Policy

How Synthia Agents uses AI

Which models are called, what they receive, what is never done with your data, and how people stay in charge of what a coworker does.

Which models are called

A coworker answers with the model your administrator chose for it, from the model providers your workspace is set up with. Each coworker can use a different provider and model, and an administrator can limit which models a group may use. See Model providers.

To answer, the model provider receives the messages and the page or file content the coworker needs for that answer. The workspace sends nothing to a model provider for any other purpose.

Your data does not train models

Synthia Agents does not use your prompts, conversations, files or results to train AI models. What a model provider may do with what it receives is set by that provider's terms with your organisation.

What a coworker learns stays in your workspace as memory you can read: facts it noticed wait for review, and anything you forget stays forgotten. A Private chat is never saved to memory or used for learning. See Memory and standing instructions.

People stay in charge

  • Your team's boundary rules are checked before every action, and an action a rule forbids is refused. See Approvals and boundary rules.
  • An action that matters waits for a person's approval, and a request nobody answers expires rather than going ahead.
  • A password or verification code field, and a pay or checkout button, are always handed to a person.
  • You can watch a coworker's computer live and take control of it yourself.
  • Spending on models can be capped per coworker each month. See Budgets and cost control.

What a coworker reads is not an order

Text a coworker finds on a web page or in a document is treated as information, not as instructions. Attempts to slip instructions in are detected and flagged, and risky actions still wait for a person.

Everything on the record

Every action a coworker took, and every one the rules refused, is in an audit trail your administrators can search and export, with secrets masked. See Audit trail and compliance exports.

Check what matters

AI models can be wrong. Read a coworker's work before you rely on it, especially numbers, names and anything you send outside your organisation. Report a harmful or wrong answer to your administrator.

More on what is kept and who receives it is in the privacy notice and Where your data is kept; what the workspace may be used for is in the acceptable use policy.