"Can we let users request warranty repairs simply by sending a photo?" "Can we delegate Tier 1 call center support to AI?" Inquiries like these from consumer-facing sectors—such as manufacturing, retail, and insurance—have surged since April. The driver behind this trend is that models like Claude Sonnet 4.6, Gemini 2, and the GPT-5.4 series have matured in multimodal capabilities, processing text, images, audio, and video in a single pass.
Combined with standardized business API integrations spurred by MCP's transition to the Linux Foundation, automated customer support—executing photo diagnostics, repair assessments, CRM ticket creation, dispatch scheduling, and customer notifications end-to-end—has finally graduated from PoCs into production readiness. This article details practical design patterns for delivering these systems to clients.
Why conventional chatbots were never enough
Traditional customer support AI faced three fundamental limitations:
- Over-reliance on text: Forcing users to describe issues in text when sending a photo or voice memo would be vastly faster
- Isolation from core business systems: Capable of answering FAQs, but unable to access orders, shipments, or billing records
- Crude escalation mechanisms: A binary choice between answering "I don't know" or blindly dumping tickets onto human staff
Pairing multimodal AI with MCP resolves all three issues simultaneously. The cross-modal search between images and documents discussed in our article on multimodal embedding rerankers served as a foundational step toward this architecture.
Standard architecture for client development
Component architecture
| Layer | Role | Technology selection guidelines |
|---|---|---|
| Ingestion channels | LINE / web chat / phone (speech-to-text) | LINE Messaging API / Twilio |
| Multimodal understanding | Interprets images and audio to extract customer intent | Claude Sonnet 4.6 / Gemini 2 |
| MCP toolset | Core backend API invocations | Custom MCP servers + official connectors |
| HITL | Human approval for high-risk actions | Slack notifications + approval UI |
| Observability and analytics | Measuring conversation quality and resolution rates | Langfuse / Datadog |
Processing flow (warranty repair example)
[1] 顧客が LINE で故障した製品の写真 + 「動かなくなった」を送信
[2] マルチモーダル AI が
- 製品モデル特定（写真）
- 故障モード推定（写真+テキスト）
- 顧客特定（LINE ID → CRM 照合 / MCP）
[3] MCP 経由で
- 保証期間チェック
- 修理可否の判定
- 在庫照会
[4] HITL（金額 1 万円超 / 例外的判定 → オペレーター承認）
[5] 承認後、配送業者 API で集荷予約 → 顧客に予約番号通知Across these five steps, integrating step [4] HITL from the very start is the most critical factor for successful production deployment.
Cost estimates (10,000 monthly tickets)
From an engineer's perspective, here is an estimated token cost breakdown for handling 10,000 tickets per month.
| Item | Units | Monthly |
|---|---|---|
| Multimodal inference (1 photo + 5 conversation turns) | 8K tokens × 10,000 tickets | 80 million tokens |
| MCP invocations (average 3 tools / session) | 2K × 30,000 calls | 60 million tokens |
| HITL summary generation (triggered on 10% of tickets) | 1K × 1,000 tickets | 1 million tokens |
| Total | Approx. 140 million tokens |
Estimated with Claude Sonnet 4.6 (assuming $3/1M input, $15/1M output, and a 70% input ratio), total monthly costs come to around $1,500 (approx. 220,000 JPY). This is far lower than the labor cost of hiring one or two human operators.
However, because multimodal inference token usage expands significantly with the number of images, architecture designs must include constraints such as automatic image resizing and capping uploads at three images per session.
PII and personal data handling
Customer support channels are filled with sensitive personal data. At minimum, the following three safeguards must be designed into the system:
- Ingestion masking: Replacing phone numbers, addresses, and credit card details with temporary tokens before data reaches the AI
- Non-transmission field designations for AI vendors: Utilizing vendor opt-out features for PII that cannot be masked
- Encrypted audit log retention: Storing conversation logs using AES-256 encryption with strict access controls for seven years (varies by industry)
As highlighted in our article on Cloudflare AI Labyrinth, the boundary design for data exposed to AI is subjected to increasingly strict regulatory scrutiny every year.
Summary — sending a photo will become the customer touchpoint standard
While multimodal AI and MCP are frequently discussed in isolation, combining them within customer support creates immense synergy. An experience where "sending a single photo resolves everything" is no longer just an experimental showcase; it is rapidly becoming the industry baseline.
In this architecture, scope and delivery timelines depend heavily on three factors: how open existing CRM APIs are, what threshold triggers HITL approval, and at which layer PII is stripped. Even if you are simply looking to automate initial LINE responses or run a targeted pilot for photo diagnostics, sharing your existing channels and core backend systems will allow us to map out the best starting point. Please reach out via our inquiry form.
