クイック回答
first-response metricのeligible conversation、start/stop event、human/bot actor、office-hours、timezone、exclusion、cohort、aggregationをfreeze。AIでclassify、risk detect、incident duplicate、minimal context、skill/authority route、capacity balance、grounded draft、narrow resolve/lossless handoff、breach predict。同等cohortでquality/safetyと共に比較。
queue最適化前にmetric contract
clock定義
eligibility、start、first public human reply/stop event、calendar/business time、office-hours、timezone、reopen、bot handling、no-reply exclusionを記録。platform definitionは異なり変更もある。
cohort freeze
same date rule、channel、locale、product、priority、segment、support hour、incident、staffing、automation eligibilityのconversationを比較。no-reply ticketをsilent dropしない。
summary選択
eligible/replied/no-reply count、median、upper percentile/time bucket、SLA breach、必要なconfidence interval。meanはoutlier影響、different date anchorはincomparable population。
minimum metric record
metric_id · definition_version · eligible_population · start_event · stop_event · actor_type · bot_time · office_hours · time_zone · exclusion · date_anchor · cohort_fields · aggregation · no_reply_count · data_refresh · owner · change_log
avoidable waitingを減らす9 AI triage play
各playはspecific delay mechanismとcontrol/measurementを示します。可能ならindependent deployしcorrection/failureをattribute。
intakeでintent・language・product・channelをnormalize
first customer messageとpermitted metadataをcontrolled routing envelopeへ：primary/secondary intent、language、product/version、channel constraint、segment、confidence。original wording保持、unknown/multiple許可。
- required control
- current taxonomy/example、language detection、product/account evidence、confidence threshold、fallback queue、correction history。
- measure / guard
- arrival-to-classification、coverage、human correction、abstention、misrouteをmeasure。classificationをresponse扱いしない。
safety・security・privacy・incident signalを先にdetect
account takeover、credential exposure、payment harm、threat、regulated request、vulnerable user、widespread failure等をhigh-recall screen。approved specialistへrouteしpriority scoreでcritical gateを薄めない。
- required control
- signal definition/version、exact evidence、scope、false-positive route、specialist availability、containment、safe message。
- measure / guard
- signal別detection/miss、specialist acceptance、containment time、false-positive burdenをmeasure。toneだけでtrait/dangerをinferしない。
duplicateをgoverned incident laneへ
product、environment、error signature、time、customer-visible symptom、counterexampleでactive incidentと比較。probable duplicateをlinkしてもreview前close/root cause claim禁止。
- required control
- incident ID/status/version、similarity feature、counterexample、impact、confidence、unlink、mass-update approval。
- measure / guard
- duplicate precision、unlink rate、incident-lane FRT、reach、incorrect closureをmeasure。shared keyword≠verified incident。
first decisionに必要なcontextだけenrich
routing/first safe responseを変えるminimum permitted account、entitlement、plan、region、product state、event、incident、prior contextを取得。stale/unavailable/conflicting/restrictedをmark。
- required control
- field purpose/owner、system of record、retrieval time、freshness、permission、sensitivity、fallback、source pointer。
- measure / guard
- availability/latency、stale correction、privacy exception、reassignment/clarification削減をmeasure。more data≠better。
skillとauthorityを持つteamへroute
intent/riskをversioned routing matrixへ。knowledge、language、product、jurisdiction、permission、tier、action authority、queue hour、fallbackを含みreceiving team acceptance確認。
- required control
- rule ID/version、skill/permission、destination、queue state、tie-break、owner、acceptance、timeout/escalation。
- measure / guard
- first-touch accuracy、pre-response transfer、blind handoff、acceptance latency、correctionをmeasure。confidence/speedだけでrouteしない。
capacity/deadlineに対しqueue loadをbalance
eligible skill、active workload、schedule、concurrency、channel、predicted handling band、backlog、SLA risk、availabilityでassignment提案。変化時recompute、manual override/fairness review保持。
- required control
- capacity definition、workload source、schedule/timezone、limit、prediction/uncertainty、SLA clock、reassignment、override。
- measure / guard
- queue wait、FRT distribution、breach、reassignment、workload concentration、qualityをcomparable cohortでmeasure。一queueのため他をstarveしない。
agent open前にgrounded first-response draft
verified issue、decision-changing question、current approved knowledgeのsafe step、internal evidence、truthful checkpointを含むdraft。unsupported/stale/conflicting/permission-sensitive claimをagentへ表示。
- required control
- approved knowledge/version、customer words、verified context、tool state、policy、provenance、risk flag、language reviewer。
- measure / guard
- section acceptance、factual correction、stale citation、unsafe suggestion、edit time、quality、customer-visible FRTをmeasure。bad fast draftはnegative capacity。
narrow eligible requestをresolve、またはlossless handoff
stable/low-risk/well-evidenced intentでapproved automationがbounded answer/action。eligibility、identity、permission、confidence、confirmation、failure、rollback、human handoffを定義。conversation/action/unresolved need/preference保持。
- required control
- automation boundary/version、source evidence、action receipt、customer-visible result、abstention、handoff trigger、summary、acceptance。
- measure / guard
- automated resolutionとhuman FRTを分離報告。platformはbot reply/resolvedを除外する場合。reopen、correction、escalation、abandonment、outcomeをmeasure。
breach riskをpredictしtimer前にrecover
unanswered eligible conversationをcorrect SLA clock、office-hours、channel、priority、capacity、owner、dependency、confidenceと比較。current owner/escalationへact可能なlead timeでalertし結果記録。
- required control
- SLA rule/version/start、counted/excluded time、remaining band、owner、queue/dependency、threshold、acknowledgement、recovery action。
- measure / guard
- alert precision/recall、lead time、ack、recovered breach、fatigue、silent failure、post-response qualityをmeasure。predictionはempty reply許可ではない。
worked example：fastest queue≠best route
hypothetical exampleでOpenMax resultではありません。customerはurgent/users cannot sign in。General Supportがshortest queue。authentication languageだがcontextはrecent enterprise SSO change、account takeover/widespread outage evidenceなし。
- uncertainty保持。 multi-user impactはcustomer-stated、SSO change verified、outage not observed、security compromise unknownでfalseではない。
- speedでなくauthority route。 identity teamはSSO/tenant permissionあり。General Supportは不可、shortest queueはtransferを生む。
- bounded first response draft。 sign-in impact、exact error/IdP connection質問、outage claimなし、unsafe resetなし、real checkpoint。
- complete pathをmeasure。 classification latency、acceptance、human FRT、transfer、correctness、safe resolution、reopen、later incident linkageを記録。fast acknowledgementだけはsuccessでない。
speed・quality・safety・capacityを併測
responsiveness view
eligible start、human reply、no reply、median/tail FRT、bucket、breach、queue wait、acceptance、transfer。sample/definitionが比較可能時のみsegment。
quality/safety counter-metric
correctness/completeness、unsupported claim、unsafe step、privacy/permission、specialist miss、clarification、resolution、reopen、escalation、complaint、accessibility、satisfaction response。
evaluation design
staged rollout、predeclared cohort、holdout/comparable queue、risk-stratified review、human adjudication、follow-up。staffing、incident、seasonality、channel mix、release、policyをconfounder記録。
OpenMaxによるAI triage coordination
OpenMaxはpermitted intake、critical gate、incident link、minimal context、skill/authority rule、capacity-aware assignment、grounded draft、lossless handoff、SLA risk、human correction/outcomeをcoordinate。metric ownership、staffing、specialist judgment、consequential action、promise、fairness、override、acceptanceはhuman。
measurement・workforce・customer safety boundary
- clock、actor、office-hours、exclusion、date anchor、channel mix、automation eligibility、aggregation変更後にimprovement claim禁止。definition versionとseries reconciliation。
- opaque surveillance/automated employment consequence禁止。necessity、transparency、access、retention、contestability、calibration、bias review、human authorityを定義。
- emotion、accent、grammar、language、channelだけでseverity、vulnerability、fraud、trait、intent、valueをinferしない。defined evidence/specialist review。
- first-reply chartのためcorrect resolution、privacy、security、accessibility、consent、honest expectation、safe escalationを犠牲にしない。
情報源、編集方法、制限
OpenMax編集部はZendesk first-reply definition/metric/SLA、Intercom responsiveness/bot-time/office-hours/reporting/SLA、NIST AI RMF role/oversight/measurement/monitoringを確認し、9 playとhypothetical SSO caseを作成。2026年9月3日再確認。
- Zendesk — Native Support time duration metrics
- Zendesk — Support metrics and attributes
- Intercom — Responsiveness reporting
- Intercom — Set SLAs for conversations and tickets
- NIST — AI RMF Core
よくある質問
auto acknowledgementはfirst response？
platform/report次第。timer停止設定でもeventをhonest labelしmeaningful human response、bot resolution、acknowledgementを分離報告。
averageかmedian FRT？
distributionを使用。medianはmeanよりoutlierに強いがtailは別。eligible/no-reply count、bucket/upper percentile、breachを含める。
AIはfastest available agentを選ぶ？
speedはone input。skill、language、product、permission、jurisdiction、action authority、capacity、accepted ownershipも必要。
bot-resolved conversationのreport？
separate eligible populationとしてresolution/handoff定義。human denominatorから除外しただけでhuman operation fasterとclaimしない。
OpenMaxで何をautomate？
classification、risk gate、incident proposal、context、routing、capacity、draft、handoff、SLA alert、evidence、review、correction、measurementをcoordinateしconsequential authorityはhuman。