Rules vs. Goals vs. Cases: Which Policy Model Should Govern an AI Agent?

An AI agent cannot operate reliably on instructions such as “handle refunds appropriately” or “protect the customer relationship.” Those statements communicate intent, but they do not define how to decide. Operators must convert policy into a model the agent can apply under real conditions.

Three models dominate: explicit rules, ranked goals, and precedent cases. Rules specify permitted behavior. Goals tell the agent what outcomes to optimize. Cases demonstrate how judgment was applied before. Each resolves a different kind of uncertainty, and each fails in a different way.

The right choice depends less on model sophistication than on the shape of the business decision: how stable the policy is, how costly errors are, how often exceptions occur, and what evidence an auditor will expect.

The three policy models

Rule-based policy

A rule maps defined conditions to an action: if the order was delivered fewer than 30 days ago and the item is unopened, approve the return. Rules may also prohibit actions, set limits, require approvals, or route exceptions.

This model makes authority explicit. Its weakness appears at the boundary. Real cases rarely match every condition cleanly, and a growing rule set can accumulate contradictions, gaps, and brittle exceptions.

Goal-based policy

A goal model defines desired outcomes and their priority. A support agent might be told to resolve valid customer problems, minimize avoidable concessions, comply with consumer law, and protect account security—in that order.

Goals help an agent adapt when no exact rule exists. They also create interpretation risk. “Protect the customer relationship” can support several incompatible actions unless the agent has measurable constraints and a clear hierarchy.

Case-based policy

A case model gives the agent approved examples containing facts, decision, rationale, and outcome. When a new situation arrives, the agent retrieves similar cases, identifies meaningful differences, and proposes a consistent response.

Cases capture tacit judgment that is difficult to express as rules. Yet precedents can be obsolete, biased, or falsely similar. Retrieval quality becomes part of policy enforcement.

Head-to-head comparison

Where rules win—and where they break

Rules are strongest when a wrong action is expensive and the relevant facts can be verified before execution. Payment limits, access controls, regulated disclosures, approval thresholds, and data-retention requirements belong here. The agent should not “reason around” a prohibition because another outcome appears beneficial.

Consider an accounts-payable agent. A supplier requests a bank-account change and asks for an urgent payment. A reliable rule requires verification through an independently sourced contact channel before changing payment details. The agent may recognize the urgency, but urgency cannot override the control.

The failure mode is rule accretion. One exception produces a new clause; that clause conflicts with another; operators then add precedence rules. Eventually, nobody can predict the combined behavior. A rule system therefore needs more than prose:

• Scope: the processes, users, regions, and transaction types covered.
• Inputs: facts that must be present and their trusted sources.
• Priority: which rule prevails when conditions overlap.
• Version: when the rule became effective and what it replaced.
• Fallback: the action when facts are missing or rules conflict.

Choose rules when operators can define the decision boundary more confidently than the agent can infer it.

Where goals outperform scripts

Goals are useful when the path varies but success can be evaluated. Scheduling, inventory balancing, account planning, and incident response often fit this pattern. The agent must adapt to changing facts rather than follow one approved sequence.

Suppose an agent reschedules field-service visits after a technician becomes unavailable. A rigid script might move appointments chronologically. A goal-based system can account for safety severity, contractual response commitments, travel time, required skills, and customer availability. It can generate a better schedule because it understands what the schedule is for.

The design challenge is preventing one attractive objective from consuming the others. “Minimize travel” could delay a critical repair. “Maximize customer satisfaction” could trigger costly concessions. Goals need a hierarchy: hard constraints first, required service outcomes second, economic preferences third. Where two goals remain comparable, the agent needs an explicit tie-breaker or escalation condition.

Goal systems also require outcome feedback. If operators only review whether an action completed, they cannot detect a policy that consistently sacrifices a less visible objective. Review should examine the selected action, rejected alternatives, governing constraints, and observed result.

Where cases capture judgment better

Cases work well when experts agree on decisions more readily than they can articulate universal rules. Customer remediation, underwriting exceptions, procurement negotiations, and complex support escalations often contain this kind of knowledge.

Imagine a customer seeking compensation after several service failures. The formal refund rule covers the latest incident, but it ignores the sequence: two missed appointments, an incorrect charge, and a failed prior promise. A precedent showing how a similar cumulative failure was resolved may guide judgment better than an expanding catalogue of incident-specific rules.

A useful case is not a transcript. It is a structured decision record containing:

• Material facts at the time of the decision.
• The action selected and alternatives rejected.
• The policy and authority level involved.
• The rationale, including decisive distinctions.
• The eventual outcome, if known.
• An expiry condition or review date.

The key control is not similarity alone. The agent must state why the precedent applies and which differences could invalidate it. A case involving a strategic customer may be operationally similar but commercially inappropriate for a standard account. Without explicit distinctions, precedent becomes accidental policy.

A worked decision: handling a disputed renewal

Assume a software customer says an annual subscription renewed unexpectedly. The contract permits renewal, the reminder was sent, the product has not been used since renewal, and the account manager wants to preserve the relationship.

A rule-based agent checks notice timing, contractual terms, usage, refund window, and approval limits. If policy permits a refund for unused renewals within a defined period, it approves or routes the request. The result is consistent, but a novel contractual discrepancy may stop the workflow.

A goal-based agent weighs contractual compliance, retention value, concession cost, and fairness. It might propose a refund, credit, or revised term. This is more flexible, but the agent must not treat relationship value as permission to exceed financial authority.

A case-based agent retrieves prior disputes with similar notice, usage, and account history. It can explain that previous unused renewals were converted into shorter commitments, while cases involving active usage were denied. The risk is importing a precedent created under an older contract or by an unusually senior approver.

The strongest design combines them: rules establish legal and financial boundaries; goals rank acceptable remedies; cases calibrate judgment among those remedies. The agent should cite all three in its decision record.

How to choose the governing model

Use five questions to classify the workflow:

1. Can the decisive conditions be enumerated? If yes, begin with rules.
2. Does the environment change faster than policy can be rewritten? If yes, goals become more important.
3. Do experts rely on analogy when explaining decisions? If yes, cases likely contain essential knowledge.
4. Must an auditor reproduce the decision? Favor rules, or cases with explicit citations and version history.
5. Can outcomes reveal whether judgment was good? If yes, goal calibration is practical; if not, constrain discretion.

Do not force the entire agent into one model. Classify individual decisions. Identity verification can be rule-based while remedy selection is goal-based and unusual concessions are supported by approved precedents.

Which approach should you pick?

Pick rules when the workflow is regulated, permission-sensitive, financially bounded, or stable enough to specify. They are the correct foundation for prohibitions and transaction limits.

Pick goals when the agent must adapt plans to changing conditions and outcomes can be evaluated. Keep hard constraints outside the optimization logic, and define priority among competing objectives.

Pick cases when expert judgment depends on patterns, cumulative context, or distinctions that resist clean codification. Curate precedents as policy assets rather than dumping historical conversations into retrieval.

Pick a hybrid for most consequential operational agents. Use rules to define the perimeter, goals to choose within it, and cases to make the choice consistent with institutional judgment. The governing principle is simple: rules determine what the agent may do, goals determine what it should prefer, and cases show what good judgment has looked like before.

This post was drafted with AI assistance and reviewed against our editorial policy before publication. Corrections are made at the source, on the page, with the date shown.