- Published
- Updated
An expert may disagree with an automated recommendation. This does not automatically establish either a system error or noncompliance by the person. A reasoned override record should explain which additional information changed the decision and who took responsibility for it.
Scope of the decision
Do not reduce the record to selecting an other category. Different reasons may include missing context, incorrect input, an appropriate exception or model error. The purpose is not to force agreement with automation, but to make the divergence understandable and available for appropriate review.
Worked review
For example, a triage tool recommends low priority while a reviewer requests prompt assessment because of new source information. The record identifies that evidence and the decision owner. Later oversight examines why comparable information was unavailable to the tool instead of treating a high override rate as employee failure.
Practical checks
- Preserve the automated recommendation, final human decision and reason separately.
- Link additional evidence used for the authorised decision appropriately.
- Examine repeated differences for connections to inputs, rules or use boundaries.
Acceptance and limits
Acceptance requires an explainable authorised decision and its supporting evidence. Long free text is not inherently a better rationale; describe the important difference concretely. Override trends may inform improvement, but should not reduce clinical or quality judgement to one percentage.
Frequently asked questions
- Does a high override rate mean poor performance?
- Not on its own. Additional information and appropriate exceptions can also raise the rate.
- Is overwriting the recommendation sufficient?
- No. Keep the relationship between the original advice and final decision traceable.