Responsible AI Policy

One task at a time

Every agent we deploy has a defined input, a defined output and a bounded scope. We do not deploy general assistants into regulated workflows.

A person approves

Actions with external effect wait for a named approver. Gates are removed only when logs and evaluation results justify it.

Everything is logged

Inputs, tool calls, outputs and approvals are recorded in a form the client’s audit function can read, and decisions can be reconstructed from the log.

Measured, not demonstrated

We publish evaluation pass rates by failure class and agreement rates across repeated runs, and rerun them on every model or prompt change.

Client data stays with the client

Work happens in the client’s environment. We do not train models on client data and do not hold it on our systems.

Incidents

We disclose incidents affecting a deployment to the client promptly and support their regulatory reporting obligations.