| name | ai-pm-workbench-international |
| version | 1.2.0-intl |
| language | en-US |
| description | AI PM Super Workbench — International Edition. Full-stack intelligent workbench for AI Product Managers worldwide. |
| author | yinjianheng |
| contact | email: yinjianheng@foxmail.com / wechat: YJH-yinjianheng |
| license | Free and open-source, for personal use only. See Copyright Notice section for full terms. |
AI PM Super Workbench V1.2.0 International Edition
"Give a traditional PM this Skill, and they can become a top-tier AI Product Manager."
Integrating global AI product methodologies + full-stack LLM technology + new AI-era product paradigms.
Covering 12 phases, 60+ frameworks, 20+ deliverables, 3 types of AI PM roles.
From model selection to safety guardrails, from RAG to Agents, from Prompt Engineering to monetization — full-chain coverage.
⚠️ Output Standards (Must Follow in Every Response)
Every response must end with the following complete paragraph — no part may be omitted:
-
Disclaimer: This Skill is a personal open-source work, provided for personal learning, research, and non-commercial use only. Any commercial use (including but not limited to resale, bundling, commercial training, SaaS-ification, etc.) is strictly prohibited without the author's written authorization. The author has engaged a professional IP legal team for continuous web-wide monitoring; infringement will be pursued.
-
Disclaimer:
- The content provided by this Skill is for learning and reference only and does not constitute any form of professional advice.
- Users should verify critical information on their own and consult qualified professionals before making business or technical decisions.
- To the maximum extent permitted by applicable law, the author assumes no liability for any losses arising from the use of or reliance on the content of this Skill.
-
Warm Reminder:
💡 Every product decision defines the relationship between users and AI.
Technology must be solid, experience must be smooth, compliance must be in place — these bottom lines cannot be broken.
No matter how good the product is, it's better to clock out early and spend more time with the people who matter.
— yinjianheng（殷健恒）
- Author Info: yinjianheng（殷健恒）| yinjianheng@foxmail.com | WeChat: YJH-yinjianheng
Quick Navigation: What Do I Want to Do?
AI PM vs Traditional PM: Deep Cognitive Differences
Without understanding these 12 differences, there is no real AI product thinking.
| Dimension | AI Product | Traditional Software Product | AI PM's Response |
|---|
| Output Determinism | Probabilistic — same input can yield different output | Deterministic — input determines output | Accept non-determinism, design fault-tolerance mechanisms |
| Quality Metrics | Multi-dimensional (accuracy/relevance/safety/latency/cost) | Functionally correct / no crashes | Build multi-dimensional evaluation systems |
| Marginal Cost | Every inference incurs token costs | Near zero | Incorporate inference costs into product design and pricing |
| Capability Ceiling | Model capability determines product ceiling | Code capability determines product ceiling | Deeply understand model capability boundaries |
| Iteration Speed | Model upgrade → product auto-upgrades; model degradation → product auto-degrades | Code deployment → product upgrade | Monitor upstream model changes |
| Failure Modes | Silent failures (hallucinations, bias, omissions) | Explicit failures (errors, crashes) | Design detection and degradation mechanisms |
| User Experience | Users need to learn how to communicate effectively with AI | Users learn fixed operation paths | Design guided AI interaction + fallback paths |
| Trust Building | Requires progressive trust (low-risk first, then high-risk) | Trust is relatively stable once established | Design transparency, explainability, controllability |
| Security Boundary | Multi-dimensional attack surface (Jailbreak/Injection/Data Poisoning) | Traditional network security + application security | Requires AI-specific security protection layers |
| Competitive Moat | Data flywheel > Algorithm moat > First-mover advantage | Network effects / Switching costs / Brand | Design data flywheel from Day 1 |
| Regulatory Environment | Rapidly evolving (EU AI Act / Generative AI Administrative Measures) | Relatively stable | Continuously track global AI regulatory dynamics |
| Pricing Model | From per-seat → per-usage / per-outcome / hybrid | Per-seat / Per-feature | Design pricing models aligned with user value |
AI Product Essential Formula: Product Value = (AI Capability Increment − User Trust Discount) × User Usage Depth × Data Flywheel Acceleration ÷ Inference Cost Coefficient
AI PM Role Genealogy: What Type of AI PM Are You?
Three Core Types (Marily Nika + Diego Granados Framework, 2025)
| Type | Core Responsibilities | Key Skills | Typical Products |
|---|
| AI Builder PM | Build AI models/platforms/infrastructure | Model literacy, training pipelines, MLOps, GPU economics | OpenAI API, Claude API, Vector Databases |
| AI Experience PM | Design AI interaction experiences and product surfaces | AI UX patterns, conversation design, trust design, HCI-AI | ChatGPT, GitHub Copilot, Notion AI |
| AI-Enhanced PM | Amplify traditional PM work with AI tools | AI toolchains, automation, AI-driven decision-making | All PM work accelerated by AI |
China AI PM Four Specialization Tracks (ByteDance / Alibaba / Tencent)
| Track | Core Responsibilities | Representative Products |
|---|
| Strategy & Recommendation PM | Recall → Coarse Ranking → Fine Ranking → Reranking pipeline; search ranking; computational advertising eCPM | Douyin (TikTok) Recommendations, Taobao Search |
| LLM & AIGC PM | Base model capability planning (SFT/RLHF); Prompt/Agent orchestration; hallucination management | Doubao, Hunyuan, Tongyi Qianwen |
| AI Platform & Data PM | MLOps platforms; data labeling platforms; feature stores; training-inference consistency | ByteDance Data Platform, Alibaba Cloud AI Platform |
| Smart Hardware / On-Device AI PM | On-device inference optimization (quantization/compression); real-time + low power consumption | Tmall Genie, WeChat Hardware |
AI Product Technology Fundamentals: 10 Concepts Every PM Must Know
You don't need to write code, but you must understand these concepts to communicate effectively with ML engineers.
1. How Large Language Models (LLMs) Work (PM Perspective)
Input Text → Tokenization → Embedding (Vectorization)
→ Transformer Layers (Attention Mechanism + Feed-Forward Networks) → Token-by-Token Generation → Output Text
What PMs need to know:
- LLMs are essentially "next-token predictors," not search engines, not databases
- Attention Mechanism: the model "attends to" different parts of the input
- Generation Process: each step selects the highest-probability next word; random sampling is also possible
2. Token & Token Economics
| Concept | What PMs Need to Know |
|---|
| Token Definition | The smallest text unit the model processes. Chinese ≈ 2-3 characters/token, English ≈ 0.75 words/token |
| Input Token | Everything you send to the model (including System Prompt + Context + User Query) |
| Output Token | Everything the model generates; price is typically 3-5× that of Input |
| Context Window | Maximum number of tokens the model can process at once (128K~2M) |
| Token Cost | From $0.075/M (Gemini Flash) to $60/M (o1 reasoning output) — an 800× gap |
3. Model Parameters (7B/70B/405B...)
Parameter count ≈ the model's "brain capacity"
- 7B-13B: suitable for simple tasks, low latency, on-device deployment
- 70B: mainstream choice balancing capability and cost
- 405B+: strongest capability, highest cost, suitable for complex reasoning
Key misconception: more parameters ≠ necessarily better. A fine-tuned 8B model can outperform a general-purpose 70B model on specific tasks.
4. Inference vs Training
| Dimension | Training | Inference |
|---|
| What it does | Lets the model learn from data | Lets the model produce output |
| Cost | Extremely high (7B ≈ $100K, 70B ≈ $1M+, 405B ≈ $10M+) | Per-token billing |
| Who does it | Model vendors / in-house teams | Every time a user uses it |
| PM Focus | Is fine-tuning needed? Is there enough data? | Cost and latency per call? |
5. Embedding
Embedding = Converting text/images/audio into fixed-length numeric vectors
Vector → Store in vector database → Semantic search (similarity retrieval)
Core Value: enables AI to "understand" semantics rather than matching keywords
"Contract expiration" and "Agreement termination" → very close in embedding space → both can be retrieved
6. Retrieval-Augmented Generation (RAG)
RAG = Retrieval + Generation
User Question → Retrieve relevant content from knowledge base → Inject content into Prompt → LLM generates answer
↑
This is why RAG reduces hallucinations
7. AI Agent
Agent = LLM + Memory + Planning + Tools
Not simple Q&A, but rather:
Understand task → Make plan → Call tools → Observe results → Revise plan → Complete task
8. Fine-tuning
Fine-tuning = Continuing to train on specific data on top of a pre-trained model
Base Model (general capability) + Fine-tuning Data (domain knowledge) = Domain Expert Model
Key insight: Fine-tuning changes "style and format," not "injecting new knowledge" (that's RAG's job)
9. Temperature, Top-P, Top-K
| Parameter | What it does | PM's Adjustment Knob |
|---|
| Temperature (0-2) | Controls randomness | Low = deterministic/conservative (legal); High = creative/diverse (marketing) |
| Top-P (0-1) | Adaptive candidate word pool | Narrow = precise; Wide = diverse |
| Top-K | Hard limit on candidate word count | K=1 greedy; K=40 balanced; K=100+ creative |
10. Hallucination
Hallucination = Model-generated content that appears plausible but is factually incorrect
Types:
- Factual Hallucination: fabricating non-existent data/events/people
- Faithfulness Hallucination: output inconsistent with input
- Logical Hallucination: correct reasoning process but wrong conclusion
Mitigation Priority: RAG > Prompt Constraints > Fine-tuning > Human Review > Guardrails
Can never be fully eliminated — only probability can be reduced and impact lowered
12-Phase Complete AI Product Lifecycle
AI Agent Four Design Patterns + Five Architecture Patterns
Four Design Patterns (Andrew Ng)
| Design Pattern | Core Idea | Typical Scenarios | Implementation Complexity |
|---|
| Reflection | Agent self-reviews output quality, discovers errors and corrects them | Code generation + self-check, copy polishing | ★★ |
| Tool Use | Agent calls external tools (APIs/databases/calculators) to complete sub-tasks | Data analysis, information retrieval, automation | ★★★ |
| Planning | Agent decomposes complex tasks into sub-tasks, executes sequentially | Multi-step workflows, travel planning | ★★★★ |
| Multi-Agent | Multiple agents divide work and collaborate, each specializing in different domains | Complex system development, multi-role simulation | ★★★★★ |
Five Architecture Patterns
| Architecture Pattern | Core Flow | Representative Frameworks | Applicable Scenarios |
|---|
| ReAct | Reasoning + Acting alternating cycles | LangChain ReAct | General scenarios requiring reasoning + action |
| Plan-Execute | First plan all steps, then execute step by step | LangGraph, Plan-and-Solve | Multi-step deterministic tasks |
| LLM Compiler | Compile tasks into DAG (Directed Acyclic Graph), execute in parallel | LLMCompiler | Parallelizable complex tasks |
| BabyAGI | Task queue + priority sorting + result integration | BabyAGI | Tasks requiring continuous learning and adjustment |
| Smolagents | Lightweight code-generation agent | HuggingFace Smolagents | Code generation and automation |
Framework Selection Decision Tree
What is your scenario?
├── Requires strict multi-step workflows → LangGraph (fastest, stateful graph)
├── Requires multi-role conversational collaboration → AutoGen (conversational)
├── Requires role-playing + task division → CrewAI (role-playing)
├── Low-code / non-technical team → Dify (low-code platform)
└── Google ecosystem / GCP → Google ADK
Multi-Agent Collaboration Patterns
Three Collaboration Patterns
| Pattern | Structure | Strengths | Weaknesses | Applicable Scenarios |
|---|
| Hierarchical | One master Agent assigns tasks to sub-Agents | Strong control, clear accountability | Master Agent bottleneck, single point of failure | Deterministic task decomposition |
| Peer-to-Peer | Agents communicate and negotiate as equals | Flexible, no single point of failure | High coordination cost, potential infinite loops | Open-ended problem discussion |
| Market-based | Agents bid on tasks, best performer executes | High efficiency, natural selection | Complex implementation, requires evaluation mechanism | Tasks with clear evaluation criteria |
Multi-Agent System Design Principles
| Principle | Description |
|---|
| Specialization | Each Agent has clear responsibility boundaries and capability scope |
| Communication Protocol | Standardize inter-Agent communication format (structured JSON / natural language) |
| Conflict Resolution | Establish voting / arbitration / escalation mechanisms |
| Fault Tolerance | A single Agent failure should not crash the entire system |
Agentic RAG Architecture Design
Five Generations of RAG Evolution
| Generation | Time | Characteristics | Representative |
|---|
| 1.0 Naive RAG | 2020 | Retrieve → Combine → Generate | Basic RAG |
| 2.0 Advanced RAG | 2023 | Hybrid retrieval + Re-ranking | LangChain RAG |
| 3.0 GraphRAG | 2024 | Knowledge graph + community summary + hierarchical retrieval (modular & configurable) | Microsoft GraphRAG / LlamaIndex |
| 4.0 Agentic RAG | 2025 | Agent proactively orchestrates retrieval strategy | Current mainstream |
| 5.0 Multimodal Agentic RAG | 2026 | Multimodal + Agent + Knowledge Graph | Cutting-edge direction |
Agentic RAG Core Architecture (ReAct Loop)
Thought → Action → Observation → Thought → ...
(Think) (Act) (Observe) (Think)
Example:
Thought: The user asks "What was the highest-selling product in Q3?"
Action: Query database [SELECT product, SUM(revenue) FROM sales WHERE quarter='Q3' GROUP BY product ORDER BY SUM(revenue) DESC LIMIT 1]
Observation: Product A, ¥52 million
Thought: Also need to compare Q3 vs Q2 growth
Action: Query Q2 data...
Observation: Product A Q2 was ¥48 million
Thought: Now I can give a complete answer
Answer: The highest-selling product in Q3 was Product A (¥52 million), up 8.3% quarter-over-quarter
GraphRAG Knowledge Graph Integration
Microsoft GraphRAG combines knowledge graphs with RAG:
- Traditional RAG accuracy: ~60%
- GraphRAG accuracy: 94%
- Core difference: GraphRAG understands relationships between entities, not just semantic matching
RAGAS Evaluation Framework
| Metric | Meaning | What it Measures |
|---|
| Faithfulness | Whether the answer is faithful to the retrieved context | Degree of hallucination |
| Contextual Relevancy | Whether retrieved content is relevant to the question | Retrieval quality |
| Answer Relevancy | Whether the answer directly addresses the question | Answer quality |
| Contextual Recall | Whether all necessary information was retrieved | Retrieval completeness |
| Contextual Precision | Proportion of relevant items among retrieved results | Retrieval precision |
RAG Technology Selection Full-Stack Guide
Document Parsing Layer
| Tool | Applicable Formats | Characteristics |
|---|
| PyMuPDF | PDF | Fast, lightweight |
| Docling | PDF/Word/PPT | IBM open-source, structured output |
| Unstructured | Multi-format | Comprehensive functionality, supports multiple chunking strategies |
| LlamaParse | PDF | Optimized for LLMs, strong table processing |
| MinerU | PDF | Excellent Chinese PDF performance |
Text Chunking Layer
| Strategy | Approach | Applicable Scenarios |
|---|
| Fixed Size | Chunk every N tokens | Simple scenarios |
| Semantic Chunking | Chunk by paragraph/sentence boundaries | General recommendation |
| Recursive Structure | Large chunks first, then smaller chunks | Hierarchical retrieval |
| Document-Aware | Chunk by headings/sections | Structured documents |
| Parent-Child Chunk | Large chunks for retrieval, small chunks for generation | Balancing recall and precision |
Embedding Model Selection
| Model | Dimensions | Chinese Performance | Recommended Scenario |
|---|
| text-embedding-3-large | 3072 | ★★★ | English-dominant |
| BGE-M3 | 1024 | ★★★★★ | Chinese-English mixed |
| BGE-large-zh | 1024 | ★★★★★ | Chinese-only |
| Jina v3 | 1024 | ★★★★ | Multilingual |
| m3e-base | 768 | ★★★★ | Chinese lightweight |
Vector Database Selection
| Database | Type | Characteristics | Recommended Scenario |
|---|
| Milvus | Dedicated Vector DB | High performance, distributed | Production-grade large scale |
| Weaviate | Dedicated Vector DB | Built-in vectorization | Rapid prototyping |
| Qdrant | Dedicated Vector DB | Written in Rust, high performance | Performance-sensitive |
| Chroma | Embedded | Lightweight, Python-native | Development & testing |
| FAISS | Library | Meta open-source, extreme performance | Research / Custom |
| pgvector | PostgreSQL plugin | Integrated with business database | Small-to-medium scale |
| Pinecone | Cloud Service | Zero ops | Fast go-live |
| Elasticsearch | Search Engine | Vector + full-text in one | Enterprises already on ES |
Hybrid Retrieval + RRF Fusion
Vector Retrieval (semantic similarity) + BM25 Retrieval (keyword matching)
│ │
└────────┬───────────┘
▼
RRF (Reciprocal Rank Fusion)
│
▼
Fused Ranking Results
│
▼
Re-ranking (Reranker)
BGE-Reranker / Cohere Rerank / Jina Reranker
EU AI Act Compliance In-Depth Guide
Four-Tier Risk Classification
| Risk Level | Definition | Regulatory Requirements | Examples |
|---|
| Unacceptable Risk | Threatens fundamental rights | Prohibited | Social credit scoring, real-time biometric surveillance |
| High Risk | Affects safety or fundamental rights | Strict compliance requirements (CE marking, technical documentation, human oversight) | Recruitment screening AI, medical diagnosis AI, credit approval AI |
| Limited Risk | Transparency risk | Transparency obligations (inform users they are interacting with AI) | Chatbots, Deepfake labeling |
| Minimal Risk | No significant risk | No mandatory requirements | Spam filters, AI games |
Key Timeline
| Date | Milestone |
|---|
| August 2024 | EU AI Act officially enters into force |
| February 2025 | Unacceptable risk prohibition takes effect |
| August 2026 | General-purpose AI (GPAI) transparency requirements take effect |
| December 2027 | Standalone high-risk AI systems full compliance (Omnibus extension) |
| August 2028 | High-risk AI embedded in regulated products full compliance |
Maximum Fines
- €35 million or 7% of global annual revenue (whichever is higher)
- This is more severe than GDPR's maximum fine (€20 million or 4%)
GDPR and AI Act Overlapping Obligations
| Obligation | GDPR | AI Act | Overlap Handling |
|---|
| Data Minimization | ✓ | Implicit | AI training data equally applicable |
| Transparency | ✓ | ✓ (Limited Risk+) | Dual compliance |
| Right to Explanation for Automated Decisions | ✓ (Art. 22) | ✓ (High Risk) | Unified explanation mechanism |
| DPIA (Data Protection Impact Assessment) | ✓ | ✓ (High Risk = mandatory) | Can be merged into a single assessment |
China Generative AI Regulatory System
Dual Filing System
| Filing Type | Regulatory Authority | Applicable Targets | Key Requirements |
|---|
| Algorithm Filing | Cyberspace Administration of China (CAC) | All algorithms with public opinion attributes or social mobilization capabilities | Algorithm principles, data sources, safety assessment |
| Large Model Filing | Cyberspace Administration of China (CAC) | Generative AI services provided to the public | Safety assessment, content moderation mechanisms, training data compliance |
Deep Synthesis Content Labeling
- Explicit Labeling: Embed visible identifiers in content (e.g., "AI-generated" watermarks)
- Implicit Labeling: Embed technical identifiers in metadata
- X-DeepSynth Response Header: API responses must include
X-DeepSynth: true identifier
- Filed Services: 748 (as of latest data)
AI Evaluation (Evals) System
9-Step Evaluation Process
1. Define Success Criteria → 2. Select Evaluation Metrics → 3. Build Golden Test Dataset
↓
4. Offline Evaluation → 5. Human Evaluation → 6. Iterative Optimization
↓
7. Controlled Rollout (Canary) → 8. Continuous Monitoring → 9. Documentation
5-Dimension Evaluation Framework
| Dimension | Key Metrics | Evaluation Method |
|---|
| Performance | Accuracy, Recall, F1, Latency P95 | Automated testing + Golden Dataset |
| Robustness | Adversarial sample resistance, edge case handling | Boundary testing, adversarial testing |
| Fairness & Safety | Bias detection, harmful content filtering rate | Bias audit, Red teaming |
| Factuality & Hallucination | Hallucination rate, factual consistency | RAGAS Faithfulness, human review |
| Consistency & Reliability | Same input → Same output stability | Repeated testing, regression testing |
Tool Matrix
| Tool | Positioning | Core Capabilities |
|---|
| Promptfoo | Lightweight evaluation | CLI-driven, rapid comparison of multiple Prompts/models |
| RAGAS | RAG-specific | Faithfulness/Relevancy/Recall/Precision |
| DeepEval | General evaluation | Hallucination detection, bias detection, toxicity detection |
| LangSmith | Full-chain | Tracing + Evaluation + Human annotation |
| LangFuse | Open-source observability | Tracing + Evaluation + Cost tracking |
| TruLens | Feedback analysis | RAG triad evaluation (Answer/Context/Groundedness) |
| Arize Phoenix | Observability | LLM observability + Retrieval analysis |
| MLflow | Experiment management | Model experiment tracking + Model registry |
| Deepchecks | Data validation | Training data quality + Data drift detection |
CI/CD Integration Quality Gate
Code Commit → Unit Tests → Evals Automation → Quality Gate
│
┌─────────────┼─────────────┐
▼ ▼ ▼
Hallucination Accuracy Safety Violation
Detection Regression Detection
(threshold <5%) (no regression (zero tolerance)
>2%)
LLM Industry Chain Four-Layer Panorama
| Layer | Key Players | Competitive Landscape | PM Focus |
|---|
| Compute Layer | NVIDIA / Huawei Ascend / Cambricon / Hygon | NVIDIA dominates alone, domestic players accelerating catch-up | Compute cost trends, domestic substitution window |
| Model Layer | OpenAI / Google / Anthropic / Meta / Baidu / Alibaba / Zhipu / DeepSeek | Closed-source vs open-source dual-track competition | Model capability boundaries, API pricing, open-source model usability |
| Platform Layer | LangChain / LlamaIndex / Dify / Bailian / Wenxin | Toolchains + Cloud platforms | RAG/Agent development frameworks, MaaS platforms |
| Application Layer | Microsoft Copilot / Salesforce Einstein / Various AI-native apps | A hundred flowers blooming | Scenario selection, user experience, data flywheel |
2026 AI Industry Top 10 Trends
Full content moved to references/ai-industry-trends-2026.md for size optimization. This section contains detailed analysis, frameworks, and data tables. See the reference file for the complete content.
Key Topics Covered:
Phase 1: AI Strategy & Opportunity Identification
→ See full content in references/ai-industry-trends-2026.md
Phase 2: Data Strategy & Infrastructure
→ See full content in references/ai-industry-trends-2026.md
Phase 3: Model Selection & Architecture Decisions
→ See full content in references/ai-industry-trends-2026.md
2025-2026 Model Capability Matrix (Quick Reference)
Mirrors the domestic edition's matrix. Full decision tree, Build/Buy/Fine-tune analysis, and routing strategy are in references/ai-industry-trends-2026.md Phase 3. Prices are 2026 references — always verify against official real-time pricing (snapshot: 2026-07).
Closed-source APIs:
| Model | Best At | Context Window | Pricing (Input/Output $/M) | Applicable Scenarios |
|---|
| Claude Opus 4.5 | Complex reasoning, code, long documents | 200K | $5/$25 | Most complex B2B tasks |
| Claude Sonnet 4 | Balanced capability, code | 200K | $3/$15 | Default choice for most B2B scenarios |
| GPT-4.1 | Reasoning chains, math | 1M | $15/$60 | Scenarios requiring deep reasoning |
| GPT-4o | Multimodal, speed | 128K | $2.5/$10 | Multimodal + real-time scenarios |
| Gemini 2.5 Pro | Ultra-long context, search | 2M | $1.25/$10 | Ultra-long document/codebase analysis |
| Gemini Flash / GPT-4o-mini / Claude Haiku | Speed + cost | 1M / 128K / 200K | $0.075–0.30 / $0.15–0.60 / $0.25–1.25 | High-throughput simple tasks |
Open-source Models (for private deployment / fine-tuning):
| Model | Parameters | Strongest Capability | Applicability |
|---|
| Llama 4 | 8B/70B/405B | General-purpose, strong ecosystem | English-primary |
| Qwen 3 | 7B/72B | Best for Chinese, multimodal | Top choice for Chinese scenarios |
| DeepSeek V3/R1 | 671B (MoE) | Reasoning, Chinese, extreme cost-performance | Cost-sensitive + strong reasoning |
| Mistral Large | 123B | Multilingual, speed | European market |
Phase 4: Prompt & Context Engineering
→ See full content in references/ai-industry-trends-2026.md
Phase 5: RAG Design & Implementation
→ See full content in references/ai-industry-trends-2026.md
Phase 6: Agents & Multi-Agent Systems
→ See full content in references/ai-industry-trends-2026.md
Phase 7: Model Fine-tuning & Adaptation
→ See full content in references/ai-industry-trends-2026.md
Phase 8: AI UX & Interaction Design
→ See full content in references/ai-industry-trends-2026.md
Phase 9: Evaluation Systems & Quality Assurance
→ See full content in references/ai-industry-trends-2026.md
Phase 10: Safety Guardrails & Red Teaming
→ See full content in references/ai-industry-trends-2026.md
Phase 11: AI Observability & Production Operations
→ See full content in references/ai-industry-trends-2026.md
Phase 12: AI Commercialization & Monetization
→ See full content in references/ai-industry-trends-2026.md
AI PM Workflow System
Dual-Track AI Product Development
Discovery (AI Capability Exploration):
Hypothesis → Prompt Prototype → Golden Dataset Evaluation → Alpha → Beta → A/B Validation → Launch
Delivery (AI Product Delivery):
Review → Prompt/Model Change → Offline Evaluation → Gradual Rollout (5%→25%→100%) → Monitoring → Iteration
AI PM Standard Weekly Cadence
| Day | Agenda |
|---|
| Monday | AI Metrics Review + Weekly Planning |
| Tuesday | User Research + Bad Case Deep Analysis |
| Wednesday | Prompt/RAG/Agent Design (Deep Work) |
| Thursday | Cross-team Alignment + Safety Review |
| Friday | Golden Dataset Maintenance + AI Knowledge Sharing |
AI Product Pre-Launch Checklist
□ Golden Dataset evaluation passed (core metrics not degraded)
□ Red team testing completed and high-risk items fixed
□ Safety guardrails deployed and tested
□ Cost model updated and reviewed
□ Monitoring alerts configured
□ Degradation/rollback plan prepared
□ Help documentation updated (users need to know how to interact with AI)
□ Gradual rollout plan confirmed
□ Legal/compliance signed off
Competency Model & Career Development
AI PM Capability Pyramid
┌────────────────────┐
│ AI Business │ ← AI Monetization / Token Economics / Market Judgment
│ Thinking (25%) │
├────────────────────┤
│ AI Technical │ ← Model Capabilities / RAG / Agent / Prompt / Evaluation
│ Literacy (30%) │
├────────────────────┤
│ AI Product │ ← AI UX / Trust Design / HITL / Interaction Patterns
│ Design (25%) │
├────────────────────┤
│ Product │ ← User Research / Requirements Analysis / Data Analysis
│ Fundamentals │
│ (20%) │
└────────────────────┘
From AI PM to Chief AI Officer
| Level | Experience | Core Competencies |
|---|
| Junior AI PM | 0-2 years | Prompt engineering basics, AI evaluation execution, AI feature PRD writing |
| Mid-level AI PM | 2-5 years | RAG/Agent solution design, Golden Dataset construction, AI UX design |
| Senior AI PM | 5-8 years | Model selection decisions, AI product strategy, safety system design, AI commercialization |
| AI Product Director | 8-12 years | AI product portfolio, Build/Buy decisions, AI team building |
| Chief AI Officer | 12+ years | Company AI strategy, AI governance, AI culture, AI investment portfolio |
AI PM Essential Technical Knowledge Checklist
Must Understand (able to have effective dialogue with ML engineers):
□ How LLMs work (Transformer / Attention Mechanism / Tokens)
□ Advanced Prompt Engineering (CoT / ReAct / Few-Shot)
□ RAG Architecture (Chunking / Retrieval / Reranking / Evaluation)
□ Agent Architecture (Tool Calling / Memory / Planning / HITL)
□ Model Evaluation Methods (Golden Dataset / LLM-as-Judge / A/B Testing)
□ Token Economics (Cost Estimation / Model Routing / Caching Strategy)
□ AI Security Basics (Injection / Jailbreak / Guardrails / Red Team Testing)
Bonus Items:
□ Fine-tuning Basics (SFT / RLHF / DPO / LoRA)
□ MLOps & AI Observability
□ GPU Economics & Inference Optimization
□ AI Governance & Compliance (EU AI Act / China Administrative Measures)
□ Multimodal AI Basics
Document Factory
AI Product Professional Documents
| Document | Audience | Detailed Template |
|---|
| AI Product PRD | Dev/ML Team | references/templates/ai-prd-template.md |
| AI Strategy Document | Management/Investors | references/templates/ai-strategy-template.md |
| RAG Design Document | ML/Backend Team | references/templates/rag-design-template.md |
| Agent Design Document | ML/Backend Team | references/templates/agent-design-template.md |
| Prompt Engineering Document | Product/ML Team | references/templates/prompt-engineering-template.md |
| AI Evaluation Plan | Product/QA/ML | references/templates/ai-evaluation-template.md |
| AI Security Plan | Security/Legal/ML | references/templates/ai-safety-template.md |
| AI Product Pricing Plan | Management/Finance | references/templates/ai-pricing-template.md |
| AI Competitive Analysis | Product/Marketing | references/templates/ai-competitive-template.md |
Diagram Factory
Must-Draw Diagrams for AI PMs
| # | Diagram Type | Purpose | Tool |
|---|
| 1 | RAG Architecture Diagram | RAG pipeline overview | drawio-skill |
| 2 | Agent Architecture Diagram | Agent / Multi-Agent system | drawio-skill |
| 3 | Model Routing Flowchart | Multi-model routing decisions | drawio-skill |
| 4 | AI Evaluation Pipeline Diagram | Evaluation process + data flow | drawio-skill |
| 5 | Safety Guardrail Layered Diagram | Multi-layer safety protection | drawio-skill |
| 6 | Data Flywheel Diagram | User → Data → AI Improvement Loop | excalidraw-diagram |
| 7 | AI Product Full-Stack Architecture Diagram | Product technical architecture | drawio-generator-pro |
Prototype Factory
"Generate an HTML prototype for an AI customer service chatbot"
→ Chat interface + Confidence display + Source citations + Human handoff + Empty state guidance
"Generate an HTML prototype for an AI contract review tool"
→ Upload contract → AI flags risky clauses → User confirms/modifies → Export report
"Generate an HTML prototype for an AI data analysis agent"
→ Natural language input → Agent thinking steps display → Visualized results → Download & share
AI Governance & Compliance
Global AI Regulatory Landscape
EU AI Act (Phased Implementation)
| Risk Level | Requirements | Product Examples |
|---|
| Unacceptable | Completely prohibited | Social credit scoring, real-time remote biometric identification |
| High Risk | Compliance assessment + human oversight + transparency + EU registry | Medical AI, recruitment AI, credit approval |
| Limited Risk | Inform users "you are interacting with AI" | Chatbots, AI-generated content |
| Minimal Risk | No additional obligations | AI filters, AI recommendations |
"Deployer Trap": Enterprises using third-party AI APIs may also bear obligations.
China AI Regulatory System
| Regulation | Core Requirements |
|---|
| Generative AI Service Administrative Measures | Safety assessment + algorithm filing + content moderation + training data compliance |
| Deep Synthesis Administrative Provisions | Synthetic content labeling + user real-name verification + review mechanism |
| Personal Information Protection Law | PII compliance in training data |
| Algorithm Recommendation Administrative Provisions | Algorithm filing + user right to know + opt-out mechanism |
China AI "Triple Registration": Algorithm Filing → AI Safety Assessment → Content Safety Review
AI Product Anti-Patterns Encyclopedia (25 Common Mistakes)
Strategy Category
| # | Anti-Pattern | Correct Approach |
|---|
| 1 | "Just Stuff AI In First" — AI for AI's sake | First ask whether AI truly solves the problem |
| 2 | Competing with OpenAI at the Model Layer | Build proprietary data and experience moats at the application layer |
| 3 | Ignoring the Data Flywheel | Design implicit feedback collection mechanisms from Day 1 |
| 4 | Pursuing SOTA Instead of Good Enough | Model routing: simple → small model, complex → large model |
| 5 | "AI Will Optimize Itself" | Establish a continuous loop of evaluation → analysis → optimization |
Technical Category
| # | Anti-Pattern | Correct Approach |
|---|
| 6 | Defaulting Everything to Agent | First evaluate with simple LLM calls, upgrade only if insufficient |
| 7 | Ignoring Token Costs | Monitor inference cost per interaction from Day 1 |
| 8 | Context Window Abuse | Curate context, don't pile everything in |
| 9 | RAG Only Using Vector Retrieval | BM25 + Vector + Reranker is the production baseline |
| 10 | Evaluation Set Made Up by PM | Build from real user query sampling |
UX Category
| # | Anti-Pattern | Correct Approach |
|---|
| 11 | Black-box AI — Not showing reasoning process | Show reasoning steps + cite sources |
| 12 | No Exit Design | Always keep a manual operation path |
| 13 | Pretending 100% Certainty | Display confidence levels when uncertain |
| 14 | AI Frequently Interrupting Users | Passive assistance, not proactive interruption |
| 15 | Ignoring Loading Experience | Streaming output + progress indicators + skeleton screens |
Security Category
| # | Anti-Pattern | Correct Approach |
|---|
| 16 | "Launch First, Security Later" | At minimum deploy basic input/output guardrails |
| 17 | Not Telling Users It's AI | Clearly label AI identity |
| 18 | Releasing Without Red Team Testing | At minimum internal red team testing before launch |
| 19 | Ignoring Low-Resource Language Security | Test jailbreak risks for all supported languages |
| 20 | No Emergency Circuit Breaker | One-click stop all AI features |
Commercialization Category
| # | Anti-Pattern | Correct Approach |
|---|
| 21 | Selling AI with Per-Seat Pricing | Hybrid model (base fee + usage) |
| 22 | Unlimited Free AI Usage | Set strict usage caps on free tier |
| 23 | Not Tracking User-Level Costs | Must be clear on each user's input-output ratio |
| 24 | Underestimating Price Wars | Token prices drop 10x annually, moat is data and experience |
| 25 | AI Gross Margin < 50% | Maintain 60%+ AI gross margin |
World-Class AI PM Framework Library
Shreyas Doshi's Product Sense (AI Era Edition)
In the AI era, Product Sense is the only irreplaceable capability.
Three Pillars:
1. Cognitive Empathy — See through the human needs behind users' irrational behavior
AI is "emotionally colorblind": understands data, but not human hearts
2. Aesthetics & Taste — Sharp intuition for whether an interaction feels "right"
AI can generate ten thousand solutions, but can't tell you which one gives users a visceral resonance
3. Business Intuition — Instantly see through the value exchange model behind anything
AI lacks the ability to make correct decisions in extreme business ambiguity
Chip Huyen's AI Engineering Framework
1. Evaluation-Driven Development
→ Build the evaluation set first, then write the Prompt. Evaluation is a navigator, not a checkpoint.
2. RAG is the Default Choice, Not the Last Resort
→ Most B2B AI products start with RAG.
3. Start Simple, Progressively Complex
Simple LLM → +RAG → +Tools → +Agent → +Multi-Agent
Don't skip levels! Verify each level before upgrading.
4. Systems Thinking, Not Model Thinking
Excellent AI Product = Model (30%) + Context Engineering (25%) + Evaluation (20%) + Safety (15%) + UX (10%)
Marily Nika's AI PM Three Pillars (Google)```
"All PMs will become AI PMs"
- Technical literacy is non-negotiable — you don't need to write code, but you must understand ML concepts
- From "translator" to "builder" — use AI tools yourself to prototype, test, and iterate
- Ethical AI literacy — bias detection, data privacy, content safety are fundamental skills for AI PMs
---
---
## Global AI Governance & Regulatory Landscape
### International AI Regulatory Comparison Matrix
| Region | Key Regulation | Risk Classification | Key Requirements | Enforcement Body | Penalties |
|--------|---------------|---------------------|------------------|------------------|-----------|
| **EU** | EU AI Act (2024) | 4 tiers: Unacceptable / High / Limited / Minimal | CE marking, conformity assessment, fundamental rights impact assessment | EU AI Office + National Authorities | Up to €35M or 7% global turnover |
| **US** | EO 14179 "Removing Barriers to American Leadership in AI" (2025-01) + NIST AI RMF 1.0 | Voluntary framework (4 functions: Govern, Map, Measure, Manage) | Federal shift from mandatory reporting (EO 14110, revoked) to deregulation + AI Action Plan; practical compliance surface now at state level (e.g., Colorado AI Act, Texas TRAIGA, California AI laws) | NIST + Sector-specific agencies + State AGs | Sector-specific + State enforcement |
| **UK** | AI Regulation White Paper (2023) + AI Safety Institute | Context-specific, principles-based (5 principles) | Cross-sectoral principles, no new regulator | Existing regulators (ICO, FCA, CMA, etc.) | Sector-specific |
| **China** | Generative AI Administrative Measures (2023) + Algorithm Filing | Content-based classification | Algorithm filing, security assessment, content moderation, training data compliance | Cyberspace Administration of China (CAC) | Service suspension, fines, criminal liability |
| **Japan** | AI Guidelines for Business (2024) | Non-binding, sector-specific | Voluntary AI governance framework, human-centric principles | METI + MIC | Sector-specific |
| **Singapore** | PDPC AI Governance Framework + AI Verify | Voluntary, risk-based | AI Verify testing toolkit, model governance framework | PDPC + IMDA | PDPA enforcement |
| **Canada** | AIDA (Bill C-27, proposed) | High-impact systems | Risk assessment, transparency, human oversight | Proposed AI & Data Commissioner | Up to CAD $25M or 5% revenue |
| **South Korea** | AI Basic Law (proposed 2024) | High-risk AI systems | Safety certification, human oversight, transparency | MSIT + KISA | Up to KRW 300M or 3% revenue |
| **UAE** | UAE AI Strategy 2031 + AI Ethics Guidelines | Sector-specific | AI ethics principles, sector-specific guidelines | UAE AI Office | Sector-specific |
> ⚠️ **Corrected 2026-07 (current version prevails)**: The original table entry "AI Executive Order 14110 (2023)" was factually incorrect — **EO 14110 was revoked by EO 14179 on Jan 23, 2025**. EO 14179 shifts to a deregulatory stance and directs development of an "AI Action Plan" (released in 2025). Consequently, there is no unified federal mandatory framework akin to the EU AI Act; practical compliance responsibility shifts down to **state-level legislation**:
> - **Colorado AI Act** (SB 24-205, signed May 2024, effective Feb 2026) — the first U.S. state law targeting "high-risk AI systems";
> - **Texas Responsible AI Governance Act (TRAIGA)** (effective 2025);
> - **California**: multiple AI bills advanced in the 2024–2025 legislative cycle (including successors to the vetoed SB 1047).
> Practitioners should follow the latest legislation of the **state where the product is sold / service is provided**.
>
> > **Corrected 2026-07: EO 14110 was revoked by EO 14179 on Jan 23, 2025.** EO 14179 shifted to a deregulation stance and directed the formulation of an "AI Action Plan" (to be issued within 2025). As a result, there is no longer a unified mandatory federal framework comparable to the EU AI Act; practical compliance has shifted to **state-level legislation**:
> > - **Colorado AI Act** (SB 24-205, signed May 2024, effective Feb 2026) — the first state-level act targeting "high-risk AI systems" in the US;
> > - **Texas Responsible AI Governance Act (TRAIGA)** (effective 2025);
> > - **California**: multiple AI bills advanced through the 2024–2025 legislative cycle (including successor variants to the vetoed SB 1047).
> > Practitioners should refer to the latest legislation in the **state where the product is sold or served**.
### NIST AI RMF 1.0 — Four Functions Deep Dive
The NIST AI Risk Management Framework (AI RMF 1.0) provides a structured approach to AI risk management:
| Function | Core Activities | AI PM's Role |
|----------|----------------|-------------|
| **GOVERN** | Establish AI risk culture, policies, accountability | Define AI governance structure, risk appetite, escalation paths |
| **MAP** | Understand AI system context, identify risks | Map AI system components, data flows, stakeholders, potential harms |
| **MEASURE** | Assess AI risks using quantitative/qualitative methods | Design evaluation metrics, monitor trustworthiness characteristics |
| **MANAGE** | Prioritize and respond to AI risks | Implement guardrails, document risk decisions, continuous monitoring |
### California SB 1047 — Safe and Secure Innovation for Frontier AI Models
> ⚠️ **Corrected 2026-07 (current version prevails)**: SB 1047 was **vetoed by the California Governor in Sep 2024 and never became law**. The points below are its original proposal highlights, provided only as conceptual reference for "frontier model safety," and should not be treated as current compliance requirements.
Key provisions relevant to AI PMs (as originally proposed):
- **Safety testing**: Mandatory safety testing for models above compute threshold (10^26 FLOPS)
- **Kill switch**: Full shutdown capability required
- **Third-party auditing**: Annual third-party safety audits
- **Liability**: Developer liability for catastrophic harms
- **Whistleblower protection**: Employee protection for reporting safety concerns
### UK AI Safety Institute (AISI)
The UK's approach focuses on frontier AI safety evaluation:
- **Pre-deployment testing**: Voluntary but strongly encouraged for frontier models
- **Evaluation domains**: Cybersecurity, bio/chemical capabilities, autonomous systems, societal impacts
- **International collaboration**: Bletchley Declaration, Seoul Summit, AI Safety Summits
- **PM implications**: Prepare for pre-deployment evaluation requirements, document safety testing results
### APAC AI Governance Deep Dive
| Country | Regulation/Policy | Status | Key Requirements | PM Implications |
|---------|------------------|--------|------------------|-----------------|
| **Japan** | AI Promotion Act (2025) | Enacted June 2025 | Non-binding framework; R&D promotion; transparency goals; no criminal penalties | Light-touch compliance; focus on manufacturing/healthcare/robotics AI |
| **Singapore** | PDPC AI Governance Framework + AI Verify | Active (v2.0, 2024) | AI Verify testing toolkit; model governance; financial sector GenAI guidelines | Use AI Verify for product testing; financial services AI has extra requirements |
| **South Korea** | AI Basic Law (2024) | Enacted | High-risk AI safety certification; human oversight; transparency; KRW 300M penalty | Safety certification for high-risk products; mandatory human-in-the-loop |
| **India** | DPDP Act 2023 + National AI Strategy | Active | Data protection; emerging AI framework; no comprehensive AI law yet | Data localization considerations; large population = diverse training data needs |
| **Australia** | Voluntary AI Safety Standards + National AI Plan | Active | Ethical AI in government; voluntary standards; sector-specific guidance | Government AI procurement standards; ethical AI principles alignment |
### Middle East AI Governance
| Country | Regulation/Policy | Status | Key Requirements | PM Implications |
|---------|------------------|--------|------------------|-----------------|
| **UAE** | UAE AI Strategy 2031 + AI Ethics Guidelines | Active | Sector-specific AI ethics; government AI adoption targets; AI Office oversight | Sovereign AI focus; government procurement opportunities; bilingual (AR/EN) requirements |
| **Saudi Arabia** | SDAIA AI Ethics Principles (2023) + Vision 2030 | Active | AI ethics governance; data classification; national AI strategy | Government-driven AI adoption; large-scale smart city projects (NEOM) |
| **Qatar** | National AI Strategy (2024) | Active | AI R&D investment; education sector AI; data governance | Growing AI hub; education and healthcare AI focus |
### LATAM AI Governance
| Country | Regulation/Policy | Status | Key Requirements | PM Implications |
|---------|------------------|--------|------------------|-----------------|
| **Brazil** | AI Bill (PL 2338/2023) + AI Ethics Guidelines | Under review | Risk-based classification; transparency; user rights; algorithmic impact assessment | Modeled on EU approach; LGPD data protection alignment; growing fintech AI |
| **Mexico** | National AI Strategy (2024) | Active | AI for social good; government AI adoption; ethics framework | Government-led AI adoption; nearshoring opportunities for US companies |
| **Argentina** | AI for Social Good Guidelines | Active | AI ethics; social impact focus; public sector AI | Public sector AI projects; growing AI talent pool |
### Africa AI Governance
| Country | Regulation/Policy | Status | Key Requirements | PM Implications |
|---------|------------------|--------|------------------|-----------------|
| **South Africa** | National AI Strategy (2024 draft) | Draft | AI for development; ethical AI; skills development | Leading African AI hub; fintech and healthtech AI focus |
| **Kenya** | National AI Strategy (2025) | Active | AI for agriculture/healthcare/fintech; data protection; innovation hubs | Mobile-first AI products; M-Pesa fintech ecosystem; growing startup scene |
| **Nigeria** | National AI Policy (2024) | Active | AI for economic diversification; local language AI; skills development | Largest African market; local language AI (Yoruba, Hausa, Igbo); fintech AI |
| **African Union** | AU AI Continental Strategy (2024) | Active | Pan-African AI harmonization; capacity building; data sovereignty | Regional harmonization efforts; infrastructure challenges = edge AI opportunities |
---
## Global AI PM Career Landscape
### International AI PM Salary Benchmarks (2025)
| Region | Entry-Level (0-3 yrs) | Mid-Level (3-7 yrs) | Senior (7-12 yrs) | Director/VP (12+ yrs) |
|--------|----------------------|---------------------|-------------------|----------------------|
| **US (SF/NYC)** | $140K-180K | $200K-280K | $300K-450K | $500K-800K+ |
| **US (Other Tech Hubs)** | $120K-150K | $170K-230K | $250K-350K | $400K-600K |
| **Canada (Toronto/Vancouver)** | CAD $100K-130K | CAD $140K-180K | CAD $200K-280K | CAD $300K-450K |
| **UK (London)** | £70K-90K | £100K-140K | £150K-200K | £200K-350K |
| **Germany (Berlin/Munich)** | €65K-85K | €90K-120K | €130K-170K | €180K-250K |
| **Netherlands** | €60K-80K | €85K-115K | €120K-160K | €170K-230K |
| **Singapore** | SGD $90K-120K | SGD $130K-180K | SGD $200K-280K | SGD $300K-450K |
| **Japan (Tokyo)** | ¥8M-12M | ¥13M-18M | ¥20M-28M | ¥30M-50M |
| **UAE (Dubai/Abu Dhabi)** | AED 300K-400K | AED 450K-600K | AED 650K-900K | AED 1M-1.5M+ |
| **India (Bangalore)** | ₹15L-25L | ₹30L-50L | ₹60L-1Cr | ₹1.2Cr-2.5Cr |
| **Australia (Sydney)** | AUD $120K-150K | AUD $160K-200K | AUD $220K-300K | AUD $350K-500K |
### Big Tech AI PM Career Ladders
| Level | Google | Meta | Microsoft | Amazon |
|-------|--------|------|-----------|--------|
| **Entry** | APM / L3 | RPM | PM1 | PM (L5) |
| **Mid** | PM L4-L5 | PM IC4-IC5 | PM2 | Sr PM (L6) |
| **Senior** | Sr PM L6 | PM IC6 | Principal PM | Principal PM (L7) |
| **Staff** | Staff PM L7 | PM IC7 | Partner PM | Sr Principal (L8) |
| **Director** | Director L8 | Director M1 | Director | Director (L8+) |
| **VP** | VP L9+ | VP M2+ | VP | VP |
### International AI PM Certifications
| Certification | Issuing Body | Focus | Global Recognition |
|--------------|-------------|-------|-------------------|
| **IAPP AIGP** | IAPP | AI Governance Professional | High (US/EU) |
| **ISO/IEC 42001 Lead Implementer** | Various | AI Management System | High (Global) |
| **AI Product Manager Certificate** | Product School | AI PM Skills | Medium (US) |
| **AI for Product Management** | Reforge | AI PM Strategy | Medium (US) |
| **Google AI PM Certificate** | Google | AI PM Foundations | Medium (Global) |
| **Duke AI Product Manager** | Duke University | AI PM Certificate | Medium (US) |
---
## Global AI Startup Ecosystem
### Major AI Hubs by Region
| Region | Key Hubs | Notable AI Startups | Key Strengths |
|--------|---------|---------------------|---------------|
| **US** | SF Bay Area, NYC, Boston, Seattle | OpenAI, Anthropic, Scale AI, Databricks | Deep tech, VC funding, talent density |
| **Canada** | Toronto, Montreal, Vancouver | Cohere, Element AI, Waabi | NLP research, government funding, talent pipeline |
| **UK** | London, Cambridge, Edinburgh | DeepMind, Stability AI, Graphcore | Research excellence, AI safety focus |
| **EU** | Paris, Berlin, Amsterdam, Stockholm | Mistral AI, Aleph Alpha, Helsing | Open-source models, defense AI, regulation |
| **Israel** | Tel Aviv | AI21 Labs, Gong, Run:ai | Enterprise AI, cybersecurity, NLP |
| **China** | Beijing, Shanghai, Shenzhen, Hangzhou | Zhipu AI, Moonshot AI, DeepSeek, MiniMax, Baichuan | Scale, mobile-first, government support |
| **Singapore** | Singapore | PatSnap, Advance.AI, Taiger | SEA hub, multilingual, fintech AI |
| **India** | Bangalore, Hyderabad, Mumbai | Krutrim, Sarvam AI, Haptik | Indic languages, enterprise, cost efficiency |
| **UAE** | Dubai, Abu Dhabi | G42, Falcon AI, AI71 | Sovereign AI, government investment, compute |
| **South Korea** | Seoul | Upstage, Riiid, Liner | AI education, enterprise AI, hardware |
### Global AI Funding by Region (2024-2025)
| Region | Total AI VC Funding | YoY Growth | Top Segments |
|--------|--------------------|------------|--------------|
| **North America** | ~$80B | +35% | Foundation models, AI infrastructure, enterprise AI |
| **China** | ~$15B | +20% | LLM applications, autonomous driving, AI chips |
| **Europe** | ~$12B | +30% | AI safety, enterprise AI, open-source models |
| **Middle East** | ~$5B | +50% | Sovereign AI, smart city, fintech AI |
| **APAC (ex-China)** | ~$8B | +25% | Enterprise AI, AI education, fintech AI |
| **LATAM** | ~$2B | +40% | Fintech AI, agritech AI, healthtech AI |
| **Africa** | ~$0.5B | +60% | Fintech AI, agritech AI, healthtech AI |
---
## Global AI Adoption & Market Data
### Enterprise AI Adoption by Region (2025)
| Region | AI Adoption Rate | GenAI in Production | Top Use Cases |
|--------|-----------------|--------------------|---------------|
| **North America** | 72% | 34% | Customer service, content generation, code generation |
| **Europe** | 58% | 22% | Process automation, compliance, customer insights |
| **APAC** | 65% | 28% | Customer engagement, operations, product development |
| **China** | 70% | 30% | Content generation, customer service, enterprise search |
| **Middle East** | 55% | 18% | Smart city, government services, fintech |
| **LATAM** | 45% | 12% | Customer service, marketing, fintech |
| **Africa** | 35% | 8% | Fintech, agriculture, healthcare |
### Global AI Spending Forecast (Gartner, 2025)
| Year | Global AI Spending | YoY Growth | Key Drivers |
|------|-------------------|------------|-------------|
| 2024 | ~$200B | +25% | GenAI adoption, LLM infrastructure |
| 2025 | ~$300B | +50% | Agent AI, enterprise AI platforms |
| 2026 | ~$450B | +50% | AI-native applications, autonomous agents |
| 2027 | ~$650B | +44% | AI-first enterprises, AI hardware |
| 2028 | ~$900B | +38% | Ubiquitous AI, AI-driven business models |
## Final Reminders
> **The Ultimate AI PM Mindset (12 Iron Rules):**
>
> 1. **AI is a tool, not magic** — Start from narrow scenarios, build moats with data
> 2. **Models will commoditize, data and experience won't** — The moat is proprietary data and unique experience
> 3. **Evaluation is the foundation of everything** — Without evaluation, there is no AI product iteration
> 4. **Trust is the currency of AI products** — Lose it once, it takes 10 perfect performances to win back
> 5. **Token cost = your COGS** — A product that doesn't track costs is blind
> 6. **Safety is not a feature, it's infrastructure** — Design safety on Day 1, not as an afterthought
> 7. **Progressive trust > one-shot release** — Start with low-risk features
> 8. **Humans always stay in the loop** — Don't let AI make decisions it can't be held accountable for
> 9. **Simple > flashy** — A precise classification model can be more valuable than a hallucinating Agent
> 10. **Prompts are the UI of AI products** — Treat prompts well, iterate on them like product interfaces
> 11. **Data flywheel > model capability** — Proprietary data accumulation doesn't happen automatically
> 12. **You are the CEO of your AI product** — Not a "model API wrapper"; take full responsibility for the user's AI experience
---
> **Get started: Tell me directly what you want to do, and the Skill will automatically match the phase, methodology, and toolchain.**
> Whether you're a traditional PM transitioning to AI or new to the AI field, this Skill is your AI PM super-assistant.
---
## Tool Integration Summary
| Task | Primary Tool | Alternative |
|------|-------------|-------------|
| Draw AI architecture diagrams | `drawio-skill` | `drawio-coderknock` |
| Draw data flywheel / journey maps | `excalidraw-diagram` | - |
| Generate AI product prototypes | Generate HTML directly (Tailwind+Alpine.js) | - |
| Write AI PRD / strategy docs | Generate Markdown directly | `word-docx` |
| Make AI strategy PPTs | `ppt-generator` | `ppt-master` |
| Build evaluation datasets | `xlsx` | - |
| Prompt version comparison | Git (Markdown) | - |
---
## Usage Examples
### Example 1: Design an AI Customer Service Product from 0 to 1
User: Help me design an AI customer service chatbot product
Output:
- Phase 1: AI Opportunity Assessment (CS scenario AI feasibility + D×V×F scoring)
- Phase 3: Model Selection (Recommend GPT-4o/Gemini Flash tiered routing)
- Phase 4: System Prompt + Few-Shot example design
- Phase 5: RAG Design (knowledge base chunking + hybrid retrieval strategy)
- Phase 8: AI UX Prototype (conversation interface + confidence + handoff to human)
- Phase 9: Evaluation Plan (Golden Dataset + RAGAS metrics)
- Phase 10: Safety Guardrails (jailbreak protection + content moderation)
- Phase 12: Pricing Plan (base monthly fee + per-conversation overage billing)
### Example 2: Add AI Contract Review to Existing CRM
User: Add AI contract review functionality to my CRM
Output:
- Phase 1: AI Feasibility Assessment (task characteristics analysis for contract review)
- Phase 3: Model Selection (Recommend Claude Sonnet + long context window)
- Phase 4: Prompt Design (legal expert System Prompt + clause review instructions)
- Phase 8: AI UX Design (upload → AI flags risks → human confirms → generate review report)
- Phase 9: Evaluation Plan (contract review accuracy Golden Dataset)
- Phase 10: HITL Design (high-risk clauses: AI suggestion → legal confirmation)
---
> **Get started: Tell me directly what you want to do, and the AI PM Skill will automatically match the phase, methodology, and toolchain.**
---
## Appendix A: Embedding Model Selection Deep Dive
> Embeddings are the cornerstone of RAG. Choose the wrong embedding model, and even the best retrieval strategy is wasted.
### Mainstream Embedding Model Comparison
| Model | Dimensions | Max Input | MTEB Chinese | Cost/1M tokens | Best Scenario |
|------|-----------|----------|-------------|---------------|---------|
| text-embedding-3-large (OpenAI) | 3072/256/1024 | 8191 | Medium | $0.13 | English, multilingual |
| text-embedding-3-small | 1536/512 | 8191 | Low-Medium | $0.02 | Cost-sensitive English |
| bge-large-zh-v1.5 (BAAI) | 1024 | 512 | ⭐⭐⭐⭐⭐ | Open-source free | Top choice for Chinese |
| bge-m3 (BAAI) | 1024 | 8192 | ⭐⭐⭐⭐⭐ | Open-source free | Multilingual + long documents |
| stella-base-zh-v3-1792d | 1792 | 512 | ⭐⭐⭐⭐ | Open-source free | High-precision Chinese retrieval |
| multilingual-e5-large | 1024 | 512 | ⭐⭐⭐⭐ | Open-source free | Mixed multilingual |
| jina-embeddings-v3 | 1024 | 8192 | ⭐⭐⭐⭐ | Paid API | Long documents + multilingual |
### Selection Decision Tree
-
Primary language?
├── Mainly Chinese → bge-large-zh-v1.5 or stella-base-zh
├── Mainly English → text-embedding-3-large
├── Mixed multilingual → bge-m3 or multilingual-e5-large
└── Need on-premise deployment → bge series (open-source)
-
Document length?
├── Short (<512 tokens) → bge-large-zh-v1.5
├── Long (512-8192) → bge-m3 or jina-embeddings-v3
└── Extra-long (>8192) → Chunk first, then embed
-
Dimension preference?
├── Accuracy-first (>1024 dims) → text-embedding-3-large (3072) or stella (1792)
├── Speed-first (768 dims) → bge-base-zh-v1.5
└── Storage-first (<512 dims) → text-embedding-3-small (512)
### Embedding Quality Checklist
□ Synonym recall test: Can "contract expiration" retrieve "agreement termination"?
□ Polysemy distinction test: Is "apple" correctly distinguished in tech vs. fruit contexts?
□ Negation semantic test: Do retrieval results for "does not include XX" contain XX?
□ Cross-language test: Can Chinese queries retrieve English documents? (if needed)
□ Long document test: Does embedding quality degrade for 512+ token documents?
□ Domain terminology test: How effective is retrieval for industry-specific terms?
---
## Appendix B: Vector Database Selection Decision Matrix
### Mainstream Vector Database Comparison
| Database | Type | Scale Limit | Performance | Ops Complexity | Best Scenario |
|--------|------|---------|------|-----------|---------|
| **pgvector** | PostgreSQL Extension | <1M vectors | Medium | Low | Already have PG, modest vector volume |
| **Milvus** | Dedicated Vector DB | >1B | ⭐⭐⭐⭐⭐ | High | Large-scale production |
| **Qdrant** | Dedicated Vector DB | >100M | ⭐⭐⭐⭐ | Medium | Medium scale, good performance |
| **Weaviate** | Dedicated Vector DB | >100M | ⭐⭐⭐⭐ | Medium | Need built-in modules (text/image) |
| **Pinecone** | Cloud Service | >1B | ⭐⭐⭐⭐ | Very Low | Don't want to self-manage |
| **Chroma** | Lightweight Embedded | <100K | Low | Very Low | Prototype/dev environment |
| **ElasticSearch** | Search Engine + Vector | >100M | ⭐⭐⭐ | Medium-High | Already have ES infrastructure |
### Selection Decision
Vector count < 100K + rapid prototyping → Chroma (zero ops)
Vector count 100K-1M + already have PG → pgvector (zero additional cost)
Vector count 1M-100M + small/medium team → Qdrant or Weaviate
Vector count > 100M + enterprise-grade → Milvus (distributed scaling)
Don't want to self-manage + sufficient budget → Pinecone
Already have ES + need hybrid search → ElasticSearch
### Key Production Considerations
| Consideration | Key Question |
|--------|---------|
| High Availability | Supports master-slave replication? Failover time? |
| Backup & Recovery | Incremental backup? Full restore time? |
| Multi-tenancy | Partition/Collection-level isolation? RBAC? |
| Hybrid Query | Vector + scalar filter performance? Scalar index? |
| Cost | Storage cost/TB? Query cost/1M queries? |
---
## Appendix C: RAG Chunking Strategy Deep Dive
### Chunking Method Comparison
| Chunking Method | Principle | Pros | Cons | Best Scenario |
|---------|------|------|------|---------|
| **Fixed Size** | Split by token count (e.g., 512 tokens) | Simple, controllable, predictable | Cuts off semantics | Initial solution, uniform documents |
| **By Paragraph** | Split by natural paragraphs | Semantically complete | Uneven length | Well-structured documents |
| **By Heading Hierarchy** | Split by H1/H2/H3 levels | Preserves context hierarchy | Complex implementation | Technical docs/help docs |
| **Semantic Chunking** | Use LLM to determine optimal split points | Semantically optimal | High cost, slow | High-quality requirement scenarios |
| **Hybrid Chunking** | Paragraph + fixed size fallback | Balances semantics and control | Complex rules | Recommended for production |
| **Sentence Window** | Split by sentences + window context | Fine-grained retrieval | Large storage | Scenarios needing precise positioning |
### Chunk Size Selection Guide
Chunk size 256 tokens:
→ Pros: Precise retrieval, low latency
→ Cons: Incomplete context, easy semantic cutoff
→ Suitable for: FAQ, short answers
Chunk size 512 tokens:
→ Pros: Balanced precision and completeness ← Recommended starting point
→ Suitable for: Most RAG scenarios
Chunk size 1024 tokens:
→ Pros: More complete context
→ Cons: Retrieval may bring back irrelevant content
→ Suitable for: Complex documents, technical manuals
Chunk size 2048+ tokens:
→ Pros: Long paragraphs not cut off
→ Cons: Low signal-to-noise ratio
→ Suitable for: Legal documents, academic papers
### Chunking Optimization in Practice
Chunking optimization loop:
- Run retrieval with current chunking strategy
- Collect Bad Cases (retrieved irrelevant / failed to retrieve relevant)
- Analyze root causes:
- Semantics cut off → Increase chunk size or add overlap
- Too much noise → Reduce chunk size or add Reranker
- Heading info lost → Prepend document title to each chunk
- Adjust strategy → Re-evaluate → Compare RAGAS metrics
- Repeat until satisfied
Key techniques:
- Add metadata prefix to each chunk: [Document Title] | [Section] | [Update Date]
- Small Chunk + Large Context Window: Retrieve with small chunks + bring back adjacent chunks
- Hierarchical Indexing: First retrieve document → then retrieve specific passages within that document
---
## Appendix D: Agent Tool Design Patterns & Practices
### 5 Principles of Tool Design
-
Single Responsibility: One tool does one thing
❌ "query_database" — too vague
✅ "query_order_info" / "query_member_points" — specific
-
Self-Describing: Tool name + description lets LLM understand at a glance
✅ Tool name: cancel_order
✅ Description: Cancel a specified order, requires order number and cancellation reason, returns cancellation status
-
Verifiable Output: Return structured results so LLM can determine success
✅ {status: "success", data: {...}, error: null}
❌ "The order seems to have been cancelled" (LLM has to guess)
-
Idempotency: Same input, multiple calls produce the same result (especially for mutation tools)
✅ Create order: use idempotency_key to prevent duplicates
✅ Cancel order: already-cancelled orders return "already cancelled" rather than error
-
Error-Friendly: Error messages must have enough context for LLM to decide next steps
✅ {error: "Order ORD123 not found, possibly wrong order number or order doesn't belong to current user"}
❌ {error: "not found"}
### Tool Count Management
| Agent Type | Recommended Tool Count | Rationale |
|-----------|----------|------|
| Simple Agent | 3-5 | Reduce confusion |
| Standard Agent | 5-10 | Cover core capabilities |
| Complex Agent | 10-20 | Multi-domain |
| Super Agent | 20-50 | Requires tool grouping + dynamic loading |
**What if there are too many tools?**
- Tool grouping: Group by domain (order-related, member-related, knowledge base-related)
- Dynamic loading: Agent first determines intent → only loads relevant tool groups
- Tool naming convention: `domain_action_object` (e.g., `order_query_status`)
### Common Tool Calling Issues
| Issue | Manifestation | Solution |
|------|------|------|
| Hallucinated calls | Calls non-existent tools | Add "only use provided tools" in tool description |
| Parameter hallucination | Fabricates parameter values | Require Agent to extract parameters from user input / prior results |
| Loop calls | Repeatedly calls the same tool | Set max retry count (3 times) |
| Premature abandonment | Gives up after one failure | Include retry suggestions in tool response |
| Permission errors | Calls tools without permission | Include permission requirements in tool definition |
---
## Appendix E: AI Team Structure & Roles
### Typical AI Product Team Configuration
AI Product Team (10-30 people):
├── AI Product Manager (1-2 people)
│ └── Responsible for: AI strategy / model selection / prompt design / evaluation system / AI UX
│
├── ML Engineers (2-4 people)
│ ├── Responsible for: model fine-tuning / evaluation pipeline / model routing / RAG implementation
│ └── PM needs to align on: model capability boundaries, fine-tuning data needs, evaluation metrics
│
├── Backend Engineers (2-4 people)
│ ├── Responsible for: API / vector database / Agent tools / knowledge base ingestion pipeline
│ └── PM needs to align on: API design, data models, performance requirements
│
├── AI Safety Engineer (1 person)
│ ├── Responsible for: guardrails / red team testing / content moderation / compliance
│ └── PM needs to align on: risk matrix, safety release standards
│
├── Prompt Engineer / Content Designer (1 person)
│ ├── Responsible for: System Prompt / Few-Shot library / output quality optimization
│ └── PM needs to align on: brand tone, content standards, user feedback
│
├── Data Annotation / Evaluation Specialist (1-2 people)
│ ├── Responsible for: Golden Dataset maintenance / human evaluation / Bad Case analysis
│ └── PM needs to align on: evaluation standards, annotation specifications, data quality
│
└── AI UX Designer (1 person)
├── Responsible for: AI interaction patterns / trust design / prototypes / user research
└── PM needs to align on: interaction principles, trust-building roadmap
### Effective PM-ML Engineer Collaboration
What PMs should NOT do:
❌ "This model isn't good enough, can you make it more accurate?" (too vague)
❌ "I read a paper saying XX model has the highest SOTA, let's use that" (benchmark-only thinking)
❌ "Why can't it answer this? You're an AI!" (not understanding capability boundaries)
What PMs SHOULD do:
✅ "In this 50-item test set, the model's accuracy on refund policy questions is 70%,
the main failure modes are... I suggest optimizing from XX direction" (specific + data + suggestions)
✅ "For this scenario our latency budget is only 500ms,
which of these three models can meet that? What accuracy can we achieve?" (constraints + trade-offs)
✅ "User feedback says AI responses are too verbose, here are 5 Bad Cases,
can we adjust the Prompt or do we need fine-tuning?" (problem + evidence + options)
---
## Appendix F: AI Product Development Cadence & Milestones
### 6 Milestones for AI Product 0→1
M1: Problem Validation (1-2 weeks)
Goal: Confirm AI is the right solution
Output: AI Opportunity Assessment (including scoring matrix)
Check: Score ≥ 3.5? Data accessible?
M2: Prompt Prototype (2-4 weeks)
Goal: Prove feasibility with prompts
Output: System Prompt V1 + 50 test results
Check: Core scenario accuracy > 70%?
M3: RAG MVP (4-8 weeks)
Goal: Minimum closed loop of knowledge base retrieval + generation
Output: Basic RAG pipeline + 100-item evaluation baseline
Check: RAGAS Faithfulness > 0.80?
M4: Alpha Internal Testing (2-4 weeks)
Goal: Internal team daily use
Output: Alpha usage feedback + Bad Case logs
Check: Internal tester satisfaction > 3.5/5?
M5: Beta External Testing (4-8 weeks)
Goal: Friendly customers real-world use
Output: Beta evaluation report + improvement checklist
Check: Willingness to pay > 50%? NPS > 30?
M6: GA Official Launch (2-4 weeks)
Goal: Public commercial availability
Output: Product launch + safety guardrails ready + monitoring ready
Check: All safety checklist items passed?
### AI Product Iteration Cadence
| Iteration Type | Frequency | Content |
|---------|------|------|
| Prompt Optimization | Weekly | Fine-tune prompts based on Bad Cases |
| RAG Optimization | Bi-weekly | Chunking strategy / retrieval parameter tuning |
| Model Upgrade Evaluation | Quarterly | New model evaluation + migration assessment |
| Fine-tuning Iteration | Quarterly | Update fine-tuning with new annotated data |
| Safety Review | Monthly | Safety metrics review + red team spot checks |
| Golden Dataset Update | Monthly | Add new queries, retire outdated queries |
---
## Appendix G: AI Cost Optimization Practical Handbook
### 4 Phases of Cost Optimization
Phase 1: Monitoring Transparency (Day 1)
□ Log token consumption for every AI call
□ Cost breakdown by feature / by user / by model
□ Free user vs. paid user cost comparison
□ Build a cost dashboard
Phase 2: Low-Hanging Fruit (Month 1)
□ Prompt streamlining (remove redundant instructions) — save 10-20%
□ Output token limits (max_tokens) — save 5-15%
□ Identical query result caching — save 20-40%
□ Common FAQ pre-generation + caching — save 30-50%
Phase 3: Architecture Optimization (Months 2-3)
□ Model routing (simple → small model) — save 40-60%
□ Semantic caching (similar query reuse) — save 20-40%
□ Batch processing (non-real-time scenarios) — save 20-50%
Phase 4: Deep Optimization (Months 4-6)
□ Fine-tune small models to replace large models — save 50-80%
□ Custom inference infrastructure — save 30-60%
□ On-premise deployment of open-source models — TCO optimization
### Cost Anomaly Detection Rules
Alert trigger conditions:
□ Single user daily cost > 200% of yesterday
□ Single user daily cost > overall P99
□ Free user daily cost > $1
□ Large model share in model routing > 40% (check if routing is broken)
□ Cache hit rate < 30% (check if cache is broken)
□ AI gross margin < 50% (check cost structure)
---
## Appendix H: AI PM Daily Practical Scenarios
### Scenario 1: Prompt Debugging
PM daily workflow:
- Discover Bad Case (user complaint / evaluation failure / self-discovery)
- Analyze: Is it a prompt problem? Retrieval problem? Model capability problem?
- If it's a prompt problem → pinpoint which specific sentence is problematic
- Modify prompt → test on the Bad Case
- Regression test on the full Golden Dataset
- Evaluation comparison report → pass → gradual rollout
Debugging tips:
- Add Few-Shot first (fastest results)
- Then modify constraint statements ("You must..." / "You must not...")
- Finally modify role description (least impact)
- Change only one thing at a time, A/B compare
### Scenario 2: Model Upgrade Evaluation
When upstream models upgrade (e.g., Claude Sonnet 3.5→4):
- Run old vs. new model comparison on existing Golden Dataset
- Focus on:
- Which scenarios improved? (don't need attention)
- Which scenarios regressed? (need focused attention!)
- Cost changes (new models are usually more expensive)
- Latency changes
- Deep analysis of regression scenarios → may need prompt adjustments
- Cost-benefit analysis → decide whether to migrate
- If migrating: 5%→25%→100% gradual rollout
### Scenario 3: User Says "AI Isn't Smart Enough"
User's "not smart enough" could mean:
├── Didn't understand their intent → Query rewriting / clarification mechanism
├── Response too generic → Retrieval not precise enough / missing user profile
├── Response incorrect → Knowledge base outdated / hallucination / insufficient model capability
├── Wrong format → Prompt output format constraints insufficient
├── Response too slow → Model too large / no streaming output
├── Missing citations → RAG didn't return sources
├── Unnatural tone → Prompt role description needs adjustment
└── Didn't remember context → Conversation history management issue
PM's diagnostic process:
- Look at actual conversation logs (don't just rely on user reports)
- Reproduce the issue yourself
- Pinpoint the specific link in the chain
- Apply targeted remedy, not generic "optimize AI"
---
## Appendix I: AI Product Technology Radar
### AI Technology Trends Worth Watching in 2025-2026
| Technology | Maturity | Impact on PMs | Action Recommendation |
|------|--------|-----------|---------|
| **Long Context Windows (1M+)** | Available | May simplify RAG design | Evaluate ROI of long context vs. RAG |
| **Multimodal (Vision + Voice)** | Rapidly maturing | Expands product scenarios | Think about scenarios multimodal can unlock |
| **AI Agent Standardization** | Early stage | Agents will become as common as APIs | Watch MCP/A2A and other Agent protocols |
| **Small Language Models (SLM) Boom** | In progress | On-device + low-cost inference | Evaluate feasibility of small models in specific scenarios |
| **AI Code Generation** | Mainstream | Changes software development | Use AI tools to accelerate prototype validation |
| **Real-time AI (Voice/Video)** | Rapidly maturing | Voice Agents become new entry points | Evaluate applicable scenarios for voice interaction |
| **AI Safety Automation** | Early stage | Reduces safety ops costs | Watch automated red team testing tools |
| **RAG 2.0** | Early stage | GraphRAG/Agentic RAG maturing | Try new patterns in complex scenarios |
### Impact of Technology Trends on Product Strategy
Trend 1: Accelerating model commoditization
→ The moat is not in model choice, but in data flywheel and user experience
→ Design model-switchable architecture (don't lock into one vendor)
Trend 2: Rapidly declining inference costs
→ AI can unlock more and more scenarios
→ But also means competitors can quickly replicate your AI features
→ Differentiation lies in proprietary data and deep integration
Trend 3: Agents moving from demo to production
→ 2025-2026 is the critical year for Agents from experimentation to production
→ Agent reliability, safety, and cost remain major challenges
→ PMs need to think: Do users really need an Agent? Or just a better UI?
Trend 4: AI shifting from "assisting" to "doing"
→ AI from assistive tool → autonomously completing work
→ Sequoia: Service-as-a-Software
→ PM's reflection: Does your product help users do, or do it for users?
Appendix J: 2025-2026 Leading Model Capability Matrix
