Falaah Falaah AI
Trust Center

AI governance

Which models Muin uses, what happens to your data, what the AI can do on its own, and what always waits for a person.

Trust Center Security pack
What happens to your data when Muin's AI works A request from your organization is scoped to your data, sent to a model on AWS Bedrock that does not train on it, logged as a hash, and any payment waits for a person. 1 YOU A request From a person in your organization. 2 ISOLATION Scoped to your data A query with no organization filteris refused. 3 MODEL Runs on AWS Bedrock Your data never trains or tunes amodel. 4 AUDIT Logged A hash of prompt and response, notthe text. 5 A PERSON Money waits for you Payments and refunds wait forapproval.
What happens to your data when Muin's AI works

Models and where they run

Muin's AI features run on foundation models served through AWS Bedrock, including Anthropic Claude models.

AI processing:
United States (AWS Bedrock, US regions)
AWS services used:
AI processing (Bedrock), document OCR (Textract), transactional email (SES), key management (KMS)

Your organization's data and AI processing stay in the United States.

Your data does not train models

Muin does not train or fine-tune any model on your organization's data.

AWS states that Amazon Bedrock does not use prompts and responses to train AWS models and does not share them with third parties. Read AWS's data protection statement.

A person approves every payment

Any AI action that pays or creates a liability, such as approving or submitting an invoice, scheduling a payment, a payout, or a refund, is held until a person approves it in Muin.

A standing automation rule cannot approve it. When the platform cannot tell whether an action involves money, it treats it as money and holds it.

Chat changes are off by default

The chat assistant only reads unless your organization turns on chat write actions. If that setting cannot be checked, changes stay off.

Every AI call is on the record

Each AI call is written to an append-only audit log with the organization, the user, who the AI was acting for, the model requested and the model used, token counts, and cost.

The log keeps a fingerprint (SHA-256 hash) of the prompt and the response, not their text. Default retention for the ai audit log: 90 days.

Agents are tested before and after launch

Agents built in Agent Studio can be run against evaluation datasets: test cases scored by exact or tolerance matching, or graded by a model against an expected answer, with each run compared to the last.

An agent's owner can also turn on review of its real runs. Every hour a sample of finished runs (5% by default) is graded, including whether each tool call was the right one and allowed for that agent.

Your organization's data stays yours

Database queries are checked for an organization filter, and on the web request path an unscoped query is refused. While an AI assistant is working for your organization, the platform refuses any attempt to read across organizations.

What Muin's AI never does

  • Train or fine-tune a model on your organization's data.
  • Pay, refund, or create a liability without a person approving it.
  • Let a standing automation rule approve a payment.
  • Read another organization's data while working for yours.
  • Make changes from chat unless your organization has turned chat write actions on.

Reviewing Muin's AI for your organization?

Send your AI or security questionnaire to contact@falaah.ai.

See also: Trust Center | Security & Privacy | Sub-Processors | DPA