Live in production on our own 3 services — the chat at the bottom right of this site is the real thing

Refunds, invoices and delivery inquiries,
completed in the conversation.

Skix CS is a customer support AI agent platform that does more than answer inquiries: it carries out identity verification, approval, execution and recording and completes the task. Approval rules are enforced in code, not prompts, so a conversation can never break them.

See how it works

With business APIs, onboarding takes 0.5–1 day. Without APIs, you can start with FAQ answers and human handoff.

Diagram of a refund conversation and its audit log A customer's refund request passes identity verification and a human approval triggered by the amount threshold, then completes, while each decision event is recorded in the append-only audit log on the right. Skix CS agent Active Please refund order #48213 For identity verification, please share your phone's last 4 digits. 7031 ✓ Identity OK — deterministic match ¥18,400 ≥ approval threshold → human approval ✓ Approved by an operator Your refund is being processed. We'll notify you as soon as it's complete. Idempotent task execution on a workflow engine Audit events append-only 12:04:01 turn 12:04:12 intent 12:04:15 state_transition 12:04:15 policy_decision 12:05:02 approval 12:05:04 tool_call 12:05:05 gateway_hop
Skix CS promotional image: three steps that automate refunds, invoices and delivery inquiries from conversation to completed task, passing identity verification and the audit log
Received in the conversation → passes the control gates → completes the task. The LLM is a component that assists decisions; it holds no authority to execute.
3 servicesLive in production on
Skix's own services
3 languagesConversation language auto-follows
Japanese, Chinese and English
Full auditEvery conversation and decision
stored as append-only events
Refusal by designBelow the evidence threshold,
it refuses honestly instead of generating
Customers

We use it ourselves first.
The chat on this site is Skix CS, too.

Skix's three services run in production on Skix CS, each as an independent tenant. The Web widget at the bottom right is one of them — not a diagram or a demo video, but the live production environment you can touch directly.

Skix CS (this site)

Answers product questions and takes demo and quote requests. Answers are returned with citations, grounded in this user guide; without evidence it refuses and hands off to a staff member.

Web widgetKnowledge baseHuman handoff
Skix Pulse

Inquiry desk for service users and visitors. Serves Japanese, Chinese and English users, automatically matching the conversation language.

Web widgetMultilingualKnowledge base
Skix One

Support desk for an enterprise service. The knowledge base gives the first answer; inquiries with no evidence or that need a judgment call go to the handoff queue in the Operator workbench.

Web widgetHuman handoffAudit log
You can start with a PoC on your own refund flow.
Features

The reasons you can trust it
are built into the structure.

We make the automation of refunds, invoices and delivery inquiries something you can delegate — not by "trusting the AI", but because controls, identity verification and auditing work as structure.

Approval rules run as code

Approval thresholds and refund limits are implemented in code, not prompts. Amounts above the threshold go to human approval; anything over the limit is rejected immediately. Clever conversations and prompt injection cannot move them.

→ "The AI issued a refund on its own" never happens

Identity gate

Deterministic matching of the order number and the last 4 digits of the phone number. Default-deny design: an unverified conversation never touches customer data. Even an anonymous visitor on LINE or the web can proceed as the verified customer once the match passes.

→ Prevents impersonation-based data leaks by structure

Full audit and replay

Every conversation and every decision is recorded as append-only events, and any session can be replayed exactly as it happened. "When, what and why" can be explained after the fact. A GDPR erasure API is also provided.

→ Answer customer disputes and audits with the log

Refusal-by-design knowledge base (RAG)

Below the evidence threshold it does not generate — it honestly says "I don't know" and connects you to a person. Answers carry source citations, and your existing FAQ can be bulk-imported.

→ Never lose trust to a plausible wrong answer

Human handoff and Operator workbench

When in doubt it hands off to a person, and the AI stays silent while a human is handling the conversation. Approval queue, handoff queue, knowledge drafts and quality review in one workbench. Neglected approvals get reminders; expired ones are cancelled automatically and the customer is notified.

→ Keep your existing support team as it is

Multilingual and long-term memory

Automatically follows Japanese, Chinese and English, and keeps up when the customer switches. Long-term memory behind a PII gate enables responses informed by each customer's context.

→ The same quality for inbound and cross-border customers
How it works

Every request passes
through the same controls.

Whichever channel it comes from, a request passes identity verification, policy decision and audit recording before it reaches the tenant's business APIs. The LLM is "a component that assists decisions"; it has no authority to execute tasks.

Skix CS architecture diagram A three-tier structure: channels (REST API, Web widget, LINE) pass through the controller and policy gateway, then connect to the tenant's business APIs via connectors. CHANNEL Channels REST API Web widget LINE Auto-follows Japanese, Chinese and English All requests POLICY GATEWAY Controller Identity gate (default-deny) Amount threshold → human approval / reject Append-only audit events Workflow engine — idempotent execution Verified only TENANT Tenant business APIs Connected via 5 connectors Built-in circuit breaker Orders, refunds, shipping, invoices, customers Isolated by DB row-level security Per-tenant policies and budget limits applied as code
Policy is code. Because rules are not written in prompts, a conversation cannot break them. Approval, rejection and recording run as structure.
Onboarding

With business APIs, live in
0.5–1 day.

No large project required. Four steps take you through acceptance testing. If you don't have business APIs yet, skip step 2 and start with FAQ answers and human handoff.

  1. Create a tenant

    Create a dedicated tenant on the multi-tenant platform. Set approval thresholds, refund limits and budget limits per tenant.

  2. Connect connectors

    Connect your business APIs (orders, refunds, shipping, invoices, customers) through five connectors. JSON plus a single auth header is all it takes, with a circuit breaker built in.

  3. Load your FAQ

    Bulk-import your existing FAQ into the knowledge base (with a dry run). Answers carry citations and are refused when evidence is insufficient.

  4. Acceptance test

    Verify identity verification, the approval flow and handoff with real data, validate the audit log by replay, and go live.

0.5–1 day in total — the standard onboarding time when business APIs are available.

Start without APIs — begin with FAQ answers + human handoff, connect business APIs later.

FAQ

Frequently asked questions

Can we adopt it without business APIs?

Yes. You can start with knowledge-base answers from your existing FAQ and human handoff alone (this site itself runs that configuration). Automatic execution of refunds, invoices and delivery inquiries becomes active once your business APIs are connected through connectors.

What is the pricing?

An individual quote based on monthly conversation volume, connected channels and the number of connected systems. We welcome inquiries starting from a small PoC. Tell the chat at the bottom right "I'd like a demo", or contact [email protected].

How do you prevent wrong answers and prompt injection?

Two layers of structure protect you. First, the knowledge base (RAG) is refusal by design — below the evidence threshold it does not generate but honestly refuses, and answers carry citations. Second, policies such as approval thresholds and limits are implemented in code, not prompts, so prompt injection cannot move them. Amounts above the threshold go to human approval; anything over the limit is rejected immediately. Uncertain cases are handed off to an operator.

How do you connect to our existing systems?

We connect to your business APIs (orders, refunds, shipping, invoices, customers) through five connectors, with a circuit breaker built in. Tasks such as refunds, invoices and delivery inquiries run idempotently on a fault-tolerant workflow engine, so a mid-way failure never causes double execution. On the customer side we support REST API, Web widget and LINE.

How does it work alongside our existing human support team?

Operators handle the approval queue and the handoff queue in the workbench. A person replies in conversations the AI has handed off, and can return them to the AI once resolved. While a human is handling a conversation, the AI does not speak. Answers born from human support can be promoted to the knowledge base as FAQ drafts.

How is our data isolated from other companies?

On the multi-tenant platform, tenants are isolated by database row-level security (RLS). Policies and budget limits are also set per tenant. In addition, each conversation passes an identity gate: unless the deterministic match of the order number and the last 4 digits of the phone number passes, no customer data is accessed at all (default-deny).

Can you handle data erasure requests (GDPR etc.)?

Yes. A GDPR erasure API is provided. Every conversation and every decision is also recorded as append-only events, and any session can be fully replayed, so "when, what and why" can be verified after the fact. Personalization through long-term memory is also behind a PII gate.

Which channels and languages are supported?

Three channels: REST API, Web widget and LINE. The Web widget is just an iframe to embed; LINE only needs the Webhook setting on your official account. Languages: Japanese, Chinese and English, automatically following the customer's language.

Start by trying it
on your own refund flow.

Tell the chat at the bottom right "I'd like a demo" — that itself is a demo of Skix CS. With business APIs, it's live in 0.5–1 day.

Contact us by email

Existing customers can log in at the Tenant console.