Typed Decisions: Why "Allow, Deny, Condition, Escalate" Beats a Score

In most credit operations, the central object is a number. A score of 0.73. A probability of default of 2.4%. A risk rating of B+. Systems compute it, dashboards display it, analysts discuss it.

But a number is not a decision. At some point, someone or something has to turn it into an action: approve or not, how much, under what terms, now or later. In most institutions, that translation is the least structured part of the process. It lives in cutoffs, informal ranges and analyst judgment.

There is a better way to design it. Instead of treating decisions as points on a continuum, treat them as types: a small, defined set of distinct outcomes, each with its own meaning, consequences and requirements. In a governed decision system, there are four.

The four types

• Allow. The action proceeds as requested. The evidence is sufficient, the case is within policy and within the authority of whoever, or whatever, is deciding.

• Deny. The action does not proceed. The case conflicts with policy, or required evidence is missing and can't be obtained. The reason is recorded and can be explained.

• Condition. The action may proceed only once a defined requirement is met: additional evidence, a human confirmation, a reduced amount, collateral, a step-up authentication. The decision specifies exactly what the condition is and what happens when it is or isn't met.

• Escalate. The decision is referred to a human or body with the appropriate authority, because the case exceeds the current decider's authority, reaches a defined threshold, falls outside known patterns or requires judgment the policy reserves for people.

These are not gradations. A conditioned decision is not "a weaker allow," and an escalation is not "almost a deny." Each type is a distinct outcome with its own downstream process.

Why scores alone don't work as decisions

Using a score as the decision, with an implicit cutoff, creates four problems.

• The middle disappears. A cutoff divides cases into two groups. But many real cases belong in neither: the risk is acceptable if a document is updated, or the amount is fine at a lower level, or the case is unusual enough to need a human look. A binary cutoff forces them into approve or deny, and both answers are wrong.

• Uncertainty is discarded. A score of 0.73 with high confidence and a score of 0.73 with low confidence look identical at a cutoff. With typed decisions, low confidence can lead to a condition or escalation, while high confidence leads to allow or deny.

• Consequences are implicit. A score doesn't say what should happen next. A typed decision does: an allow triggers execution, a condition triggers a request, an escalation routes the case to a specific queue with specific authority.

• Consistency can't be tested. It is hard to verify that equivalent cases received equivalent treatment when treatment is a continuous number filtered through informal judgment. With typed outcomes, consistency becomes measurable: for equivalent evidence and the same policy version, did cases receive the same type?

How the types connect to the rest of the system

Typed decisions are not just labels. They are the connective tissue between evidence, policy, authority and execution.

• Evidence → type. Missing evidence doesn't silently degrade a score. It maps to a defined type: condition when the evidence can be obtained, deny when it is essential and unavailable, escalate when a human should judge.

• Uncertainty → type. The policy can use the uncertainty attached to model signals. A central estimate within appetite with an upper bound outside it can map to condition, such as a partial approval, instead of a full allow.

• Authority → type. When a case exceeds the authority of the current decider, whether a system operating under a mandate or a junior analyst, the type is escalate, routed to the level that holds the necessary authority.

• Type → execution. Only allow, and a condition once satisfied, lead to execution. Execution systems can enforce this: an action without an authorizing decision of the right type doesn't run.

An illustrative example

The following cases are illustrative.

An institution's policy for small-business credit-line increases uses a model's default estimate and its uncertainty range, plus evidence requirements:

Case	Evidence and signal	Type	What happens
A	Complete, current evidence; estimate 1.2%, narrow range	Allow	Increase executed
B	Complete evidence; estimate 2.1%, range up to 6.8%	Condition	Partial increase now; remainder after updated statements
C	Financial statements missing and not obtainable	Deny	Reason recorded and explained to customer
D	Sector outside known patterns; request above branch authority	Escalate	Routed to regional credit committee

Four cases, four different outcomes, each with a clear meaning and a defined next step. A single cutoff would have turned them into two.

Typed outcomes for agents

Typed decisions become even more important when AI agents act on behalf of customers.

An agent proposing a payment needs a clear answer it can act on:

• Execute (the agentic form of allow): the action proceeds within the mandate.

• Condition: the action waits for a specific requirement, such as the principal's confirmation.

• Escalate: the action goes to a human with authority.

• Block (the agentic form of deny): the action does not proceed, and the reason is recorded.

An agent can't interpret an ambiguous score. It can act correctly on a typed outcome.

Measuring with types

Once decisions are typed, a set of operational metrics becomes available:

• Type distribution by product, segment and channel: how many decisions are allowed, denied, conditioned, escalated?

• Condition fulfillment rate: how often conditions are met, and how long it takes.

• Escalation resolution: how escalated cases are ultimately decided, and whether escalation thresholds are well calibrated.

• Consistency: whether equivalent cases receive the same type under the same policy version.

• Outcome by type: how decisions of each type perform over time.

These metrics show not only how well the institution decides, but where its policy needs adjustment.

Decisions as commitments

A score describes risk. A decision commits the institution to a course of action. Treating that commitment as a typed object, with defined meaning and defined consequences, is what makes decisions consistent, explainable, governable and measurable.

A score is a number. A decision is a commitment. Commitments deserve structure.

Want to see how typed decisions would work in your credit operation? Talk to our team →