A hands-on course in the AI that companies actually use — with every module grounded in e-commerce work: product listings, customer support, marketing, ops data, and the internal tools that tie them together. No academic theory beyond what you need to make good decisions.

---

Course Overview — 課程簡介

Learning outcomes — 學習成果

By the end you will be able to:

1. Explain what LLMs can and cannot do, and pick the right tool for a business task.
2. Write reliable, reusable prompts and system instructions for your team.
3. Generate and QA product copy, emails, ads, and images at scale with a brand voice.
4. Analyse sales, inventory, and customer data with AI — and verify the results.
5. Build a knowledge bot over your own docs (RAG) for support and staff.
6. Automate multi-step workflows (orders, reviews, tickets, leads) with Zapier / Make / n8n and APIs.
7. Use AI coding assistants to build and maintain your own tools.
8. Run models locally for private data.
9. Ship a simple AI web app that calls an LLM API.
10. Govern AI use: security, privacy, cost, and ROI measurement.

Schedule — 時間表

---

Prerequisites — 先修要求

• Comfortable with spreadsheets (Google Sheets or Excel) and basic e-commerce admin (e.g. Shopify, WooCommerce, Amazon Seller Central).
• Optional but helpful for M7, M10, M11: basic command line and reading simple Python or JavaScript.
• Accounts: one paid AI assistant (ChatGPT Plus, Claude Pro, or Gemini Advanced/Google AI Pro), a free Zapier/Make or self-hosted n8n account, a GitHub account, and an API key from OpenAI, Anthropic, or Google AI Studio (budget a few dollars of usage).
• A sample dataset: your store's order/product exports, or a public e-commerce dataset (e.g. the Online Retail dataset on the UCI Machine Learning Repository, or Olist on Kaggle).

---

Module 1 — AI Foundations for Business · 商業AI基礎

Learning objectives

• Understand LLMs at a practical level: tokens, context windows, temperature, model tiers (fast vs. reasoning models).
• Recognise failure modes: hallucination, stale knowledge, prompt injection, bias.
• Classify company data by sensitivity and know which tools are safe for which data.

Lessons

1. What an LLM actually does (next-token prediction, why it sounds confident when wrong).
2. Model landscape: OpenAI GPT models, Anthropic Claude, Google Gemini, Meta Llama, Mistral, Qwen, DeepSeek — closed vs. open-weight.
3. Limits & hallucination: when to trust, when to verify, grounding with sources.
4. Data privacy: consumer vs. business/enterprise plans, training opt-outs, data retention, customer PII and GDPR/CCPA basics.
5. Where AI pays off in e-commerce: high-volume, text-heavy, repetitive, reviewable tasks.

Hands-on exercise: Ask three assistants the same five questions about your niche (one factual, one numerical, one about a recent event, one policy question, one creative). Record and score errors.

Tools covered: ChatGPT, Claude, Gemini, Microsoft Copilot.

Deliverable: A one-page "AI opportunity map" listing 10 tasks in your business, scored by volume, risk, and data sensitivity.

---

Module 2 — Prompting & AI Assistants · 提示工程與AI助手

Learning objectives

• Write structured prompts that produce consistent, reviewable output.
• Configure persistent assistants (Custom GPTs, Claude Projects, Gemini Gems).
• Choose the right assistant for a task.

Lessons

1. Prompt anatomy: role, context, task, constraints, examples, output format.
2. Few-shot examples, step-by-step reasoning, and asking for structured output (tables, JSON).
3. Iteration: critique-and-revise loops, self-check prompts, comparing outputs.
4. Persistent assistants: Custom GPTs, Claude Projects, Gemini Gems, Microsoft Copilot in Microsoft 365.
5. Building a team prompt library (naming, versioning, owners).

Hands-on exercise: Build a "Store Ops Assistant" project with your brand guide, shipping/returns policy, and product catalogue attached; test it on 10 real staff questions.

Tools covered: ChatGPT (Custom GPTs, Projects), Claude (Projects), Gemini (Gems), Microsoft 365 Copilot.

Deliverable: A prompt library of 10 tested prompts (in a shared doc or Notion) with input/output examples.

---

Module 3 — AI for Writing, Content & Product Listings · AI寫作、內容與商品文案

Learning objectives

• Produce on-brand product descriptions, emails, and social posts at scale.
• Localise content (e.g. English ↔ Traditional Chinese) with quality checks.
• Set up a human review step that catches errors before publishing.

Lessons

1. Brand voice guides as prompts: tone, banned words, reading level.
2. Product listings: titles, bullets, descriptions, SEO meta, alt text — from a spec sheet or CSV.
3. Email and social: launch sequences, abandoned-cart copy, captions, ad variants.
4. Translation and localisation; cultural review.
5. QA checklist: factual claims, sizes/materials, compliance wording, duplicates.

Hands-on exercise: Take a 20-row product CSV and generate titles, descriptions, and SEO meta for all rows; import into a Shopify development store or a spreadsheet.

Tools covered: ChatGPT / Claude, Shopify Magic, Grammarly, DeepL, Jasper (optional), Google Docs with Gemini.

Deliverable: Brand voice prompt + 20 published-ready listings + a QA checklist.

---

Module 4 — AI for Data Analysis & Spreadsheets · AI數據分析與試算表

Learning objectives

• Use AI to clean, analyse, and chart sales, inventory, and customer data.
• Write formulas, pivots, and simple Python analyses with AI help.
• Verify AI-generated numbers.

Lessons

1. Uploading data to ChatGPT (data analysis), Claude (analysis tool), and Julius AI; what happens to your file.
2. Copilot in Excel and Gemini in Google Sheets: formulas, pivots, summaries.
3. Core e-commerce analyses: AOV, repeat rate, cohort retention, RFM segments, SKU ABC analysis, stock-out risk.
4. Charts and narrative summaries for weekly reports.
5. Verification: recompute key figures, sanity checks, reproducible code.

Hands-on exercise: With a 12-month order export, produce an RFM segmentation, top/bottom SKUs, and a monthly trend chart. Re-check three numbers manually.

Tools covered: ChatGPT data analysis, Claude, Julius AI, Microsoft Copilot in Excel, Gemini in Google Sheets, Python (pandas) in Google Colab.

Deliverable: A one-page weekly KPI report template with AI-written commentary, plus the analysis notebook or prompt chain.

---

Module 5 — Research & Knowledge Management (RAG) · 研究與知識管理（RAG）

Learning objectives

• Use AI research tools with citations for market and competitor research.
• Understand retrieval-augmented generation (RAG): chunking, embeddings, vector search.
• Build a searchable knowledge bot over internal docs.

Lessons

1. Research with citations: Perplexity, ChatGPT search / deep research, Gemini Deep Research.
2. NotebookLM for source-grounded Q&A over PDFs, docs, and websites.
3. How RAG works: chunking, embeddings, vector databases, re-ranking, citing sources.
4. Enterprise search options: Glean, Microsoft 365 Copilot, Notion AI, Google Agentspace.
5. Keeping knowledge fresh: ownership, update cadence, evaluation questions.

Hands-on exercise: Load your policies, supplier docs, and FAQ into NotebookLM or a Claude Project; then write a 20-question evaluation set and score answers for accuracy and citation.

Tools covered: Perplexity, NotebookLM, ChatGPT search, Gemini Deep Research, Notion AI, Glean (overview), Chroma / pgvector (concept).

Deliverable: A competitor research brief with citations + an internal knowledge bot with an evaluation scorecard.

---

Module 6 — Workflow Automation & AI Agents · 工作流程自動化與AI代理

Learning objectives

• Design trigger → AI step → action workflows.
• Connect store, email, sheets, and chat tools via no-code platforms and APIs.
• Understand agents (tool use, multi-step tasks) and where to keep a human in the loop.

Lessons

1. Automation building blocks: triggers, actions, filters, webhooks, JSON.
2. Zapier (incl. Zapier AI actions), Make, and n8n (self-hostable) compared.
3. E-commerce recipes: new review → sentiment + reply draft; new order → fraud/notes check; low stock → supplier email draft; tickets → classify and route.
4. Agents: tool calling, MCP (Model Context Protocol), agent builders (OpenAI Agents SDK, n8n AI Agent node); approval steps.
5. Reliability: error handling, logging, rate limits, cost per run.

Hands-on exercise: Build a workflow: new product review (Shopify/Judge.me or a form) → LLM classifies sentiment and topic → log to Google Sheets → Slack alert for negative reviews with a suggested reply.

Tools covered: Zapier, Make, n8n, Slack, Google Sheets, OpenAI / Anthropic APIs, webhooks.

Deliverable: Two working automations with a diagram, a run log, and an estimated monthly cost.

---

Module 7 — AI Coding Assistants · AI編程助手

Learning objectives

• Use AI coding tools to build scripts, small apps, and integrations — even with limited coding background.
• Review, test, and safely run AI-written code.

Lessons

1. The landscape: Cursor, GitHub Copilot, Claude Code, OpenAI Codex, Replit, Lovable, v0, Bolt.
2. Working effectively: describe the goal, give context files, iterate in small steps.
3. Practical builds: Shopify Admin API scripts, CSV transformers, scrapers for your own data, dashboards.
4. Git basics for non-engineers: commits, branches, reverting mistakes.
5. Safety: secrets in .env, never paste API keys into prompts, code review, testing.

Hands-on exercise: With Cursor or GitHub Copilot, build a script that pulls the last 30 days of orders via the Shopify Admin API (or reads a CSV) and outputs a daily sales summary to Slack.

Tools covered: Cursor, GitHub Copilot, Claude Code, Replit, GitHub.

Deliverable: A Git repository with a working script, README, and .env.example.

---

Module 8 — Customer Support, Sales & Marketing · 客戶服務、銷售與市場推廣

Learning objectives

• Deploy AI support with safe escalation paths.
• Use AI for personalisation, segmentation, and ad creative.
• Measure the impact of AI in customer-facing work.

Lessons

1. AI support agents: Intercom Fin, Zendesk AI, Gorgias AI Agent, Tidio Lyro; grounding on your help centre; handoff rules.
2. Order-status, returns, and sizing questions — what to automate and what never to automate (refund exceptions, angry VIPs, legal).
3. Email/SMS marketing AI: Klaviyo AI features, subject-line testing, segmentation.
4. Ads and creative: Meta Advantage+, Google Performance Max, generating and testing ad variants.
5. Sales and B2B: lead research and outreach drafts (HubSpot Breeze, Apollo), with personalisation that is accurate.

Hands-on exercise: Write the knowledge base and guardrails for a support bot; test it with 25 real tickets (including 5 adversarial ones) and grade the answers.

Tools covered: Intercom Fin, Zendesk AI, Gorgias, Tidio, Klaviyo, Shopify Inbox, HubSpot, Meta Ads Manager, Google Ads.

Deliverable: A support bot configuration with escalation rules + a test report; one AI-assisted email campaign brief.

---

Module 9 — Meetings, Voice, Image & Video Tools · 會議、語音、圖像與影片工具

Learning objectives

• Automate meeting notes and action items.
• Produce product photos, lifestyle images, and short videos with AI.
• Understand rights, disclosure, and brand consistency for generated media.

Lessons

1. Meetings: Otter.ai, Fireflies.ai, Granola, Zoom AI Companion, Microsoft Teams Copilot, Google Meet with Gemini.
2. Product imagery: background removal and scene generation (Photoroom, Canva Magic Studio, Adobe Firefly, Shopify Magic image editing).
3. Image generation: ChatGPT image generation, Midjourney, Ideogram, Google Imagen — prompts for consistent style.
4. Video and voice: Runway, Google Veo, CapCut, HeyGen, Synthesia, ElevenLabs — product demos, UGC-style ads, multilingual voice-overs.
5. Rights & disclosure: commercial licences, likeness consent, platform labelling rules, not misrepresenting the product.

Hands-on exercise: Turn 3 plain product photos into a set of marketplace-ready images (white background + 2 lifestyle scenes) and one 15-second vertical video ad with voice-over.

Tools covered: Otter.ai, Fireflies.ai, Granola, Photoroom, Canva, Adobe Firefly, Midjourney, Runway, CapCut, HeyGen, ElevenLabs.

Deliverable: A creative asset pack + a reusable image-prompt style guide.

---

Module 10 — Local & Private Models · 本地與私有模型

Learning objectives

• Run open-weight models on your own machine or server.
• Decide when local beats cloud (privacy, cost, offline) and when it doesn't (quality, speed).
• Connect a local model to your tools via an OpenAI-compatible API.

Lessons

1. Open-weight models: Llama, Qwen, Mistral, Gemma, DeepSeek; parameter size vs. hardware (RAM/VRAM); quantisation (GGUF).
2. Ollama and LM Studio: install, pull, chat, serve an API.
3. llama.cpp: lightweight, high-control inference; vLLM for server deployments.
4. Private RAG: local embeddings + Chroma, Open WebUI or AnythingLLM as a team front end.
5. Trade-offs: quality benchmarks on your tasks, maintenance, licences.

Hands-on exercise: Install Ollama, run two models, and compare them against a cloud model on 10 of your Module 2 prompts (e.g. classifying customer emails that contain PII).

Tools covered: Ollama, LM Studio, llama.cpp, Open WebUI, AnythingLLM, Hugging Face.

Deliverable: A local-vs-cloud comparison table with a recommendation for which tasks stay local.

---

Module 11 — Build a Simple AI App with an API · 用API構建簡單AI應用

Learning objectives

• Call LLM APIs directly: messages, system prompts, structured outputs, tool calling.
• Build and deploy a small web app for a real business task.
• Track cost and add basic evaluation.

Lessons

1. API fundamentals: OpenAI, Anthropic, and Gemini APIs; keys, tokens, pricing per million tokens, rate limits.
2. Structured outputs (JSON schemas) and function/tool calling.
3. A minimal app with Streamlit (Python) or Next.js; optional frameworks (LangChain, LlamaIndex, Vercel AI SDK).
4. Adding RAG: embeddings, a vector store (pgvector, Chroma), citations.
5. Deploy & operate: Streamlit Community Cloud, Vercel, or Render; logging, caching, evals, cost caps.

Hands-on exercise: Build a "Listing Generator" app: upload a product spec CSV → app returns structured title/bullets/description/SEO JSON → export CSV ready for Shopify import.

Tools covered: OpenAI API, Anthropic API, Google AI Studio / Gemini API, Streamlit, Vercel, Cursor.

Deliverable: A deployed app URL + repo + a cost-per-100-products estimate.

---

Module 12 — Governance, Security & ROI Measurement · 管治、安全與投資回報衡量

Learning objectives

• Write a practical AI usage policy.
• Identify and mitigate AI-specific security risks.
• Measure ROI with baselines rather than assumptions.

Lessons

1. AI usage policy: approved tools, data classes allowed per tool, review requirements, disclosure.
2. Security: prompt injection, data leakage, over-permissioned agents, API key management, vendor review (SOC 2, DPA).
3. Regulation awareness: GDPR, CCPA, EU AI Act, consumer-protection rules on reviews and advertising, Hong Kong PDPO guidance on AI.
4. ROI: baseline time/cost per task, measure after, include quality and error costs, subscription + API spend.
5. Change management: champions, training, feedback loops, retiring tools that don't deliver.

Hands-on exercise: Pick two automations from earlier modules; measure the before/after time per task and error rate over one week.

Tools covered: Policy templates, Google Sheets/Excel ROI model, vendor trust centres, API usage dashboards (OpenAI, Anthropic).

Deliverable: A one-page AI policy + an ROI sheet for two use cases.

---

Hands-on Track — AI Infrastructure for Real Business Problems · 實作專題：AI基礎架構

A code-first companion to the modules. Ten labs build a working AI platform for one fictional store, Kettle & Kiln (an online tea-ware shop), on one shared dataset (hands-on/data/). Every lab has:

• a lesson with a business problem, an architecture diagram, exact setup commands, tests, cost notes, common failures, a stretch exercise and a deliverable checklist
• runnable code under hands-on/labs/NN/: Python 3.12, FastAPI, Docker Compose, .env.example, tests

All tests run offline with a mock model; live results in the lessons are single CPU runs with small local models, labelled as such. No prices or benchmarks are quoted. Start with the track overview.

---

Capstone Project — 畢業專題 (Weeks 9–10)

Brief: Design, build, and present an AI solution that solves a real problem in an e-commerce business (yours or a realistic case). It must combine at least three course elements (e.g. RAG + automation + API app) and include governance and ROI.

Example projects

• AI support agent grounded on the help centre with order-status lookup and human escalation.
• Product catalogue pipeline: supplier spec → listings + images + translations → Shopify import, with QA gates.
• Weekly ops copilot: pulls sales/inventory data, flags anomalies, drafts reorder emails for approval.
• Review intelligence: collects reviews across channels, clusters themes, feeds product and marketing teams.

Required submissions: problem statement & baseline metrics · architecture diagram · working demo (video or live) · repo or workflow export · AI policy notes for this use case · ROI estimate · 10-minute presentation.

Grading rubric — 評分準則

Pass mark: weighted average ≥ 2.5.

---

30-60-90 Day Company Rollout Plan — 30-60-90天公司推行計劃

---

Tool Cheat Sheet — 工具速查表

Cost tiers: Free = usable free tier · $ = roughly under US$30/user/month · $$ = higher subscription or usage-based · $$$ = enterprise/sales-led. Pricing changes often — check vendor sites.

---

Glossary — 詞彙表

• LLM (Large Language Model / 大型語言模型): A model trained on large text corpora to predict and generate text.
• Token: A chunk of text (part of a word) — the unit for model limits and API pricing.
• Context window: How much text a model can consider at once.
• Hallucination (幻覺): Confident but false or unsupported output.
• Prompt / System prompt: Instructions to the model; the system prompt sets persistent behaviour.
• Few-shot: Including examples in a prompt to show the desired output.
• RAG (Retrieval-Augmented Generation / 檢索增強生成): Retrieving relevant documents and giving them to the model so answers are grounded and citable.
• Embedding: A numeric vector representing meaning, used for semantic search.
• Vector database: Storage optimised for similarity search over embeddings (e.g. pgvector, Chroma).
• Agent (AI代理): An AI system that plans and takes actions using tools, often over multiple steps.
• Tool / function calling: A model outputting a structured request to run a function or API.
• MCP (Model Context Protocol): An open standard for connecting AI apps to tools and data sources.
• Fine-tuning: Further training a model on your examples to change its behaviour.
• Open-weight model: A model whose weights can be downloaded and run yourself (e.g. Llama, Qwen, Mistral).
• Quantisation: Compressing a model (e.g. GGUF 4-bit) to run on smaller hardware.
• Prompt injection: Malicious text in inputs (emails, web pages, reviews) that tries to hijack the model's instructions.
• Eval (評估): A repeatable test set to measure AI output quality.
• Human-in-the-loop: A person reviews or approves AI output before it takes effect.
• PII: Personally identifiable information (names, addresses, emails, phone numbers).
• ROI: Return on investment — value gained versus total cost (tools, usage, time, errors).