Separate the draft from the action
An AI-generated summary and an instruction to change an ERP record are different levels of responsibility. Keep the draft step separate from the action step. The person reviewing a suggestion should be able to see exactly what would change before authorizing it.
Define the boundary in ordinary language: the assistant may classify or draft, but it may not release a payment, change access, or commit a customer-facing response without the specified approval. Then enforce that boundary through permissions and workflow logic, not just a prompt.
Make review meaningful
A review screen should help a person make a decision, not merely present an approval button.
- Show the proposed action and the affected record.
- Include relevant source material and known limitations.
- Allow editing, rejection, and escalation.
- Identify who is allowed to approve each action type.
- Require fresh review when the proposed action changes.
- Record the approved action and its outcome without unnecessary sensitive content.
Evaluate the uncomfortable cases
Test ambiguous requests, contradictory source material, missing information, and instructions embedded in untrusted documents. An assistant should not treat a document's contents as permission to perform an unrelated action.
Use representative examples and keep a separate evaluation set. Track unsupported claims, incorrect classifications, rejected suggestions, and review time. A fluent answer can still be wrong; a low correction rate on easy examples does not prove readiness for every workflow.
Keep a fallback and a change process
Define when the workflow should stop and hand the task to a person. Missing evidence, unavailable systems, and actions outside scope should have explicit handling. Do not interpret silence as approval.
Repeat the evaluation after changing prompts, models, tools, or source data. Keep the pilot narrow until its quality and controls are demonstrated. These practices are a design checklist, not a guarantee that an AI system will be error-free or meet a particular compliance standard.