Published October 12, 2026
Published October 12, 2026
Define quality before selecting a sample
Sampling cannot rescue an unclear standard. For each work type, describe the few observable conditions that make an output acceptable: correct recipient, complete required fields, supported claims, authorized action, readable format, and traceable source. Separate critical failures from minor presentation issues.
The definition should be available to the assistant before execution. A review rubric revealed only after delivery becomes a memory test rather than a control.
Match the sample to the risk
Review more work when a process is new, changed, high-consequence, or showing exceptions. Stable, reversible work can use a smaller sample. Always inspect rare critical events even if random selection would miss them.
Avoid choosing only the easiest files or the items already suspected to be wrong. Use a transparent method such as every tenth completed item plus all policy exceptions. Record the population and selection rule.
Review evidence, not style preferences
Score against the stated acceptance conditions. If a reviewer prefers a different phrase but the output is accurate, authorized, and clear, do not label it a defect. Mixing preference with failure teaches the assistant to optimize for a reviewer’s mood.
When judgment is genuinely part of quality, give examples of acceptable variation and define who resolves disagreement.
Classify defects by consequence
A typo, missing attachment, unauthorized commitment, and privacy exposure should not share one error count. Use severity categories tied to required response: correct in the next cycle, correct before release, pause the process, or escalate immediately.
Record whether the defect reached its audience or was caught internally. This helps distinguish prevention from recovery without minimizing the underlying weakness.
Find the process cause
Ask whether the error came from missing input, confusing instructions, tool behavior, workload, inadequate training, or a decision outside the assistant’s authority. “Be more careful” is rarely a durable corrective action.
Update the smallest control that addresses the cause: a required field, example, validation rule, access change, or clearer escalation trigger. Do not rebuild the whole workflow around one unusual event.
Close the feedback loop
Share sampled findings quickly and with source evidence. The assistant should know what passed as well as what failed. Invite clarification when the rubric does not fit a real case.
Track corrective actions to completion and sample the affected condition again. A review that identifies a problem but never verifies the fix creates paperwork, not assurance.
Adjust sampling over time
Increase frequency after a material change or critical defect. Reduce it only after several comparable cycles meet the standard and the underlying process is stable. Document why the level changed.
Sampling is a way to learn about a process with less review effort. It should preserve accountability, reveal risk early, and give the assistant a fair, repeatable quality standard instead of constant surveillance.
This guide was published on October 12, 2026. For adjacent workflows, explore the Delegation Assistant blog or review the deeper evidence in Research.