Offer covered help to frustrated customers

Learning objective

When a customer expresses frustration without asking for a human, acknowledge it, offer an in-policy solution, and escalate if subsequently requested.

Core idea

Frustration is evidence about the customer’s experience, not a choice of support destination. If policy and tools support a useful resolution, acknowledge the frustration and offer that resolution. If the customer then asks for a person, switch to handoff immediately.

These are sequential decisions, not competing personality styles. Empathy belongs in both automated help and escalation. The question is whether the customer has requested a person and whether the proposed action is authorized and executable. See 5.2.1 for independent escalation triggers and 5.2.2 for the human-request path.

How it works

Acknowledge the problem without changing authority

A brief acknowledgment should identify the experience: repeated late deliveries, a confusing charge, or an unresolved return. It should not promise an exception, invent facts about the order, or imply that a human is already available.

For example, “Two late deliveries is frustrating. I can check whether this order qualifies for a delivery-fee refund” acknowledges the complaint and names an available next step. If the relevant eligibility facts are already verified, the offer can be more direct: “This order qualifies for a delivery-fee refund. Would you like me to submit it?”

The difference is evidence, not tone. Do not assert eligibility before the lookup returns it. Equally, do not ask the customer to repeat verified details merely to sound attentive. The response should move toward an actual permitted solution.

Offer the specific in-policy action

In the Messages API, the top-level system parameter can state the standing response rule: acknowledge frustration, offer the available covered remedy, and honor later human requests. The tool’s description should identify when the remedy is appropriate rather than treating any negative emotion as a reason to transfer.

Suppose your application defines lookup_order, process_refund, and escalate_to_human as custom client tools in tools. Each uses an input_schema for the operation’s arguments. These names are retailer-specific examples, not Anthropic built-ins. The actual refund authority remains with the application’s rules and execution checks; see 1.5.3.

An offer is not a completed refund. If your workflow requires customer confirmation, wait for it before executing process_refund. If the tool then succeeds, describe the confirmed result. If it fails and recovery cannot advance the case, the separate blocked-progress trigger applies. A warm response does not convert a failed operation into resolution.

The exam objective asks for an in-policy solution, not an arbitrary concession. Giving a competitor price match to calm the customer is still outside a policy that covers only own-site price changes. Frustration does not expand the agent’s authority.

Re-evaluate the next customer turn

When the customer accepts the offer, proceed through the authorized workflow. When the customer provides missing details, use them to determine the covered next step. When the customer asks for a representative, initiate escalate_to_human rather than trying another offer.

A refusal of one remedy is not always a human request. “That won’t fix the damaged item” supplies a reason the proposed remedy is insufficient. Check whether another covered remedy applies. “No, get me a representative” specifies the destination and ends the attempt to resolve autonomously.

Do not require a second request when the first explicit request appears. The exam guide’s wording about a customer reiterating their preference describes the transition from offered resolution to requested escalation. It does not authorize an agent to ignore the first unambiguous request for a person.

Keep emotion and routing distinct

A frustrated customer can still have a policy-gap request or a case that cannot progress. Then escalate for that independent reason. The rule is not “keep angry customers automated at all costs.” It is “do not infer a transfer request from frustration alone.”

When a message hints at a person without clearly requesting one, such as “Can someone actually sort this out?”, treat it as frustration: acknowledge the problem and offer the covered remedy. Do not replace that offer with a question about whether the customer wants a human. If the reply restates the wish for a person—“No, I mean a real person”—initiate handoff at once without another offer.

This sequence follows the exam guide’s instruction to acknowledge frustration, offer an in-capability resolution, and escalate if the customer reiterates their preference. It does not apply to a first unambiguous request for a person, which already requires immediate handoff.

Key distinctions and trade-offs

The following customer replies call for different next actions:

Offering immediate covered help can avoid unnecessary waiting. Persisting after a human request imposes a different cost: the system continues a workflow the customer has declined. Good routing preserves both the chance of useful automation and the customer’s choice.

Anthropic’s customer-support guide separates sentiment maintenance from escalation efficiency. Measure whether the agent acknowledges frustration and whether it chooses the right destination as distinct outcomes. Positive wording alone does not establish either a correct refund or an honored request.

Architecture / decision scenario

A retailer allows delivery-fee refunds for verified late arrivals. A customer writes, “Second late parcel this month. I’m fed up.” The agent acknowledges the repeated inconvenience and, using the established order evidence, offers the eligible delivery-fee refund. It does not call escalate_to_human merely because the message is negative.

Run three continuations. In the first, the customer accepts, so the application executes process_refund and the agent reports its result. In the second, the customer says the item is damaged; the agent checks the applicable damage remedy instead of assuming the fee refund resolves the problem. In the third, the customer says, “No, I want a person.” The next operation is handoff, with the prior offer included in the known context.

A fourth variant asks for a competitor price match. Even with the same frustrated tone, that request falls outside the retailer’s supplied adjustment rule. Route it for a policy-gap decision. This prevents the empathy instruction from becoming an authorization to invent concessions.

Inspect each trace after the reply. The correct first-turn offer does not excuse a later attempt to persuade the customer away from a representative. Conversation state changes when the customer changes the request.

Common exam traps

• Escalating every negative message without checking the requested destination.
• Ignoring frustration while making a technically correct offer.
• Offering an unauthorized concession to improve sentiment.
• Continuing with another remedy after the customer requests a person.
• Treating rejection of one remedy as automatic proof of a human request.

What to remember

Acknowledge frustration and offer a concrete, authorized remedy when one is available. Re-evaluate each reply. A later explicit request for a person triggers immediate handoff, while policy gaps and blocked progress remain independent reasons to escalate.

Knowledge check

References

• Anthropic, Claude Certified Architect – Foundations Exam Guide
• Anthropic, Create a Message
• Anthropic, Define tools
• Anthropic, Tool use with Claude
• Anthropic, Customer support agent