Enterprise AI orchestration: authority before automation

Specify the action contract

Begin with the institutional task and the person accountable for its outcome. Separate reading, drafting, recommending and executing into different capabilities. A model that can propose a response should not automatically be able to send it, change a record or approve a payment. Record the data scope, action owner, expiry and approval requirement for each capability.

Test failures before expanding autonomy

Test missing identity, revoked access, conflicting sources, an unavailable dependency and a repeated request. Make retries safe with operation identifiers and distinguish an attempted action from a confirmed action. Treat retrieved text as information, not permission. Require authorization in the service performing the operation, not only in the interface or model prompt.

Keep the institution able to operate

Define a fallback when the assistant is unavailable. Preserve source references, the approved action and the resulting record without collecting unnecessary sensitive content. Evaluate on representative tasks chosen before testing; report failures as well as successes. These are proposed design controls, not evidence that a particular Roadyom deployment has passed an independent security evaluation.

Continue the work