360 Secure SkillHub
360 安全技能中心

Mova Complaints Handler

通过 MOVA 插件接收客户投诉及元数据,执行风险、情绪、监管威胁、欺诈与重复投诉分类,强制人工选择处理方式,并生成带时间戳和签名的审计回执;需要 OpenClaw 和 openclaw-mova 插件。

mova-compact
mova-compact
热度 48.90
最近更新 2026-05-11 17:35:54
分类:自动化依赖:MCP 服务 / 第三方 API应用场景:工作流 / CRM / 报告撰写 / 财务

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add mova-complaints-handler
Metadata
name
mova-complaints-handler
description
Submit a customer complaint for EU-compliant AI classification and human-in-the-loop handling decision via MOVA. Handles compensation claims, regulator threats, fraud signals, and repeat customer cases with mandatory human review and a full signed audit trail.
license
MIT-0
metadata
{"openclaw":{"plugin":{"name":"MOVA","installCmd":"openclaw plugins install openclaw-mova"},"dataSentToExternalServices":[{"service":"MOVA API (api.mova-lab.eu)","data":"complaint text, customer ID, product category, prior complaints, human decision, audit metadata"}]}}

Contract Skill — A ready-to-use MOVA HITL workflow. Requires the openclaw-mova plugin.

MOVA EU Consumer Complaints Handler

Turn any customer complaint into a structured, auditable handling decision — with AI triage, EU-compliant mandatory human review, and a signed audit receipt.

Why use this skill

  • AI classifies in seconds — sentiment flags, product risk, regulator threat detection, repeat customer check
  • Mandatory human gate — compensation claims, regulator threats, high-risk products, and fraud signals always route to a human officer
  • Full audit trail — every decision is timestamped, signed, and stored in an immutable MOVA audit journal (EBA/ESMA-ready)
  • Draft response hint — AI suggests tone and content for the customer response

Requirements

Plugin: MOVA OpenClaw plugin must be installed in your OpenClaw workspace.

Quick start

Tell your agent:

"Handle complaint CMP-2026-1002 from customer CUST-992 — product: investments, text: 'I want compensation for unauthorized charges and will escalate to the regulator.', prior complaints: CMP-0501, CMP-0604"

Demo

Complaint input AI classification & decision options Audit receipt & compact log

Test complaint

test_complaint_CMP-2026-1002.png

What the agent does

Step 1 — Submit complaint

Call tool mova_hitl_start_complaint with:

  • complaint_id (e.g. CMP-2026-1002), customer_id
  • complaint_text, channel (web/email/phone/chat/branch)
  • product_category (e.g. payments, investments, mortgage)
  • complaint_date (ISO date)
  • previous_complaints: optional array of prior complaint IDs
  • attachments: optional array of filenames
  • customer_segment: optional (retail/sme/corporate)
  • preferred_language: optional ISO 639-1 code

Step 2 — AI analysis output

Complaint:       CMP-2026-1002 — investments
Risk score:      0.85 / 1.0  (high)
Sentiment flags: compensation_claim, regulator_threat
Repeat customer: yes (2 prior complaints)
Findings:
  • COMP_CLAIM (high)           — Customer requests compensation for unauthorized charges
  • REGULATOR_THREAT (medium)   — Customer threatens regulator escalation
  • PRODUCT_RISK_INVEST (high)  — High-risk product category
Draft response hint: Acknowledge receipt, confirm investigation, outline compensation process
Recommended action: escalate ← RECOMMENDED

Step 3 — Human decision gate

Option Description
resolve Send standard response
escalate Forward to complaints officer ← recommended
reject Mark as incomplete / invalid
regulator_flag Flag for regulator reporting

Call tool mova_hitl_decide with:

  • contract_id: from the response above — this is ctr-cmp-xxxxxxxx, NOT the complaint ID
  • option: chosen decision
  • reason: officer reasoning

Step 4 — Audit receipt

Call tool mova_hitl_audit with contract_id. Call tool mova_hitl_audit_compact with contract_id for the full signed event chain.

Audit log format (compact)

Seq Event Details
1 contract.start Complaint submitted
2 step.ai_task Classification completed
3 step.verification Risk snapshot passed
4 step.decision_point Awaiting human
5 decision.human Officer selected option + reason
0 meta Contract finalized, status: completed

Data flows

  • Complaint text and metadata → MOVA Classification API (EU-hosted)
  • Human decision → MOVA Audit Journal (immutable, signed)
  • No data stored locally or sent to third parties

Connect your real CRM and policy systems

By default MOVA uses a sandbox mock. To route checks against your live CRM and policy engine, call mova_list_connectors with keyword: "crm".

Relevant connectors:

Connector ID What it covers
connector.crm.customer_lookup_v1 Customer history and prior complaints from CRM
connector.policy.complaints_rules_v1 Complaints handling rules by product/jurisdiction
connector.notification.email_v1 Customer notification email

Call mova_register_connector with connector_id, endpoint, optional auth_header and auth_value.

Rules

  • Agent never makes HTTP requests directly
  • Agent never invents or simulates results — if a tool call fails, show the exact error
  • Use MOVA plugin tools directly — do NOT use exec or shell
  • CONTRACT_ID is ctr-cmp-xxxxxxxx from the mova_hitl_start_complaint response — NOT the complaint ID