Governance

Control is part of the architecture.

Permissions, approval, attribution and undo accompany agent actions. The rules protecting the software also apply to the AI working inside it.

Explore

Inherited, capped permissions.

An agent acts for an identified person. Its scope is bounded by that person’s rights and those of the assigned task. An exception must be approved, time-limited and recorded.

A record that explains the action.

Who asked? Which agent acted? Who approved? Which rules applied? What state changed? These details make the action attributable and verifiable.

The record follows the business object. The team can review an action instead of reconstructing events from scattered technical logs.

Preserve the previous state.

Reversible actions retain what they changed. Undo restores the previous state while keeping a record of the initial action and its correction.

External effects, such as a payment or a sent email, require their own approval and compensation rules. The scope of undo is defined with the business process.

Received data does not authorize an action.

An email, ticket or document can contain a hostile instruction. Even if the model follows it, application tools must still check permissions, allowed values and approval steps.

A chosen boundary for your data.

European hosting or your own servers, local or remote models: the project determines where data may be processed. Business documents and requests are not used to train a third party’s model.

04 — Governance

Intelligence moves forward.
Control stays with you.

Record historyDemo
14:32
Agent · Renewal

Draft added to the Dupont record.

ATTRIBUTED · REVERSIBLE
Understand the safeguards

The same permissions.

Never more access than the person the agent works for. Every exception requires human approval.

An action. A record.

Who asked, who approved, what changed. A readable record, attributed and retained.

The right to go back.

The previous state is retained. A mistake can be corrected without losing the history.

Your data stays yours.

European hosting or your own servers. Your data is never used to train anyone else’s model.

It starts with
one process.

The one your team does by hand, every week.
What if your software took care of it?