Two ways to reach a decision
A deterministic authorization system evaluates a request against explicit rules and known inputs, and produces the same answer every time those inputs recur. A probabilistic system — a risk score, a machine-learning classifier, a language model asked to judge a request — produces an answer that depends on a model's internal state, which can shift with retraining, prompt changes, or randomness in inference, even when the visible inputs look identical.
Both can be useful. A risk score is a reasonable way to prioritize which transactions a human reviews. It is a poor way to decide, on its own, whether a consequential action is permitted, because the same request can plausibly receive different scores at different times for reasons nobody can fully reconstruct.
Reproducibility is what makes a decision reviewable
The practical reason this distinction matters is auditability. If a decision is deterministic, a reviewer can take the recorded inputs, the policy version that was in force, and the outcome, and confirm independently that the outcome was correct — without needing to trust whoever operated the system at the time. If a decision was probabilistic, the best a reviewer can do is note what the model returned; they cannot rerun the model against the same inputs and expect a guaranteed identical result, especially once time has passed and the model has changed.
This is why architectures built around explicit fail-closed rules keep AI and probabilistic judgment out of the specific step that decides whether an action is permitted, even when they use such tools elsewhere — to help a person draft a policy, or to investigate an incident after the fact. Judgment and investigation can tolerate uncertainty. The gate in front of a consequential action is a poor place for it.
Where probabilistic tools still belong
None of this argues against probabilistic tools generally. Anomaly detection, fraud scoring, and natural-language triage are legitimately useful for surfacing what a human should look at next. The distinction is about where in the decision path they sit: upstream of a human or a deterministic gate, informing a review, rather than standing in as the gate itself.