AI Business Platform — Product Guide

AI Business Platform
│
├── Platform Console
│   ├── Dashboard
│   ├── Companies
│   ├── Monitoring
│   ├── Logs
│   ├── Support
│   └── Settings
│
├── Company Workspace
│   ├── Dashboard
│   ├── Knowledge
│   ├── Conversations
│   ├── Employee AI
│   ├── Analytics
│   └── Settings
│
├── Core AI Engine
│   ├── Knowledge Engine
│   ├── AI Router
│   ├── Human Handover
│   ├── Cost Optimization
│   └── Token Usage Policy
│
└── Channel & Control Layer
    ├── WhatsApp
    ├── Notifications
    ├── Security
    ├── Customer-Owned API
    └── Future Integrations

How To Read It

Build a reusable Multi-Tenant AI Business Platform that companies can deploy as their own internal AI system.

打造一个可复制的多租户 AI 商业平台，让不同公司都能部署属于自己的内部 AI 系统。

The platform is designed to solve two major business problems:

1. Employees constantly asking senior staff repetitive questions.
2. Customers requiring fast and accurate responses through WhatsApp.

Core Objectives

目标不是做聊天机器人，而是做企业真正会长期使用、可控、可商业化的 AI 工作平台。

Two Main Platforms

整体结构先分成两个平台层：一个给平台运营方管理使用，一个给客户公司日常使用。

AI Business Platform
│
├── Platform Console
│   └── Used by platform operator
│
└── Company Workspace
    └── Used by customer company staff

Platform Console

Used by the platform operator. This platform is never exposed to customers.

这是平台运营方使用的管理后台，负责客户开通、监控、日志、支援和系统管理，不直接开放给客户。

Responsibilities: Customer Management, Workspace Creation, Deployment, Monitoring, System Health, License Management, AI Monitoring, Logs, Backup, Updates, Support.

Used by customer company staff. V1 should keep this workspace simple and easy to understand. No complicated permission or department system in V1.

这是客户员工每天会真正使用的工作区。第一版强调简单、清楚、容易上手，不先做太复杂的权限与部门层级。

Company Workspace
│
├── Dashboard
│
├── Knowledge
│   ├── Documents
│   ├── FAQ
│   ├── SOP
│   ├── Product Information
│   └── AI Suggestions

├── Conversations
│   ├── Needs Attention
│   ├── Active
│   ├── Completed
│   └── All Conversations

├── Employee AI
├── Analytics
└── Settings

Conversations
├── Needs Attention
├── Active
├── Completed
└── All Conversations

Employee AI
└── Ask a Question

Analytics
└── Simple Results

Settings
├── WhatsApp
├── Staff Notifications
├── AI
├── Usage and Budget
└── Company

1. Knowledge

Knowledge is the only customer-facing name for a company's approved information. Internally, a Knowledge Engine prepares, indexes, retrieves, and versions it.

客户只需要看到 Knowledge。解析、整理、搜索、版本和发布都由后台自动完成。

Only Published Knowledge is available to AI. Old conversations can create suggestions, but never become official Knowledge without human approval.

• PDF
• Word
• Excel
• SOP
• FAQ
• Product Manual
• Internal Documentation
• HR Policy
• Company Policy
• Website (Future)
• Approved Conversation Insights

Knowledge is versioned, reviewed, published, and audited. AI must not guess or invent company information.

2. AI Employee Assistant

Target User: Employees

UI Name: Employee AI

Employees should ask AI instead of asking senior staff.

重点是把 senior 脑中的重复知识，变成新人和员工可以随时向 AI 查询的内部能力。

• How do I explain Product A?
• Where is this SOP?
• What is the leave application process?
• How do I reply to this customer?
• What is our warranty policy?
• How does this workflow work?

If information cannot be found: Inform the employee, suggest the relevant department, and in the future notify the responsible person.

3. AI Customer Assistant

Target User: Customers

UI Location: Conversations

Primary communication channel: WhatsApp

There is only ONE AI Customer Assistant. Do not create separate Sales AI and After Sales AI in V1.

对客户来说只会看到一个统一的 AI 助手，不会感觉自己被转来转去；销售、售后等逻辑在系统内部处理。

• Sales Enquiry
• Product Information
• After Sales Support
• Warranty
• Appointment
• Complaint
• General Enquiry

This is a V1 core feature. Customers should never be asked to contact staff manually.

这是 V1 核心功能。客户不需要自己找员工，系统会自动暂停 AI、通知员工并保留原本的 WhatsApp 对话。

Internal workflow:

Customer
↓
AI detects escalation
↓
Create Case
↓
Notify responsible staff
↓
Staff accepts case
↓
Staff continues the original conversation
↓
Customer remains inside WhatsApp
↓
Case is resolved and closed
↓
AI resumes only when allowed

Notification Channels

The conversation never leaves WhatsApp.

Do NOT create a standalone AI Learning Center in V1. Learning from old conversations should be part of the Knowledge module.

第一版不另外做独立的 Learning Center，而是把历史对话学习、FAQ 建议、话术整理这些能力都收进 Knowledge 模块里。

Import Existing WhatsApp Chat
↓
AI Background Analysis
↓
AI Generates Suggestions
↓
Notify Responsible Person
↓
Review Suggestions
↓
Approve / Edit / Reject
↓
Approved Content Updates Knowledge

AI Can

• Analyze conversations
• Detect common questions
• Generate FAQ suggestions
• Extract reply style
• Identify sales flow
• Identify after-sales flow
• Suggest new knowledge
• Suggest SOP improvements

AI Cannot

• Directly update official knowledge
• Directly change official reply style
• Learn from conversations without approval

也就是说，AI 可以协助整理和建议，但不能自己修改正式知识库，最终决定权仍然在人手上。

AI Principles

这一段定义的是平台底线：宁可回避，也不要乱答。

• Answer using approved, published Knowledge.
• Never hallucinate.
• Never guess.
• Never invent prices.
• Never invent policies.
• Never expose confidential information.

If verified information does not exist: Escalate.

Knowledge Strategy

New information becomes available to AI only after review, processing, and publishing complete. Customers do not manage these technical steps.

资料必须经过审核和后台处理后才会对 AI 生效，客户不需要理解或手动操作技术步骤。

Approved Knowledge
↓
Knowledge Engine
↓
AI Router
↓
Employee AI
Customer AI
Future AI Modules

所有 AI 功能共用同一套已批准 Knowledge；后台自动完成处理，客户界面保持简单。

AI Router

One of the system's core modules.

后台会先检查安全、当前是否有人接手、规则、资料是否足够和风险，再决定回复方式。

• Security and prompt-injection check.
• Active Human Case check.
• Hard rules, intent, and risk classification.
• Knowledge coverage and retrieval confidence check.
• Select reply method or Human Handover.

Safe and supported → AI reply
Unclear or risky → Human Handover
Unsafe request → Blocked response

Company Admin controls only AI On / Paused, Monthly Budget, and Warning Level. Model selection, risk checks, handover timing, and token protection stay in the background.

Possible Strategies

成本控制不是事后报表，而是要在系统流程里先天就考虑进去。

Token Usage Policy

The system must be designed to minimize unnecessary AI token usage.

这部分的重点是：不是每一句都丢给大模型，而是先做规则判断、摘要、缓存和检索，尽量让 AI 成本保持在可预测范围内。

1. Do not call LLM for every message.
2. Use rule-based checks before AI when possible.
3. Use rules, templates, cache, and relevant Knowledge before model generation.
4. Never send full documents to the AI model.
5. Never send full conversation history to the AI model.
6. Use conversation summaries instead of long chat history.
7. Cache repeated FAQ answers.
8. Use the lowest-cost safe response method.
9. Use larger models only when complex understanding is genuinely required.
10. Stop AI generation when Human Handover is triggered.
11. Track token usage per Company, Model, and Feature.
12. Limit input, output, retrieved Knowledge, Context, Attachments, and Retries.
13. Prevent duplicate Calls and Retry Loops.
14. Pause or limit AI when abnormal usage is detected.

Target Architecture

Customer Message
↓
Rule / Fast Check
↓
Intent Detection
↓
Knowledge Retrieval
↓
Cache Check
↓
Need AI?
↓
Yes → System selects response method → Generate Reply
↓
No → Template Reply / Human Handover

这个流程图表达的是“先判断、再检索、最后才决定要不要真的调用模型”。

Cost Control Features

• Per-company usage dashboard
• Monthly Budget
• Warning Level
• Monthly Hard Limit
• Estimated Cost
• Duplicate and Retry Protection
• Abnormal Usage Detection
• Automatic AI Pause when required

When the Monthly Hard Limit is reached, paid AI Requests stop, Customer Conversations move to Human Handover, and Employee AI shows Usage Limit Reached.

客户只需要看到用了多少、预算警告和 AI 是否暂停；系统不会要求客户切换模式或选择模型。

Customer-Owned AI

Each company owns its own AI API. The platform should not require the platform operator to provide AI credits.

这一点是商业模式关键：客户自己提供 AI API Key，平台提供的是基础设施、流程和知识执行能力，而不是代付所有模型费用。

• API Key
• AI Cost
• Company Knowledge
• Company Conversations

The platform only provides the infrastructure.

WhatsApp Strategy

WhatsApp is the primary customer communication channel.

客户沟通的第一主场是 WhatsApp，因此所有自动回复、转人工、继续跟进都要尽量留在同一个对话场景里完成。

Customer
↓
WhatsApp
↓
AI Customer Assistant
↓
Intent Detection
↓
Approved Knowledge
↓
Reply
↓
Need Human?
↓
Create Case
↓
Notify Staff
↓
Staff Takes Over
↓
Continue WhatsApp Conversation

Future channels: Website Chat, Facebook Messenger, Telegram, Instagram, Email.

Customer data must never be shared across companies.

安全不是后面再补，而是第一版就必须把公司资料隔离、API Key 加密、对话保护和审计记录一起纳入。

Platform Console

平台运营管理层。

• Dashboard
• Customer Management
• Workspace Creation
• Deployment
• Monitoring
• Logs
• Backup
• Support

Company Workspace

客户实际使用层。

• Dashboard
• Knowledge
• Conversations
• Employee AI
• Analytics
• Settings

Core Engine

统一 AI 执行层。

• Knowledge processing
• AI routing
• Human Handover
• Cost Optimizer
• Security

这些属于后续扩充方向，先不进入第一版核心开发，但它们已经代表平台长期可以延伸的能力边界。

Current Phase

Research & Architecture

Development should NOT begin until the technical architecture is finalized.

当前阶段仍然是研究与架构确认，不是正式大规模开发阶段。

Current Goals

• Validate technical feasibility.
• Evaluate AI technologies.
• Select the most suitable architecture.
• Decide on AI stack.
• Decide on RAG implementation.
• Decide on WhatsApp integration.
• Decide on Security architecture.

这部分定义的是以后做每一个功能时都要回头检查的原则，避免系统越做越重、越做越散。