes are not model inputs. Every credit decision is made by your own analysts and approvers.

Will you tell us if an AI provider or its terms change?

Yes. We notify you in writing before a provider that receives customer data is added or removed, before your retention configuration changes, and before any change to a provider's retention or training terms takes effect. A model version change within a provider under unchanged terms runs under our change management process and is not separately notified.