From Anadoris. This guide comes from the team behind Digital & Technology: websites, ERP, CRM and integrations, built by people who run businesses. About the group.
Describe actions precisely
Avoid broad permissions such as “manage advertising” or “handle orders.” Break them into specific actions: read performance data, draft a proposed budget change, update an approved field or send a customer-facing message. Each action has a different consequence and may need a different level of oversight.
Identify the person authorised to approve the action and the limits of that authority. A warehouse user may correct a packing note without being authorised to change a refund. Keep the permission model aligned with actual responsibility rather than granting broad access because it makes an integration easier to build.
Make the review meaningful
Present the proposed change, affected records, supporting information and expected effect together. The reviewer should be able to understand what will happen without searching through several screens. Flag uncertainty and stale inputs before requesting approval. A confirmation button is not meaningful if the person cannot see the scope of the action.
Record the approved version. If the proposal changes materially after approval, return it for review according to the agreed rule. Include a clear expiry or revalidation condition when the underlying information is time-sensitive. An old approval should not quietly authorise a substantially different action.
Keep execution and recovery observable
Log the approved action, the actor or service that performed it, the time and the outcome. Give users a way to distinguish requested, running, completed and failed work. Where reversal is possible, explain its limits; a sent message cannot be treated like an unsaved draft.
Begin with one workflow that crosses a meaningful boundary, such as changing a commercial setting or contacting a customer. Map its permissions and test both allowed and disallowed paths. This is a proposed design standard for accountable automation, not a claim that a particular product already supports every control. The objective is to reduce repetitive effort while preserving clear human responsibility for consequential decisions.
