A customer demands a full refund for a product they used for 11 months. Your policy says no refunds after 30 days. The product worked fine. They just don't need it anymore.

If you refund them, you keep one customer happy. But 5 other customers who were denied refunds under the same policy will feel cheated if they find out. And if word spreads that you refund upon request regardless of policy, your refund rate will climb, potentially costing tens of thousands per year.

If you don't refund them, you lose one customer and potentially get a negative review. But you maintain policy consistency, which protects your margins and treats all customers equally.

This is the trolley problem of customer support. Every policy decision involves trade-offs between individual outcomes and systemic consequences.

Individual Justice vs Systemic Fairness

Most support interactions pit two values against each other:

Individual justice. This specific customer has a legitimate grievance. They should get the outcome that's fair for their situation, regardless of what the policy says.

Systemic fairness. Policies exist to treat all customers equally. Making exceptions for loud or persistent customers rewards the wrong behavior and punishes the quiet, patient majority.

These values conflict constantly. The customer who escalates to the CEO gets a refund. The customer who accepted the policy didn't. Both had the same situation. One got a better outcome because they were louder. That's individually just (the loud customer's situation warranted a refund) and systemically unfair (the quiet customer's identical situation didn't get one).

The Precedent Problem

Every support decision creates a precedent. "We refunded a customer who asked nicely after 6 months" becomes "customers who ask nicely get refunds regardless of the time window." Maybe not explicitly. But agents remember. The next time a similar request comes in, they think: "We did it last time, why not now?"

Over time, exceptions become the rule. The 30-day refund policy becomes a 90-day refund policy (in practice) because enough exceptions were made. The policy on paper no longer matches the policy in practice. Nobody updated the documentation. Agents apply it inconsistently, which creates the worst outcome: some customers get refunds and some don't, for no clear reason.

Decision Frameworks

The utilitarian approach: make the decision that produces the best outcome for the greatest number. A $50 refund to a persistent customer that prevents a negative review (protecting 100 future customers who'd read it) might be utilitarian. A $500 refund to a customer who used the product for 11 months is harder to justify on utilitarian grounds (the precedent cost exceeds the individual benefit).

The Rawlsian approach (the "veil of ignorance"): design the policy as if you don't know whether you'd be the customer or the company. Behind the veil, you don't know if you're the customer asking for a refund or the company paying it. Would you design a 30-day refund policy? Probably. Would you design an exception process for unusual circumstances? Also probably. The Rawlsian policy has clear rules and a clear exception process.

The virtue ethics approach: what decision would a good company make? Not "what does the policy say?" but "what's the right thing to do?" This approach often favors the customer but has the weakness of being subjective. Two people can disagree on what a "good company" would do.

The Practical Framework

For most support teams, a three-tier decision model works:

Clear policy (80% of cases). The policy applies straightforwardly. Refund within 30 days? Yes. Refund after 6 months? No. The agent applies the policy. No judgment needed.

Manager exception (15% of cases). The situation has legitimate nuance. The customer's circumstances are unusual. The agent can't justify the exception but can't justify the denial either. A manager reviews and makes a judgment call, documented in the ticket for precedent reference.

Executive override (5% of cases). High-value account, potential PR issue, legal threat, or a situation that exposes a genuine policy gap. An executive makes the call, with the understanding that this is a one-time decision, not a policy change.

The key: document every exception with the reason. "Refunded $500 for [customer] because their onboarding was mishandled by our sales team, which led them to use the wrong plan for 6 months." That's not a precedent for "everyone gets 6-month refunds." It's a precedent for "we fix our own mistakes."

What AI Can Do

AI can't make ethical decisions. But it can provide the data that informs them.

Supp's classification identifies the intent (refund request) and context (time since purchase, account value, sentiment). It can flag whether the request falls within policy (auto-process) or outside policy (route to manager).

The classification also provides historical context: has this customer made similar requests before? What's their lifetime value? What's their support history? This context helps the manager make a more informed exception decision.

For the 80% of cases that are clear-cut, AI handles them automatically. For the 15% that need judgment, AI provides the context. For the 5% that need an executive, AI flags the severity.

The trolley problem never gets a clean answer. But having better data, faster routing, and documented precedents makes the decision less arbitrary and more defensible. And in customer support, that's about the best you can do.