Як Travelers автоматизує подання claims через OpenAI Realtime

Після автомобільної аварії customer хоче швидко повідомити insurer, зрозуміти базовий process і отримати claim ID. Call center у звичайний день має predictable load, але catastrophe може створити масовий spike. Voice agent добре підходить для structured first notice of loss, бо збирає повторюваний набір фактів природною мовою й доступний 24/7.

Найважливіша межа — filing не дорівнює insurance decision. Agent може зібрати who/what/when/where, дати approved policy information і створити claim record. Він не повинен самостійно вирішувати coverage dispute, liability, fraud, reserve, settlement або payment лише тому, що вже має контекст розмови.

Trigger — inbound claim call або approved digital handoff. Input — authenticated policy/customer context, incident narrative, vehicle/property details, location/time, involved parties, injuries/safety signals, approved policy information і state/jurisdiction rules. AI stage виконує speech understanding, clarification, slot filling, policy-grounded explanation, consistency checks і structured FNOL creation.

Integrations — telephony, OpenAI Realtime API, policy lookup, claims orchestration, identity, notifications і digital follow-on systems. Output — confirmed claim submission, claim/reference ID, next-step notification, digital link для photos/appraisal/repairs/rental або escalation до live specialist. Claim system, а не transcript, є source of truth.

• Trigger → customer claim call.
• Input → identity/policy + incident facts + approved coverage/process sources.
• AI → understand → clarify → structure → validate → explain.
• Integrations → claims infrastructure, policy systems, notifications, digital follow-on.
• Output → submitted FNOL або live-specialist escalation; adjudication authority окремо.

1) Disclose AI and offer live-specialist option. 2) Verify identity/policy using approved factors. 3) Determine whether case is eligible for AI FNOL lane. 4) Capture incident facts in structured schema. 5) Ask targeted clarifications and detect contradictions/missing mandatory fields. 6) Answer only approved policy/process questions from current authoritative source. 7) Summarize material facts back to customer. 8) Obtain confirmation. 9) Submit claim with idempotency key. 10) Reconcile claim ID/status. 11) Send next steps or hand off.

Any injury/emergency signal, unsupported jurisdiction, policy mismatch, severe ambiguity, fraud-sensitive issue or explicit request for human help exits the autonomous lane. Good agentic design is not the art of keeping everyone inside AI; it is the art of knowing exactly when automation should stop.

A4 is appropriate for bounded FNOL submission because the assistant can complete a multi-step external action. The autonomy boundary ends at the claims-intake contract. Live specialist remains available at any point according to Travelers, while claim professionals focus on more complex cases. Human escalation is therefore a first-class route, not an error page.

Review should be risk-triggered, not random bureaucracy. High-impact flags route immediately; routine auto-damage filing can proceed if identity, eligibility, required fields, confirmation and system health all pass. If downstream claim professional later corrects a material capture error, the incident becomes labeled eval evidence and can lower automation threshold for that slice.

Insurance language is sensitive to policy version, endorsements, state and effective date. Agent must retrieve approved current source by policy context and answer only within documented scope. If question requires interpretation beyond approved explanation, system escalates instead of inventing legal certainty.

Every material answer trace stores policy/version/source ID and timestamp. Model-generated wording can be natural, but mandatory disclosures and prohibited claims are deterministic templates/rules. A fluent real-time voice is interface quality; coverage truth still comes from authoritative contract/rules.

Claims calls happen under stress, with road noise, accents, interruptions and incomplete memories. Critical fields — date, vehicle, location, injuries, contact details — need repeat-back or visual confirmation where appropriate. Low ASR confidence or contradictory answers should trigger clarification, not silent normalization.

The assistant also needs emergency and vulnerability handling. If customer indicates immediate danger or medical emergency, claim intake is no longer the primary task. Safety script and escalation path must be deterministic. Voice sentiment may help interaction design, but should not become hidden coverage or fraud scoring without a separate governed decision process.

Failure modes: policy lookup outage, wrong identity match, audio dropout, model interruption, mandatory field missing, claim-service timeout, notification failure, duplicate calls, customer correction after submit і catastrophe-scale dependency degradation. Submission states should include `NOT_STARTED`, `READY`, `SUBMITTING`, `CONFIRMED`, `FAILED`, `UNKNOWN_REQUIRES_RECONCILIATION`.

Claim creation uses idempotency key linked to authenticated policy/customer and incident intent. If tool times out after submit, assistant does not create a second claim. It queries claims system by external reference, reconciles authoritative state and either returns existing claim ID or escalates. Notification can retry separately; it must not repeat claim creation.

OpenAI notes that catastrophe events can generate more than 100,000 claims in days. For reproduction, capacity model includes telephony concurrency, Realtime sessions, policy/claims API limits, queues, notification systems, human fallback capacity and regional dependency failure. A green model endpoint is meaningless if claim system is saturated.

Degraded mode should be explicit: queue callback, collect non-consequential preliminary details, route live support selectively or ask customer to use digital path, depending on approved business rules. Never tell customer claim is filed until authoritative claim ID/status is confirmed. Runtime truth needs end-to-end postcondition, not just 200 OK from model API.

Eval slices: simple auto damage, multiple vehicles, ambiguous timeline, policy question, unsupported line, state variation, injury mention, live-agent request, noisy audio, multilingual/accent variation, contradiction, prompt injection via quoted text, policy-source outage, claim-submit timeout і catastrophe load. Multiple trials matter because voice/agent trajectories vary.

Metrics: required-field completeness, material-fact accuracy, policy-grounding correctness, unsupported-answer rate, escalation precision/recall, customer correction rate, claim-submission success, duplicate-claim rate, unknown-state reconciliation, p95 latency, live-agent transfer success і cost per verified filed claim. Reported 85–90% completion is a reference metric, not proof of all these dimensions.

Claims data can contain personal, financial, vehicle, location and injury information. Context minimization happens before model call; tools expose only fields needed for the current step. Raw audio/transcripts, policy documents, traces and model outputs have distinct retention/access policies. Training-use policy and retention are treated as separate controls.

Audit envelope records actor/session, policy reference, model/revision, prompt/config version, tool calls, confirmations, escalation, claim external reference and terminal state. Secrets/credentials never appear in model context. External/quoted content cannot expand tool scopes. Any high-severity leakage or unauthorized action triggers containment and permanent regression test.

Full cost includes Realtime audio inference, telephony, policy retrieval, claims orchestration, identity verification, monitoring, evaluation, human fallback, training/upskilling, fraud/safety controls, incident response and support. Catastrophe peak matters more than average call volume; capacity reserve and graceful degradation belong in economics.

Business KPI should separate automation from customer/claims quality: verified completion, transfer rate, recontact, corrections, duplicate submissions, claim-professional rework, cost per verified FNOL and severe-error rate. Public completion metric does not prove lower claim severity, faster settlement, cost savings or better adjudication, тому такі claims не приписуються кейсу без окремих даних.

Етап 1 — shadow transcription/structured capture поруч із live agent. Етап 2 — AI conversation + human submit. Етап 3 — bounded claim submission у кількох jurisdictions. Етап 4 — countrywide/large-scale rollout після load tests, reconciliation drills і safety gates. Етап 5 — broader claim interactions лише як окремі risk contracts, а не автоматичне розширення permissions.

Патерн підходить insurers, roadside assistance, warranty і service-intake operations з повторюваним structured first notice. 80/20 — одна claim type, current policy source, human escape і verified submit. Спроба одразу автоматизувати coverage, liability, payment і dispute resolution перетворює хороший intake agent на regulatory experiment, який ніхто не просив.