What you will be able to do
- Name the three conditions that should make a support agent escalate to a human, and tell them apart from conditions that only look like they should
- Handle an explicit request for a human differently from frustration where the customer has not asked for one
- Choose between handing off immediately and offering to resolve, and escalate when the customer repeats a preference for a person after the offer
- Recognise a policy gap and send it to a human, instead of declining it or approving it by stretching the nearest rule
- Explain why sentiment and self-reported confidence are poor escalation triggers
- Respond to a lookup that returns several matching customers by asking for another identifier
- Write escalation criteria and escalate-versus-resolve examples into a system prompt
Key concept
Escalation trigger — An observable condition in the conversation that sends a case to a human. There are three: the customer explicitly asks for a person, policy is silent or ambiguous on what they want, or the agent cannot make meaningful progress. Tone, sentiment and the agent's own sense of difficulty are not triggers.
1.The three conditions that justify a handoff
A support agent can get escalation wrong in two directions. If it escalates too often, it becomes an expensive way to route tickets to people. If it escalates too rarely, it makes decisions it has no authority to make, or keeps a customer in a loop that cannot end. Escalation accuracy measures both errors: conversations that should have reached a human but didn't, and conversations that were handed off when they didn't need to be.
Three conditions justify a handoff, and each one can be observed in the conversation itself:
- The customer explicitly asks for a human. "Let me talk to a person" decides the matter. The agent has no case to evaluate first.
- Policy is silent or ambiguous on the specific request. The question is not how complex the case is. It is whether a written rule covers it. A complicated refund that policy fully covers is still the agent's job. A simple request that no rule addresses is not.
- The agent cannot make meaningful progress. It has tried the reasonable paths, such as the lookups, the tools and the questions the customer can answer, and it still cannot move the case forward. Escalating at that point is better than repeating the same attempts.
Complexity is missing from that list on purpose. Agents handle multi-step cases well as long as policy covers each step, so "this seems complicated" is not a reason to hand off. In exam answers, the option that escalates "because the case is complex" is usually a distractor. The option that escalates because policy does not cover the request, or because the customer asked for a person, is usually the right one.
| Condition in the conversation | Escalate? | What the agent does |
|---|---|---|
| Customer says they want a human | Yes, immediately | Hands off without investigating first |
| Request falls outside anything written policy addresses | Yes | Refers the case to someone with authority to decide it |
| Reasonable attempts made, still no way forward | Yes | Hands off with a summary of what was tried |
| Customer is frustrated but has not asked for a human | No | Acknowledges the frustration, resolves the issue, offers a human as an option |
| Case has many steps but policy covers each one | No | Works through it |
Checkpoint 1 of 9· Check yourself
A team is writing the handoff rules for a new customer support agent. Which design gets escalation right in both directions?
Escalation can fail in two ways: handing off too often or too rarely. The fix is a short list of triggers you can observe in the conversation, not 'everything' or 'nothing.' Complexity isn't a trigger: if policy covers every step of a multi-step case, the agent should handle it.
“Design your agents to flag specific situations for approval rather than proceeding automatically on everything or nothing.”Source: claude.com
2.Explicit requests for a human compared with frustration
Two customers can sound equally annoyed and still need opposite handling. The first says "I want a real person, not a bot": that is an explicit demand, and the agent hands off immediately. The second sounds irritated but asks for nothing in particular, and the issue is straightforward and within the agent's capability: the agent offers to resolve it. What separates them is what the customer asked for, not how loudly.
When a customer demands a human, the agent does not try one quick fix first, ask a few diagnostic questions, or point out that the issue is easy. Suppose a customer writes "I want a real person, not a bot" about a routine password reset. They get a person, even though the agent could have reset the password in two turns. If the agent investigates first, it has overridden a clear preference the customer stated.
A frustrated customer who has not asked for a human gets three things in one reply when the issue is within the agent's capability. The agent acknowledges the frustration, resolves the issue, and offers a human as an option. Each part rules out a common wrong answer:
- Acknowledge. Skipping the tone and doing the task mechanically completes the request but ignores how the customer feels. If the team's playbook says to acknowledge frustration, leaving that out is a mistake.
- Resolve now. Sending an irritated customer to a queue for something the agent can do in one step makes them wait longer, which is exactly what upset them.
- Offer, don't require. "Happy to connect you with someone if you'd prefer" leaves the choice with the customer. "Before I continue, please confirm you don't want a staff member" makes the customer answer an extra question before getting help they never asked to delay.
The offer is made once, and the customer's answer settles what happens next. If they accept the help, the agent carries on. If they answer with "no, I want a person", or ask for a human again later, they have reiterated their preference, which makes it an explicit request, and the agent hands off without further argument. Escalation follows the stated preference, whether it comes in the first message or the fifth, and never the mood alone.
Checkpoint 2 of 9· Check yourself
Earlier in a chat, a customer sounded irritated but never asked for a person. So the agent acknowledged the frustration and offered a human as an option. The customer now replies: "No, just put me through to a person." The billing-date change they need takes one tool call. What should the agent do next?
Once the customer says they want a person, it is an explicit request, whether it comes in the first message or the fifth. The agent hands off right away. It does not do a quick fix first or ask the customer to justify the request, even though the issue is simple.
“provide straightforward pathways to human specialists when needed”Source: claude.com
Checkpoint 3 of 9· Exam question
A healthcare scheduling assistant follows a configured escalation policy: when a request is within the assistant's capability and the patient has not asked for a staff member, the assistant should acknowledge any expressed frustration, complete the request directly, and offer escalation to a staff member as an option rather than a requirement -- reserving mandatory escalation for requests it cannot complete, cases needing clinical judgment, or an explicit request for a human. A patient messages the assistant in a clearly irritated tone about needing to reschedule a routine appointment with no clinical urgency. The assistant can complete the rescheduling directly, and the patient has not asked to speak with a staff member. Following its configured policy, what should the assistant do?
Correct answer: A — Acknowledge the frustration, go ahead and reschedule the appointment now, and offer to escalate to a staff member.
- A. This is the only option that carries out all three parts of the assistant's configured policy: it acknowledges the expressed frustration, it completes the reschedule directly (the request is within capability and the patient has not asked for a human, so going ahead with the reschedule is the required action rather than merely saying it could be done), and it offers escalation as an option rather than making it mandatory or leaving it unoffered.
- B. The policy reserves mandatory escalation for three conditions: the assistant cannot complete the request, the situation needs clinical judgment, or the patient explicitly asks for a staff member. An irritated tone by itself is none of those three, so escalating immediately here goes against the policy rather than following it.
- C. The policy requires acknowledging any expressed frustration alongside completing the request, not instead of it. Proceeding with the reschedule while treating the patient's tone as irrelevant leaves out that half of the policy, even though completing the reschedule itself is correct.
- D. The policy makes escalation something the assistant offers, not a condition the patient must first rule out. Requiring the patient to confirm they don't want a staff member before the reschedule happens adds a step the policy does not call for and delays the direct action the policy calls for when a request is within capability.
Sources1
3.Policy that is silent or ambiguous
Silence in a policy does not answer the customer's question, so the agent should not treat it as an answer. Take a retailer whose written policy covers price adjustments only when the store's own website lowers a price within 14 days of purchase. A customer asks the agent to match a lower price on a competitor's site. The policy does not mention competitors at all. It neither allows nor forbids a match.
There are three tempting ways to handle this case, and all of them are wrong:
- Decline. This treats the own-site rule as the complete list of allowed adjustments. The policy never says that, so the agent has invented a prohibition.
- Approve. This extends the own-site rule to competitor prices. The agent has invented a permission and committed the company to something nobody approved.
- Verify, then approve. Asking for a screenshot of the competitor's listing confirms the price. It does not give the agent authority to match it. Checking a fact does not fill a gap in policy.
The right move is to refer the case to a supervisor or the team that owns pricing policy, since they can decide cases the written rules don't address. The agent can tell the customer that the request is being reviewed by someone who can make that call. The same reasoning applies to anything at the edge of policy, such as goodwill credits, exceptions to a return window, or a combination of discounts that no rule mentions. The agent's authority ends where the written policy ends.
Checkpoint 4 of 9· Match them up
Written policy covers price adjustments only when the store's own site lowers a price. A customer asks for a match to a competitor's lower price. Match each way the agent could respond to what that response actually does.
Tap a term, then the definition that fits it.
The policy doesn't mention competitors, so it neither allows nor forbids a match. Declining, approving, and approving after verifying the price all mean the agent decides on the company's behalf. Only referring the case sends the decision to someone with authority to make it.
“Autonomy is powerful, but some decisions require human judgment. Design your agents to flag specific situations for approval”Source: claude.com
Checkpoint 5 of 9· Exam question
A retail customer asks a support agent to match a lower price they found on a competitor's website. The store's documented policy only describes price adjustments when the store's own website lowers a price within 14 days of purchase; it does not mention competitor pricing at all. How should the agent proceed?
Correct answer: A — Refer the request to a supervisor, who can decide cases the written policy does not address
- A. Correct - the written policy covers only own-site price drops within 14 days and says nothing about competitor prices, so the agent has no basis to grant or refuse and should pass the case to someone with authority to rule on it.
- B. Incorrect - the policy describes one kind of adjustment but never prohibits others. Reading it as a complete list of permitted adjustments invents a restriction that was not written, and the agent cannot settle the gap alone.
- C. Incorrect - the own-site provision applies to the store's own price drops within 14 days. Stretching it to cover competitor prices is a policy extension the agent is not authorized to make, because the policy never addresses that case.
- D. Incorrect - verifying the competitor's price confirms a fact but does not create authority to match it. Approving afterwards still means the agent decides a policy question the documented rules do not address.
4.Why sentiment and self-reported confidence are poor triggers
Escalation decisions are often built on signals that are easy to compute but measure the wrong thing. Sentiment scoring is the most common example. An angry customer may have the simplest request of the day, like a wrong t-shirt size covered by a standard 30-day exchange. A calm, polite customer may be asking for something no policy covers. If negative sentiment sends cases to humans, the agent escalates easy cases and keeps the gap cases that actually needed a person.
The right response to an angry customer with a straightforward issue is to acknowledge the anger and offer to resolve it, since the agent is fully able to. Immediate escalation is reserved for the customer who explicitly demands a person. Tone can colour the reply but cannot decide the route. Agents tend to lean the other way unaided: when customers express dissatisfaction, the model may address the emotion at the expense of the underlying problem, so the prompt should say when to prioritise which. Sentiment is better used as a measure of outcome than as a reason to route a case. A good conversation should end with the customer feeling better than at the start. Escalation accuracy is a separate measure, judged against whether a case actually needed a human.
Asking the model to rate its own confidence and escalating below some threshold has the same problem. A raw confidence number the model reports about itself has not been checked against how often it is actually right. It may be high on a policy gap the model has quietly misread and low on a routine case worded in an unusual way. Calibrated confidence scores for review routing is a different technique: it checks scores against labeled outcomes before trusting them. An uncalibrated score, like a sentiment score, is not a reliable stand-in for how complex a case is. The three triggers work better because each one can be checked in the conversation: did the customer ask, does a rule cover this, can the agent move forward.
Checkpoint 6 of 9· Check yourself
A support team's prototype routes a conversation to a human whenever its sentiment score turns negative or the model rates its own confidence below 0.6. In testing, a furious customer with a routine 30-day size exchange gets escalated. A polite customer whose request no written policy covers stays with the agent. What should the team change?
Sentiment and raw, uncalibrated self-reported confidence don't measure whether a case needs a human, and combining two unreliable proxies doesn't make them reliable. The three triggers can be checked in the conversation itself. Sentiment is better used to judge outcomes, comparing how the customer felt at the start with how they felt at the end.
“Use sentiment analysis tools to measure sentiment at the beginning and end of each conversation. Aim for maintained or improved sentiment in 90% of interactions.”Source: platform.claude.com
5.Ask for an identifier when a lookup returns several matches
A lookup that returns more than one customer has not found the customer. Take an airline agent that searches by passenger name and route for a baggage claim and gets back three records, each with a different booking reference. It cannot tell from the records which one belongs to the person it is talking to. Any choice it makes is a guess that could expose another person's data or change their records.
None of them. Each detail is a heuristic, a plausible signal that says nothing about who is in the chat. The agent asks the customer for one more identifier, such as a ticket number, booking reference or phone number, and continues only after exactly one record matches.
Heuristic selection fails quietly, which is what makes it dangerous. Picking the most recent departure, the most complete profile, or the first result usually looks fine in testing because test data rarely contains namesakes. In production the wrong pick lets the agent discuss one passenger's baggage claim with another. Merging the records into a summary is worse, because it creates a customer who doesn't exist. Asking costs one turn. Guessing wrong can cost a privacy incident.
This is a narrow case where asking is better than acting. Normally an agent should use its tools before asking the customer for anything. Here it already used them, and the result is ambiguous in a way only the customer can resolve. The instruction belongs in the system prompt: when a lookup returns more than one match, ask for an additional identifier and never choose among the matches. The same applies when a telecom lookup by last name and area code returns two similar accounts. The agent asks for the phone number or account number before discussing either account.
Checkpoint 7 of 9· Check yourself
In production, a hotel-booking support agent searches by guest name and check-in city and gets back two reservations with different booking references. Logs show it picked the one with the most recent activity and discussed it with the wrong guest. Which system-prompt instruction fixes this?
Several matches means the customer hasn't been found yet. No heuristic shows which record belongs to the person in the chat, and summarizing every record would expose other guests' data. The agent has already used its tools, and only the customer can resolve this, so it asks for one more identifier.
“Claude asks a clarifying question only when proceeding would clearly waste effort or go in an entirely wrong direction”Source: platform.claude.com
Checkpoint 8 of 9· Exam question
An airline support agent searches for a passenger by name and flight route to check a baggage claim. The lookup tool returns three passenger records with the same name on that route, each with a different booking reference. What should the agent do?
Correct answer: A — Ask the customer for one more identifier, such as a ticket number or phone number, then continue the claim review
- A. Correct - three passengers share the name and route, so nothing in the lookup shows which record belongs to this customer. Getting a further identifier from the customer settles it before the claim is reviewed.
- B. Incorrect - all three passengers travelled the same route, and a later departure date says nothing about who the caller is. It is a guess that could open and expose another passenger's baggage claim.
- C. Incorrect - how much contact information a profile holds reflects how the account was set up, not whether it belongs to the caller. The agent would still be guessing between three same-name passengers.
- D. Incorrect - only one of the three records belongs to this customer. Merging them mixes in details of two other passengers, which gives a wrong claim summary and discloses their information.
Sources4
6.Writing escalation criteria and examples into the system prompt
An agent applies these rules consistently only if the system prompt spells them out. "Escalate when appropriate" leaves the model to supply its own threshold, and that threshold tends to drift toward tone and complexity. Write the criteria the way the earlier sections state them: explicit request means hand off immediately without investigating, silent or ambiguous policy means refer, no meaningful progress means hand off with a summary, frustration alone means acknowledge, resolve and offer, and multiple lookup matches mean ask for another identifier.
Put few-shot examples right next to those criteria, and write them in pairs that differ by one thing. One example shows an angry customer with a routine exchange: the agent acknowledges, resolves and offers a human. The paired example shows the same customer saying "get me a manager": the agent hands off immediately. A policy pair contrasts a request the policy covers with one it does not mention. A lookup example shows one match leading to the agent proceeding, and its partner shows two matches leading to the agent asking. Pairs like these teach the boundary between cases. Examples that are all escalations, or all angry customers, teach the model to associate escalation with the wrong feature.
TASK_SPECIFIC_INSTRUCTIONS = " ".join(
[
STATIC_GREETINGS_AND_GENERAL,
STATIC_CAR_INSURANCE,
STATIC_ELECTRIC_CAR_INSURANCE,
EXAMPLES,
ADDITIONAL_GUARDRAILS,
]
)The guardrail section is where escalation criteria go, and the guide's own guardrail on authority shows the pattern: the agent must not make promises or agreements it isn't authorized to make. Add the reason behind each rule too. For example, explain that policy gaps go to a supervisor because only a supervisor can commit the company to an exception. A model that knows the reason applies the rule more reliably to cases the examples didn't cover.
Checkpoint 9 of 9· Match them up
Four drafts of the escalation section of a support agent's system prompt are under review. Match each draft to the behavior it is most likely to produce.
Tap a term, then the definition that fits it.
Vague criteria leave the threshold up to the model, and one-sided examples teach the wrong feature. Contrasting pairs mark the boundary between cases, and giving the reason behind each rule lets the model generalize to new cases.
“Diverse: Cover edge cases and vary enough that Claude doesn't pick up unintended patterns.”Source: platform.claude.com
Sources5
Sources
(7)
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“This measures Claude's ability to recognize when a query needs human intervention and escalate appropriately.”
↩︎ The three conditions that justify a handoff“This assesses Claude's ability to maintain or improve customer sentiment throughout the conversation.”
↩︎ Explicit requests for a human compared with frustration“Don't make promises or enter into agreements it's not authorized to make.”
↩︎ Policy that is silent or ambiguous“Track the percentage of correctly escalated conversations versus those that should have been escalated but weren't.”
↩︎ Why sentiment and self-reported confidence are poor triggers“This measures Claude's ability to recognize when a query needs human intervention and escalate appropriately.”
↩︎ Exam trap 1“Track the percentage of correctly escalated conversations versus those that should have been escalated but weren't.”
↩︎ Exam trap 2“Don't make promises or enter into agreements it's not authorized to make.”
↩︎ Exam trap 3“Use sentiment analysis tools to measure sentiment at the beginning and end of each conversation. Aim for maintained or improved sentiment in 90% of interactions.”
↩︎ Checkpoint - 2.https://claude.com/blog/building-ai-agents-in-financial-servicesSecondary source
“Escalation protocols that automatically route complex cases to appropriate specialists based on clear criteria”
↩︎ The three conditions that justify a handoff“provide straightforward pathways to human specialists when needed”
↩︎ Checkpoint - 3.
“How are edge cases or ambiguous tickets handled?”
↩︎ Policy that is silent or ambiguous“When customers express dissatisfaction, Claude may prioritize addressing the emotion over solving the underlying problem.”
↩︎ Why sentiment and self-reported confidence are poor triggers - 4.
“Claude only asks upfront when the request is genuinely unanswerable without the missing information”
↩︎ Ask for an identifier when a lookup returns several matches“Claude only asks upfront when the request is genuinely unanswerable without the missing information”
↩︎ Exam trap 4 - 5.https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practicesOfficial docs
“Diverse: Cover edge cases and vary enough that Claude doesn't pick up unintended patterns.”
↩︎ Writing escalation criteria and examples into the system prompt“Examples are one of the most reliable ways to steer Claude's output format, tone, and structure.”
↩︎ Writing escalation criteria and examples into the system prompt
Also cited
“Design your agents to flag specific situations for approval rather than proceeding automatically on everything or nothing.”
↩︎ Key concept“Design your agents to flag specific situations for approval rather than proceeding automatically on everything or nothing.”
↩︎ Checkpoint“Autonomy is powerful, but some decisions require human judgment. Design your agents to flag specific situations for approval”
↩︎ Checkpoint“Claude asks a clarifying question only when proceeding would clearly waste effort or go in an entirely wrong direction”
↩︎ Checkpoint