AnthropicClaude Certified Architect – Foundations2026-105.2 Design effective escalation and ambiguity resolution patterns
Route on evidence rather than scores
Completed by the knowledge check
Learning objective
Avoid treating sentiment or uncalibrated self-reported model confidence as proxies for case complexity.
Core idea
A routing signal is useful only if it answers the routing question. Sentiment describes how the customer feels. A model’s self-reported confidence describes its generated assessment of certainty. Neither, by itself, establishes whether policy authorizes the requested action, whether the customer wants a person, or whether the workflow can progress.
The exam guide calls out both proxies. Replacing “escalate complex cases” with a numeric threshold can preserve the original defect while making it look measurable. Start with the explicit criteria in 5.2.1. See 4.6.3 for confidence calibration in a human-review workflow. An untested number does not establish support authority.
How it works
Separate sentiment from the requested action
An angry customer can ask for a covered delivery-fee refund. A calm customer can ask for an uncovered competitor price match. If a sentiment threshold transfers the first and keeps the second autonomous, it makes both routing decisions from the wrong evidence.
Sentiment can still be useful for evaluating conversational quality. Anthropic’s customer-support guide treats sentiment maintenance as a business-impact measure: compare sentiment at the beginning and end of a conversation. It separately describes escalation efficiency as recognizing cases that need human intervention. The separation matters because improvement in tone does not demonstrate correct policy decisions.
Do not convert every negative message into a human request. Equally, do not let positive sentiment conceal an explicit request for a representative. “Thank you; I’d still prefer to speak with a person” is polite and unambiguous. The customer’s requested destination, not the sentiment label, controls the next action.
Sentiment may inform wording without deciding who owns the case. Acknowledge inconvenience when the customer is frustrated; then apply the same policy and progress criteria used for a neutral message. See 5.2.3 for the response sequence.
Treat self-reported confidence as a claim to test
Suppose the prompt asks Claude to output confidence: 0.92. In this example, confidence is an application-defined field. Its numeric form does not make it an observed probability that the support decision is correct. An uncalibrated score has not been checked against labeled outcomes for the decisions it will control.
The failure can occur in either direction. A high-confidence interpretation can overlook the absence of a competitor price-match rule. A low-confidence interpretation can accompany a routine refund whose eligibility is already established by tool evidence. A threshold then suppresses the uncertain covered case while allowing the confident unauthorized one.
Do not make the score override a categorical trigger. If the customer asks for a person, hand off even when the agent is certain it can resolve the issue. If policy is silent, confidence in a plausible interpretation supplies no authority to approve the request.
When a model expresses uncertainty, examine what is missing. An ambiguous customer identity requires a clarifying identifier; a missing policy rule requires an authorized decision; a failed operation requires the appropriate recovery or escalation path. “Low confidence” collapses these different causes into one destination and loses the action that could resolve them.
Expose deciding evidence in the tool contract
Define an application-owned escalate_to_human client tool through the Messages API’s tools parameter. Its description states when it should and should not be used. Its input_schema can require a concrete reason and a short evidence summary rather than an unexplained confidence number. The reason values are your support system’s contract, not Anthropic enums.
For example, reason: "policy_gap" should be accompanied by the requested exception and the relevant supplied rule. reason: "human_request" should identify the customer’s request. reason: "blocked_progress" should describe what remains unresolved after permitted recovery. The application checks that evidence before performing sensitive operations; see 1.5.3 for that enforcement boundary.
These fields do not make the model infallible. They make the decision inspectable against the conversation, policy, and actual tool results. Claude emits a tool_use request for the client tool; your application performs the operation. The execution result establishes whether the transfer happened, independently of how certain the assistant sounds.
Evaluate the two errors separately
A support evaluation should distinguish unnecessary transfers from missed required transfers. An overall completion count can hide either. In particular, reducing escalation volume can improve deflection while worsening unauthorized approvals or ignored human requests.
Use cases labeled from the explicit criteria, then compare actual action sequences with those labels. Include angry covered requests, calm uncovered requests, polite human requests, and high-confidence wrong policy interpretations. Change the wording while holding policy evidence constant to check whether mood accidentally controls the decision.
This is a routing check rather than a request for the model to grade itself. Anthropic’s support guide recommends tracking correctly escalated conversations and conversations that should have been escalated but were not. Use those observed outcomes to judge the routing change, rather than the average generated confidence score.
Key distinctions and trade-offs
The following signals answer different questions:
| Signal | What it describes | What it does not establish alone |
|---|---|---|
| Sentiment score | Customer mood or its change | Need for human authority |
| Self-reported confidence | Generated certainty assessment | Calibrated correctness probability |
| Policy evidence | Whether a rule covers the action | Whether the customer prefers a person |
| Explicit destination request | Customer preference | Difficulty of the underlying issue |
| Tool and recovery results | Available progress toward resolution | Authority to create an exception |
Explicit criteria take more care to author than a single threshold. They preserve distinctions the score discards. You can retain sentiment for conversation analysis without granting it routing authority, and retain diagnostic uncertainty without pretending it is a validated classifier.
Architecture / decision scenario
A retailer routes messages below a sentiment threshold to people and permits high-confidence model decisions to remain autonomous. Review reveals two cases: a furious customer waits for a person about a covered late-delivery refund, while a calm customer receives an unauthorized competitor price match.
Replace the routing rule with criteria in the top-level system prompt and reinforce them in the custom handoff tool’s description. The furious covered case receives acknowledgment and an eligible refund offer. The calm uncovered request produces reason: "policy_gap". Add a polite request for a representative; it transfers immediately even if the agent reports certainty that a refund would work.
Replay variations of all three cases and inspect tool calls. Keep sentiment measurements for customer-experience analysis. Compare required handoffs, missed handoffs, and unnecessary transfers against labeled outcomes. A falling escalation count alone cannot show that the repair succeeded.
Common exam traps
- Treating negative sentiment as a measure of case complexity.
- Treating a generated decimal as a calibrated probability.
- Allowing high confidence to override a policy gap or human request.
- Sending every uncertain identity lookup directly to a person.
- Measuring routing quality only through deflection or completion rates.
What to remember
Mood and uncalibrated confidence answer different questions from escalation criteria. Route from explicit preference, policy authority, and observed progress. Evaluate decisions against labeled outcomes, while using sentiment separately to assess the customer experience.
Knowledge check
A retailer routes messages below a sentiment threshold to people and keeps decisions with model-reported confidence above 0.9 autonomous. Review finds a furious customer waiting for a person about a covered refund, and a calm customer granted an uncovered competitor price match at confidence 0.95. Which change addresses both failures?
A team wants every call to its custom escalate_to_human tool to be checkable against the conversation, the policy and the tool results. Which input_schema design serves that goal?
confidence field stating how sure the agent is of the transferreason from the application’s trigger values plus a short evidence summarysentiment field recording the customer’s mood at the moment of transfernote in which the agent explains the transfer its own wayAfter a routing change, a support agent’s escalation volume falls by 30%. Which evidence best shows whether the change improved routing?
References
- Anthropic, Claude Certified Architect – Foundations Exam Guide
- Anthropic, Customer support agent
- Anthropic, Define tools
- Anthropic, Tool use with Claude
- Anthropic, Create a Message