Primary Metric and Guardrail Metric

A primary metric expresses the improvement a decision is chiefly trying to achieve. Guardrail metrics track outcomes that must remain acceptable while pursuing it. These roles are chosen for a particular decision; a metric can be primary in one project and a guardrail in another.

Suppose a team runs a trial of an assistant that helps clerks process supplier invoices. The primary metric could be mean clerk handling minutes per eligible completed case. Guardrails are tracked separately: errors in the invoice packets (the invoice and its supporting documents) sent on for approval, unauthorized disclosures, and extra work pushed onto the team that receives the goods and checks the invoice against them. A fast response does not compensate for exposing another branch’s records.

Define each metric’s population, calculation, observation window, limit, and consequence before examining trial results. Missing quality records should not be treated as no errors. An observed guardrail breach may require stopping the trial, revising the proposal, or withholding a success decision, depending on the agreed rule.

Passing measured guardrails in a small sample is not proof of zero production risk. Keep uncertainty visible and assess the primary outcome and guardrails together instead of hiding a failed condition inside an average score.

Examples and calculations here are illustrative.

References:

• GOV.UK service measurement guidance