5つの不正シグナルで請求詐欺を検出します。損失の最初の通知時に1つのワークフロー。

You are integrating Didit into an insurance carrier's claim workflow at first notice of loss (motor, health, property, travel, business interruption). Goal: catch ghost claims, deepfake video evidence, forged supporting documents, sanctions hits on the claimant, and organised-injury rings before the payout enters the queue. One API call. One signed webhook. Five fraud signals.

WHY THIS SHAPE
  - The Coalition Against Insurance Fraud estimates fraud costs the US industry around $308 billion / year (2022 figures). Most of that walks in through claims with no identity check on the claimant and no document forensics on the evidence.
  - Five signals settle the question on most claims: (1) the claimant is the policyholder, (2) the claimant is alive and present (not a deepfake), (3) the supporting documents are not tampered or templated, (4) the claimant is not on a sanctions list, (5) the claimant has not already filed N near-identical claims across your portfolio.
  - One Didit /v3/session/ call bundles all five. $0.33 KYC + $0.20 AML = $0.53 per claim. Face Search 1:N is free per search. 500 verifications free every month.

PRE-REQUISITES
  - Production API key from https://business.didit.me (sandbox key in 60 seconds, no credit card).
  - A webhook endpoint with HMAC SHA-256 verification of the X-Signature-V2 header using your webhook secret.
  - A Workflow Builder workflow bundling ID Verification + Passive Liveness + Face Match 1:1 (with the policyholder portrait as comparison target) + AML Screening + Face Search 1:N (gallery scoped to prior claimants on your account).
  - Reference to the policy ID and policyholder portrait on file from your Policy Administration System (PAS) — passed as metadata on the session.

STEP 1 — Open the claim-verification session
  POST https://verification.didit.me/v3/session/
  Headers:
    x-api-key: <your api key>
    Content-Type: application/json
  Body:
    {
      "workflow_id": "<wf id bundling ID + Liveness + Face Match + AML + Face Search 1:N>",
      "vendor_data": "<your claim id, max 256 chars>",
      "callback": "https://<your-app>/claims/verify/callback",
      "metadata": {
        "policy_id": "<your policy id>",
        "line_of_business": "motor",
        "loss_date": "2026-04-12"
      }
    }

  Response: 201 Created with a hosted session URL. Send it to the claimant by email / Short Message Service (SMS) / inside the claims app. The claim stays in HOLD on your side until the signed webhook lands.

STEP 2 — Read the signed webhook
  Didit POSTs the verdict. Verify X-Signature-V2 (HMAC SHA-256 of the raw body) BEFORE reading the JSON.

  Payload (excerpted):
    {
      "session_id": "<uuid>",
      "vendor_data": "<your claim id>",
      "status": "In Review",
      "id_verification": { "status": "Approved" },
      "liveness": { "status": "Approved" },
      "face":     { "status": "Approved", "similarity_score": 0.92 },
      "aml":      { "status": "In Review", "hits": [{ "list": "PEP" }] },
      "face_search": {
        "matches": [
          { "session_id": "claim-7710", "similarity": 0.94, "vendor_data": "claim-7710" }
        ]
      }
    }

  Session status enum (exact case, Title Case With Spaces): Approved | Declined | In Review | Resubmitted | Expired | Not Finished | Kyc Expired | Abandoned.

STEP 3 — Branch on the verdict
  Approved      → release the claim into the standard payout queue.
  In Review     → route to the Special Investigations Unit (SIU) with the per-module signals + face-search match list as the case file.
  Declined      → decline + open file. Block the payout. The decision payload is the audit pack.
  Not Finished  → resend the session link.

STEP 4 — Document forensics on supporting evidence (separate sub-flow)
  For each supporting PDF / image uploaded (police report, repair quote, medical invoice, photo of loss), run Didit Document Optical Character Recognition (OCR). The OCR response surfaces:
    - Field-level extracted text (claim amount, names, dates)
    - Tamper score per region (pixel-level edits)
    - Template match against prior submissions (reused PDF skeleton)
    - EXIF / metadata mismatch (date in photo vs date of loss)

  A high tamper score or a template match against a prior claim flips the case status to In Review.

STEP 5 — Pull the full decision for the case file
  GET https://verification.didit.me/v3/session/{session_id}/decision/
  Headers:
    x-api-key: <your api key>

  Returns the full decision payload — per-module signals, raw face-similarity scores, AML hit list with source list per match, face-search candidate list with scores. Use this as the audit pack for any dispute.

WEBHOOK EVENT NAMES
  - Sessions: standard session webhook. One endpoint, status field tells you the lifecycle.
  - Verify X-Signature-V2 (HMAC SHA-256) on every payload.

CONSTRAINTS
  - Session statuses use Title Case With Spaces. Never UPPER_SNAKE_CASE — that's the Transactions API.
  - The Face Match comparison target is the policyholder portrait from your Policy Administration System (PAS). A deepfake of the policyholder cannot pass when Passive Liveness is also in the workflow.
  - Face Search 1:N gallery is scoped to YOUR account — Didit does not search across carriers. To collaborate across an industry pool, use a shared workflow_id pointing at a multi-carrier gallery you own.
  - 200+ fraud signals are surfaced on every session at no extra cost — read them off the decision payload, don't re-query.
  - Default retention is the standard 5-year insurance horizon; adjust per workflow if your jurisdiction differs.

Read the docs:
  - https://docs.didit.me/sessions-api/create-session
  - https://docs.didit.me/core-technology/face-match/overview
  - https://docs.didit.me/core-technology/aml-screening/overview
  - https://docs.didit.me/core-technology/face-search/overview
  - https://docs.didit.me/integration/webhooks

Start free at https://business.didit.me — sandbox key in 60 seconds, 500 verifications free every month, no credit card.