What is the human final call?
The human final call is the point beyond which a model recommends and a named person decides. The boundary is set in advance, with the categories of decision reserved to a person stated explicitly. It is recorded per decision, so the record shows who exercised it and what they changed.
Also called: Final human decision
Why a named person is a regulatory requirement rather than a courtesy
Several regimes governing consequential decisions require that a person, identifiable by name, is answerable for the outcome. That requirement is unsatisfied by a reviewer who approves a queue of recommendations at a rate of several a minute, which is what unconstrained automation produces in practice. Meaningful human decision needs both the information to disagree and the authority to do so, and each has to be engineered deliberately. An approval rate that nobody could sustain is the first thing an inspector looks at.
- Accountability attaches to a person, so the record has to name one.
- Rubber stamping satisfies the form of human oversight while defeating its purpose.
- A stated boundary tells reviewers which decisions are genuinely theirs.
- Override data is the most direct evidence of whether oversight is real.
How the human final call works
- Categories reserved to a human decision are defined before deployment.
- The model produces a recommendation with its confidence and its linked evidence.
- Reserved and low confidence cases are routed to a reviewer with capacity to consider them.
- The reviewer may amend or reject the recommendation, with a reason required for any change.
- The decision records the reviewer's identity alongside the recommendation they received.
- Override rates and review durations are monitored as oversight indicators.
How the human final call is evaluated
Human oversight is assessed on whether the person could realistically have decided otherwise. Measure the following on routed decisions:
- median time a reviewer spends on each routed decision
- override rate, reported separately by reviewer
- proportion of overrides carrying a recorded reason
- outcomes of overridden decisions compared against accepted recommendations
- share of reserved categories actually reaching a person
Limitations, risks and common failure modes
- Reviewers approve almost everything, which the override rate exposes immediately.
- Oversight is assigned to staff without the seniority to reject a recommendation.
- The recommendation is presented before the evidence, which anchors the reviewer.
- Volume targets make considered review impossible while leaving the policy intact.