## 📖 Quick Navigation - [What is this guide about?](#what-is-this-guide) - [TL;DR - 30 Second Summary](#tldr) - [Full Table of Contents](#full-table-of-contents) - [Start Learning](#the-landscape) --- ## What is this guide about? **This is the guide between knowing how AI works and knowing which AI to use.** You now understand how Generative AI works under the hood, what it can and cannot do, how to think with it, how to prompt it effectively, how it fits into daily work, and how to use it responsibly. What you may not have is a clear, honest map of the actual tools available — and the judgment to choose between them. **Which AI tool should you use — and for what?** The AI tools landscape is large, fast-moving, and full of marketing noise. This guide cuts through it. No hype. No brand loyalty. A clear, practical framework for understanding what the major platforms are, how they differ, and how to match the tool to the task. **In one sentence:** This guide gives you the map and the compass to navigate the AI tools landscape — so you pick the right tool for the right job, every time. --- ## TL;DR - 30 Second Summary The AI tools landscape has five layers — and most people only see one: 1. **Foundation models** — the large AI models that power everything (GPT-4, Claude, Gemini, Llama) 2. **AI assistants** — consumer and enterprise chat interfaces (ChatGPT, Claude.ai, Gemini, Copilot) 3. **AI-native apps** — purpose-built tools for specific tasks (Midjourney, ElevenLabs, GitHub Copilot) 4. **AI-embedded tools** — existing tools with AI added (Notion AI, Grammarly, Google Workspace) 5. **AI platforms & APIs** — for building custom AI solutions (OpenAI API, Anthropic API, Azure AI) ```mermaid flowchart TD A["🤖 AI Tools Landscape"] --> B["🏗️ Layer 1: Foundation Models\nThe engines underneath everything"] A --> C["💬 Layer 2: AI Assistants\nGeneral-purpose chat interfaces"] A --> D["🎯 Layer 3: AI-Native Apps\nBuilt for one specific job"] A --> E["🔌 Layer 4: AI-Embedded Tools\nYour existing tools, now with AI"] A --> F["⚙️ Layer 5: AI Platforms & APIs\nBuild your own AI-powered solutions"] style A stroke:#2980b9 style B stroke:#8e44ad style C stroke:#27ae60 style D stroke:#f39c12 style E stroke:#e67e22 style F stroke:#4a90e2 ``` **Key idea:** Most people jump straight to the tool without understanding which layer they need. The right question is not "which AI should I use?" It is "what layer of the stack does my task require?" --- ## Full Table of Contents 1. [The Landscape — A Map of What Exists](#the-landscape) 2. [Foundation Models — The Engines Underneath](#foundation-models) 3. [AI Assistants — Your General-Purpose Interface](#ai-assistants) 4. [AI-Native Apps — Purpose-Built for One Job](#ai-native-apps) 5. [AI-Embedded Tools — Your Existing Stack, Upgraded](#ai-embedded-tools) 6. [AI Platforms & APIs — Building Custom Solutions](#ai-platforms) 7. [How to Choose — The Tool Selection Framework](#tool-selection) 8. [Building Your Team AI Stack](#team-stack) [Final Mental Model](#final-mental-model) | [Concept Reference Card](#concept-reference-card) ---

# AI Tools, Landscape & Platforms ### Map First. Layers Second. Then Your Personal Stack. --- # 🌟 The Only Idea You Need There is no single best AI tool. 👉 **There is the right tool for the right task — at the right layer of the stack.** A hammer is not better or worse than a screwdriver. It depends entirely on what you are building. The same principle applies to every AI tool that exists. --- # 🗺️ 1. The Landscape — A Map of What Exists ## 🧠 Core Intuition **Problem:** The AI tools landscape feels overwhelming — hundreds of tools, new launches every week, constant claims of being "the best." Without a map, it is impossible to navigate. **What happens:** Almost every AI tool in existence fits into one of five layers. Understanding the layers makes the landscape manageable. ```mermaid flowchart LR subgraph L1["🏗️ Layer 1: Foundation Models"] FM1["GPT-4o\n(OpenAI)"] FM2["Claude 3\n(Anthropic)"] FM3["Gemini\n(Google)"] FM4["Llama 3\n(Meta — Open Source)"] FM5["Mistral\n(Open Source)"] end subgraph L2["💬 Layer 2: AI Assistants"] AS1["ChatGPT"] AS2["Claude.ai"] AS3["Google Gemini"] AS4["Microsoft Copilot"] AS5["Perplexity"] end subgraph L3["🎯 Layer 3: AI-Native Apps"] AN1["Midjourney\n(Images)"] AN2["ElevenLabs\n(Voice)"] AN3["GitHub Copilot\n(Code)"] AN4["Runway\n(Video)"] AN5["Cursor\n(Code IDE)"] end subgraph L4["🔌 Layer 4: AI-Embedded Tools"] AE1["Notion AI"] AE2["Grammarly"] AE3["Google Workspace AI"] AE4["Microsoft 365 Copilot"] AE5["Salesforce Einstein"] end subgraph L5["⚙️ Layer 5: Platforms & APIs"] PL1["OpenAI API"] PL2["Anthropic API"] PL3["Google Vertex AI"] PL4["Azure OpenAI"] PL5["AWS Bedrock"] end L1 -->|"Powers"| L2 L1 -->|"Powers"| L3 L1 -->|"Powers"| L4 L1 -->|"Accessed via"| L5 ``` **Why it matters:** Most AI tools at Layers 2, 3, and 4 are interfaces built on top of the same foundation models at Layer 1. When a new AI assistant launches, it is usually a new interface — not a new model. Understanding this prevents you from chasing every new tool release. **One-line memory:** 👉 Five layers: **Foundation → Assistants → Native Apps → Embedded → Platforms** ---

⚙️ Technical Deep Dive (Optional)

**What is it called?** 👉 The **AI stack** — the layered architecture of AI systems from raw model to end-user interface **The model vs. interface distinction:** This is one of the most important distinctions in the landscape: | Layer | What it is | Who interacts with it | |---|---|---| | **Foundation model** | The trained neural network with billions of parameters | Developers, researchers, API consumers | | **Interface / product** | The UI and system prompt layer on top of the model | End users, business teams | Most "new AI tools" are new interfaces on existing models. The underlying intelligence comes from the foundation model. The interface determines the experience, constraints, and use-case focus. **Open source vs. proprietary:** | Type | Examples | Pros | Cons | |---|---|---|---| | **Proprietary** | GPT-4o, Claude 3, Gemini | Higher performance, managed infrastructure | Cost, data privacy concerns, vendor lock-in | | **Open source** | Llama 3, Mistral, Phi | Free to run, full data control, customisable | Requires infrastructure, lower out-of-box performance on complex tasks | **How dots connect:** - Foundation models (from How LLMs Work) are the transformer-based systems trained on internet-scale data - The interface layer is where system prompts, context windows, and output formatting are configured - Governance (from Responsible AI) determines which layers your organisation can use with which data

--- # 🏗️ 2. Foundation Models — The Engines Underneath ## 🧠 Core Intuition **Problem:** When people compare "ChatGPT vs Claude vs Gemini," they are often comparing interfaces — not the models themselves. Understanding the foundation model layer helps you make better decisions about what you are actually choosing. **What happens:** A small number of foundation models power most of the AI tools the world uses today. ```mermaid flowchart TD FM["🏗️ Major Foundation Model Families"] --> OAI["🟢 OpenAI\nGPT-4o · o1 · o3\nStrong across all tasks\nWidest ecosystem"] FM --> ANT["🔵 Anthropic\nClaude 3.5 Sonnet · Claude 3 Opus\nStrong reasoning & long context\nSafety-focused design"] FM --> GOO["🔴 Google DeepMind\nGemini 1.5 Pro · Flash\nMultimodal · Deep Google integration\nLong context window"] FM --> META["🟠 Meta\nLlama 3 (Open Source)\nFree to use and fine-tune\nSelf-hosted option"] FM --> MIS["🟣 Mistral\nMistral Large · Mixtral (Open Source)\nEfficient · European data residency option"] style FM stroke:#2980b9 style OAI stroke:#27ae60 style ANT stroke:#4a90e2 style GOO stroke:#c0392b style META stroke:#e67e22 style MIS stroke:#8e44ad ``` **How the major models compare — at a glance:** | Model Family | Strongest at | Best for | |---|---|---| | **OpenAI GPT-4o / o1** | Broad capability, reasoning, coding, vision | General use, coding, complex analysis | | **Anthropic Claude 3.5** | Long documents, nuanced writing, instruction following | Document work, research, writing, safety-critical tasks | | **Google Gemini 1.5** | Multimodal (text + image + audio + video), massive context | Media-rich tasks, long document analysis, Google Workspace | | **Meta Llama 3** | Open-source flexibility, self-hosted deployment | Privacy-first deployments, custom fine-tuning, cost control | | **Mistral** | Efficient performance, EU data residency | European compliance requirements, cost-efficient deployment | **Why it matters:** For most general workplace tasks, the major models perform comparably. The decision between them often comes down to integration, data residency, cost, and specific task requirements — not raw capability. **One-line memory:** 👉 Foundation models are the engines — **pick based on fit, not brand** ---

⚙️ Technical Deep Dive (Optional)

**Model size and performance:** Foundation models are measured partly by parameter count — the number of weights in the neural network. Larger is not always better: | Model size | Typical use | Trade-off | |---|---|---| | **Small (1B–7B params)** | On-device, edge deployment | Fast, cheap, less capable | | **Medium (13B–70B params)** | Self-hosted enterprise | Good balance of cost and capability | | **Large (70B+ params)** | Cloud API, frontier tasks | Highest capability, higher cost | | **Frontier (GPT-4o, Claude 3 Opus class)** | Complex reasoning, research | State of the art, highest cost | **Multimodality:** Modern foundation models increasingly accept multiple input types: - **Text** — all major models - **Images** — GPT-4o, Claude 3, Gemini 1.5 - **Audio** — GPT-4o, Gemini 1.5 - **Video** — Gemini 1.5, Runway (specialised) - **Documents** — Claude 3 (200K token context), Gemini 1.5 (1M token context) **How dots connect:** - Tokens and context windows (from How LLMs Work) determine how much content you can give a model at once - Model selection interacts directly with data privacy requirements (from Responsible AI) — where the model runs determines where your data goes

--- # 💬 3. AI Assistants — Your General-Purpose Interface ## 🧠 Core Intuition **Problem:** Most people's experience of AI is through one AI assistant — usually ChatGPT. Understanding the differences between the major assistants helps you choose the best one for your context and avoid over-relying on a single tool. **What happens:** AI assistants are general-purpose chat interfaces that wrap foundation models with system prompts, memory, plugins, and access tools. ```mermaid flowchart LR ASS["💬 Major AI Assistants"] --> CG["🟢 ChatGPT\nOpenAI · GPT-4o\nLargest user base\nStrong plugin ecosystem\nCode interpreter built in"] ASS --> CL["🔵 Claude.ai\nAnthropic · Claude 3.5\nLongest context window\nStrong for long documents\nNuanced, careful reasoning"] ASS --> GG["🔴 Google Gemini\nGoogle · Gemini 1.5\nDeep Google Workspace integration\nMultimodal natively\nReal-time web access"] ASS --> MC["🟡 Microsoft Copilot\nMicrosoft · GPT-4o\nBuilt into Windows & M365\nEnterprise-grade compliance\nStrong for Office users"] ASS --> PP["🟣 Perplexity\nMultiple models\nSearch-first AI\nCitations by default\nStrong for research"] style ASS stroke:#2980b9 style CG stroke:#27ae60 style CL stroke:#4a90e2 style GG stroke:#c0392b style MC stroke:#f39c12 style PP stroke:#8e44ad ``` **Choosing between AI assistants:** | If you need... | Best assistant | |---|---| | General-purpose daily use | ChatGPT or Claude.ai | | Long document analysis (100+ pages) | Claude.ai | | Real-time web search with citations | Perplexity | | Google Workspace integration | Google Gemini | | Microsoft 365 integration | Microsoft Copilot | | Enterprise data compliance (M365 users) | Microsoft Copilot | | Nuanced writing and careful instruction following | Claude.ai | | Code generation and data analysis | ChatGPT (Code Interpreter) | **One-line memory:** 👉 AI assistants are the interface — **the model underneath matters more than the brand** ---

⚙️ Technical Deep Dive (Optional)

**What makes assistants different from raw models:** AI assistants add several layers on top of the foundation model: | Layer | What it does | |---|---| | **System prompt** | Sets the assistant's persona, rules, and default behaviours | | **Memory** | Stores information about you across conversations (where enabled) | | **Tools / Plugins** | Extends the model with web search, code execution, file reading, etc. | | **Context management** | Decides how conversation history is handled within the context window | | **Safety filters** | Additional content moderation on top of model-level training | **Tiers within assistants:** Most major assistants have both free and paid tiers: | Tier | Typical limitations | Typical upgrades | |---|---|---| | **Free** | Older model version, rate limits, no file upload | — | | **Pro / Plus (~$20/month)** | Latest model, higher limits, file upload, image generation | — | | **Enterprise** | Data privacy agreements, admin controls, SSO, audit logs | — | For professional use with any sensitive data, the enterprise tier is typically the minimum acceptable level — because only enterprise agreements include the data privacy protections that most organisations require. **How dots connect:** - System prompts (from Effective Prompt Engineering) are how assistants are configured at the platform level - Context window (from What is Generative AI?) is why Claude.ai handles very long documents better than others — its context window is larger - Data privacy tiers (from Responsible AI) — free tier assistants should never be used with confidential data

--- # 🎯 4. AI-Native Apps — Purpose-Built for One Job ## 🧠 Core Intuition **Problem:** General-purpose AI assistants are broad. For some tasks — image generation, voice synthesis, code completion — a purpose-built tool will outperform a general assistant every time. **What happens:** AI-native apps are products built specifically for one modality or task type, with AI at the core of the entire product experience. ```mermaid flowchart TD ANA["🎯 AI-Native Apps by Category"] --> IMG["🖼️ Image Generation"] ANA --> VID["🎬 Video Generation"] ANA --> VOI["🎙️ Voice & Audio"] ANA --> COD["💻 Code & Development"] ANA --> DOC["📄 Document & Research"] ANA --> PRES["📊 Presentations & Design"] IMG --> I1["Midjourney — Highest image quality"] IMG --> I2["DALL-E 3 — Integrated into ChatGPT"] IMG --> I3["Stable Diffusion — Open source, self-hosted"] IMG --> I4["Adobe Firefly — IP-safe, creative suite"] VID --> V1["Runway — Video generation & editing"] VID --> V2["Sora (OpenAI) — Text-to-video"] VID --> V3["Pika — Short video creation"] VOI --> VO1["ElevenLabs — Voice cloning & synthesis"] VOI --> VO2["Whisper (OpenAI) — Speech-to-text"] VOI --> VO3["Descript — Audio/video editing with AI"] COD --> C1["GitHub Copilot — Code completion in IDE"] COD --> C2["Cursor — AI-native code editor"] COD --> C3["Replit — AI coding environment"] DOC --> D1["Perplexity — AI search with citations"] DOC --> D2["NotebookLM (Google) — Document Q&A"] DOC --> D3["Elicit — Research paper analysis"] PRES --> P1["Gamma — AI presentation builder"] PRES --> P2["Canva AI — Design with AI"] PRES --> P3["Beautiful.ai — Slide design"] style ANA stroke:#2980b9 style IMG stroke:#f39c12 style VID stroke:#c0392b style VOI stroke:#8e44ad style COD stroke:#4a90e2 style DOC stroke:#27ae60 style PRES stroke:#e67e22 ``` **When to use a native app vs a general assistant:** | Situation | Use native app | Use general assistant | |---|---|---| | Need high-quality images | ✅ Midjourney, Adobe Firefly | ❌ DALL-E quality lower | | Need code completion as you type | ✅ GitHub Copilot, Cursor | ❌ Copy-paste workflow | | Need to analyse 50 research papers | ✅ NotebookLM, Elicit | ⚠️ Possible but clunky | | Need a voice-over for a video | ✅ ElevenLabs | ❌ Not designed for this | | Need a draft email | ❌ Overkill | ✅ General assistant is fine | | Need to brainstorm ideas | ❌ Overkill | ✅ General assistant is fine | **One-line memory:** 👉 Native apps beat general assistants **when the task is specialised and quality matters** ---

⚙️ Technical Deep Dive (Optional)

**Why native apps outperform general assistants on specialist tasks:** | Factor | General assistant | Native app | |---|---|---| | **Model** | General-purpose LLM | Specialised model (diffusion, voice, code) | | **Interface** | Text in, text out | Task-optimised UI and workflow | | **Fine-tuning** | Broad training | Trained specifically on domain data | | **Output format** | Text / basic images | Images, audio, video, code in native format | | **Iteration** | Re-prompt in chat | Purpose-built iteration tools | **The modality gap:** Language models (GPT-4o, Claude, Gemini) are exceptional at text and code. For images, audio, and video, dedicated models — diffusion models for images, neural TTS for voice, video generation models — significantly outperform what a language model can do via a chat interface. **How dots connect:** - Types of AI (from What is Generative AI?) — Generative AI covers not just text but images, audio, and video. Native apps are where those modalities are most powerful - Task fit (from AI Capabilities, Limitations & Use Cases) — matching the right type of AI to the task is the same principle, applied to tools

--- # 🔌 5. AI-Embedded Tools — Your Existing Stack, Upgraded ## 🧠 Core Intuition **Problem:** Most people look for new AI tools without realising the tools they already use every day have AI built in. For many tasks, the highest-value AI move is not adopting something new — it is unlocking what is already there. **What happens:** Most major productivity, collaboration, and business tools now have AI capabilities embedded directly into the product. ```mermaid flowchart LR AE["🔌 AI-Embedded Tools by Category"] --> PROD["📝 Productivity & Writing"] AE --> COLLAB["🤝 Collaboration"] AE --> CRM["💼 CRM & Sales"] AE --> DATA["📊 Data & Analytics"] AE --> DEV["💻 Development"] PROD --> P1["Microsoft 365 Copilot\nWord · Excel · PowerPoint · Outlook"] PROD --> P2["Google Workspace AI\nDocs · Sheets · Gmail · Slides"] PROD --> P3["Notion AI\nDrafting · Summarising · Database queries"] PROD --> P4["Grammarly\nWriting assistance · Tone adjustment"] COLLAB --> C1["Slack AI\nThread summaries · Search · Recaps"] COLLAB --> C2["Zoom AI Companion\nMeeting summaries · Action items"] COLLAB --> C3["Teams Copilot\nMeeting notes · Chat summaries"] CRM --> R1["Salesforce Einstein\nLead scoring · Email drafting · Insights"] CRM --> R2["HubSpot AI\nContent generation · Forecasting"] DATA --> D1["Tableau AI\nNatural language queries · Insights"] DATA --> D2["Power BI Copilot\nReport generation · Q&A"] DEV --> DV1["GitHub Copilot\nCode in your existing IDE"] DEV --> DV2["Jira AI\nTicket summarisation · Sprint insights"] style AE stroke:#2980b9 style PROD stroke:#27ae60 style COLLAB stroke:#f39c12 style CRM stroke:#e67e22 style DATA stroke:#8e44ad style DEV stroke:#4a90e2 ``` **Why embedded tools are often the highest-value starting point:** | Advantage | Why it matters | |---|---| | **No new tool to learn** | AI is where you already work — zero adoption friction | | **Context-aware** | AI has access to your documents, emails, and data | | **Enterprise-grade by default** | Data stays within your existing compliance boundary | | **Already purchased** | Most M365 or Google Workspace licences include AI features | **One-line memory:** 👉 Before adopting new AI tools — **check what is already in the tools you pay for** ---

⚙️ Technical Deep Dive (Optional)

**The Microsoft 365 Copilot model:** Microsoft 365 Copilot is the most widely deployed enterprise AI tool in the world. Its architecture is significant: - Built on GPT-4o (OpenAI) via Microsoft's Azure partnership - Data stays within your Microsoft 365 tenant — not sent to OpenAI for training - Access to your emails, documents, calendar, Teams messages as context - Enterprise data protection by default — the same compliance boundary as M365 This is a fundamentally different data model from using ChatGPT.com — and why enterprise AI governance (from Responsible AI) often centres on M365 Copilot as the approved tool. **The Google Workspace equivalent:** Google Gemini for Workspace (formerly Duet AI) follows the same model — Gemini 1.5 Pro under the hood, data staying within the Google Workspace tenant, integrated directly into Docs, Gmail, Sheets, and Meet. **How dots connect:** - Data privacy and tool tiers (from Responsible AI) — embedded enterprise tools like M365 Copilot and Google Workspace AI are typically the first "approved" AI tools in any governance framework - The 5 work modes (from AI in the Workplace) — Draft, Summarise, Research, Explain, Transform — all map directly to embedded tool features

--- # ⚙️ 6. AI Platforms & APIs — Building Custom Solutions ## 🧠 Core Intuition **Problem:** Off-the-shelf AI tools do not always fit every use case. Organisations with specific workflows, proprietary data, or compliance requirements sometimes need to build their own AI-powered solutions — or customise existing models. **What happens:** AI platforms and APIs provide the infrastructure for building custom AI applications on top of foundation models — without training a model from scratch. ```mermaid flowchart TD PLAT["⚙️ AI Platforms & APIs"] --> DIRECT["Direct Model APIs\nAccess models directly"] PLAT --> CLOUD["Cloud AI Platforms\nManaged infrastructure + models"] PLAT --> BUILD["AI App Builders\nNo-code / low-code AI apps"] PLAT --> VECTOR["Vector Databases\nFor RAG and knowledge retrieval"] DIRECT --> D1["OpenAI API — GPT-4o, o1"] DIRECT --> D2["Anthropic API — Claude 3.5"] DIRECT --> D3["Google AI Studio — Gemini"] CLOUD --> C1["Azure OpenAI — GPT-4 in Microsoft cloud"] CLOUD --> C2["AWS Bedrock — Multiple models in AWS"] CLOUD --> C3["Google Vertex AI — Gemini in Google cloud"] BUILD --> B1["Langchain — AI workflow orchestration"] BUILD --> B2["LlamaIndex — Document Q&A pipelines"] BUILD --> B3["Flowise / n8n — No-code AI workflows"] VECTOR --> V1["Pinecone"] VECTOR --> V2["Weaviate"] VECTOR --> V3["pgvector (PostgreSQL)"] style PLAT stroke:#2980b9 style DIRECT stroke:#27ae60 style CLOUD stroke:#4a90e2 style BUILD stroke:#f39c12 style VECTOR stroke:#8e44ad ``` **When you need a platform vs a product:** | Situation | Use a product | Use a platform / API | |---|---|---| | General daily tasks | ✅ ChatGPT / Claude.ai | ❌ Overkill | | Custom AI on your own data | ❌ Not possible | ✅ RAG pipeline + API | | AI embedded in your own app | ❌ Not possible | ✅ API integration | | Specific data residency requirements | ⚠️ Depends on tool | ✅ Azure / AWS / GCP deployment | | Fine-tuning on your own domain | ❌ Not possible | ✅ Fine-tuning via API | | Automated AI workflows | ⚠️ Limited | ✅ Langchain / n8n / custom | **One-line memory:** 👉 Platforms and APIs are for when **the product does not exist yet — or needs to be yours** ---

⚙️ Technical Deep Dive (Optional)

**RAG — the most important enterprise AI architecture pattern:** Retrieval-Augmented Generation (RAG) is the pattern that powers most custom enterprise AI solutions: ```mermaid flowchart LR DOCS["📄 Your documents\n(PDFs, wikis, databases)"] --> EMBED["Embed into\nvector database"] QUERY["🧑 User question"] --> SEARCH["Semantic search\nin vector DB"] EMBED --> SEARCH SEARCH --> CHUNKS["Relevant document\nchunks retrieved"] CHUNKS --> PROMPT["Injected into\nLLM prompt as context"] PROMPT --> ANSWER["✅ Grounded answer\nbased on your documents"] style DOCS stroke:#2980b9 style ANSWER stroke:#27ae60 ``` RAG solves the two biggest enterprise AI problems simultaneously: - **Knowledge cutoff** — the model answers from your documents, not its training data - **Hallucination** — grounded in real retrieved content, not pure generation **Fine-tuning vs RAG:** | Approach | When to use | Cost | |---|---|---| | **Prompt engineering** | Most tasks | Free | | **RAG** | Custom knowledge base, proprietary data | Low-Medium | | **Fine-tuning** | Specific style, format, or domain behaviour | High | | **Training from scratch** | Almost never for enterprises | Very high | **How dots connect:** - RAG (from How LLMs Work) — section 8 of that module covers the retrieval architecture in detail - Hallucination (from How LLMs Work) — RAG is one of the primary mitigation strategies - Governance (from Responsible AI) — cloud platform choice (Azure, AWS, GCP) is often the compliance decision

--- # 🧭 7. How to Choose — The Tool Selection Framework ## 🧠 Core Intuition **Problem:** With five layers and hundreds of tools, the selection decision can feel paralysing. A simple framework cuts through the noise. **What happens:** Tool selection comes down to four questions, asked in order. ```mermaid flowchart TD START["📋 I have a task to do with AI"] START --> Q1{"1. Is it already in\na tool I use daily?"} Q1 -->|"Yes"| A1["🔌 Use the embedded AI feature\n(M365 Copilot, Google Workspace AI, Notion AI)"] Q1 -->|"No"| Q2 Q2{"2. Is it a specialist task\n(image, voice, code, video, research)?"} Q2 -->|"Yes"| A2["🎯 Use a purpose-built native app\n(Midjourney, GitHub Copilot, ElevenLabs...)"] Q2 -->|"No"| Q3 Q3{"3. Is it a general language task\n(write, summarise, analyse, explain)?"} Q3 -->|"Yes"| A3["💬 Use a general AI assistant\n(ChatGPT, Claude.ai, Gemini, Copilot)"] Q3 -->|"No"| Q4 Q4{"4. Does your organisation need\na custom or proprietary solution?"} Q4 -->|"Yes"| A4["⚙️ Use an AI platform or API\n(Azure OpenAI, Anthropic API, AWS Bedrock)"] Q4 -->|"No"| A5["🔁 Re-examine the task —\nAI may not be the right tool"] style START stroke:#2980b9 style A1 stroke:#e67e22 style A2 stroke:#f39c12 style A3 stroke:#27ae60 style A4 stroke:#4a90e2 style A5 stroke:#c0392b ``` **Within each layer — how to choose between options:** | Decision factor | What to consider | |---|---| | **Data sensitivity** | Can this data go to this tool's servers? Check your org's AI policy | | **Integration** | Does it connect with the tools your team already uses? | | **Cost** | What is the per-user or per-call cost at your expected volume? | | **Performance on your task** | Test it — benchmark the tools on a real sample of your work | | **Support & reliability** | Is this a stable product with enterprise support? | | **Compliance** | Does it meet your data residency, audit, and access control requirements? | **One-line memory:** 👉 Ask four questions in order: **Embedded → Native → Assistant → Platform** ---

⚙️ Technical Deep Dive (Optional)

**The build vs. buy decision for AI:** Organisations frequently face the question of whether to build a custom AI solution or buy an off-the-shelf product. | Factor | Favour buying | Favour building | |---|---|---| | **Task specificity** | Generic task (summarisation, drafting) | Highly specific workflow | | **Data sensitivity** | Low-sensitivity data | Highly proprietary data that cannot leave org systems | | **Time to value** | Need results in days | Can invest weeks/months | | **Maintenance** | Want vendor to maintain | Have internal capability | | **Cost** | Low usage volume | High usage volume (API cheaper at scale) | **Evaluation framework for AI tools:** Before committing to any AI tool at organisational scale: 1. **Proof of concept** — test the tool on 10 representative real tasks 2. **Data assessment** — what data will it touch? Is that compliant? 3. **Security review** — does it meet your IT security standards? 4. **Cost modelling** — what is the TCO at expected usage? 5. **Governance fit** — can it be logged, audited, and controlled? **How dots connect:** - Risk tiering (from Responsible AI) feeds directly into step 2 and 3 of this evaluation - The AI governance framework (from Responsible AI) — tool selection is one of the first outputs a governance process should produce

--- # 🏗️ 8. Building Your Team AI Stack ## 🧠 Core Intuition **Problem:** Individual tool choices are one thing. Building a coherent, approved, cost-effective AI stack for an entire team or organisation is another challenge — and one that most teams have not yet done deliberately. **What happens:** A team AI stack is a curated set of AI tools — one or two per layer — that covers the team's core AI use cases, meets governance requirements, and does not create tool sprawl. ```mermaid flowchart TD STACK["🏗️ Team AI Stack"] --> T1["Layer 2: Primary AI Assistant\nThe tool everyone uses for general tasks"] STACK --> T2["Layer 3: Specialist Tools\n1-2 purpose-built tools for your team's top specialist needs"] STACK --> T3["Layer 4: Embedded AI\nEnable AI features in tools you already pay for"] STACK --> T4["Layer 5: Platform (if needed)\nOnly if building custom solutions"] T1 --> EX1["Example stacks by org type"] EX1 --> MSFT["🟡 Microsoft-first org\nCopilot (M365) · GitHub Copilot · Bing/Copilot Pro"] EX1 --> GOOG["🔴 Google-first org\nGemini for Workspace · Gemini Advanced · NotebookLM"] EX1 --> MIXED["🟢 Mixed / best of breed\nClaude.ai Pro · GitHub Copilot · Perplexity · Notion AI"] EX1 --> BUILD["🔵 Technical / building team\nAnthropic API · Azure OpenAI · Cursor · Langchain"] style STACK stroke:#2980b9 style T1 stroke:#27ae60 style T2 stroke:#f39c12 style T3 stroke:#e67e22 style T4 stroke:#4a90e2 style MSFT stroke:#f39c12 style GOOG stroke:#c0392b style MIXED stroke:#27ae60 style BUILD stroke:#4a90e2 ``` **Building your team stack — a practical process:** | Step | What to do | |---|---| | **1. Audit current tools** | What AI tools are people already using, officially or unofficially? | | **2. Map use cases** | What are the top 5 tasks where AI could save the most time? | | **3. Check governance constraints** | Which tools are approved? What data can they touch? | | **4. Select and pilot** | Pick one tool per layer. Run a 30-day team pilot. | | **5. Train and embed** | Ensure everyone knows how to use the tools in their workflow | | **6. Review and iterate** | Quarterly review — is the stack still the right one? | **One-line memory:** 👉 A team AI stack = **one tool per layer, chosen deliberately, governed consistently** ---

⚙️ Technical Deep Dive (Optional)

**Tool sprawl — the hidden cost:** Unmanaged AI tool adoption leads to tool sprawl — employees using dozens of different AI tools, often with overlapping functionality, inconsistent data practices, and no shared prompt libraries or workflows. The cost of sprawl: - Data risk — different tools with different data handling, often unchecked - Training overhead — every new tool requires relearning - No shared prompt library — every person reinvents their prompts - Budget inefficiency — paying for features that overlap - Governance gap — impossible to audit what is being used and how **The governance and stack connection:** The team AI stack is the operational output of the AI governance framework from Responsible AI module. The approved tools list (a governance output) should map directly to the team stack. | Governance output | Stack implication | |---|---| | Approved tools list | Defines which tools are eligible | | Data classification policy | Determines which tools can handle which data | | Use case registry | Identifies which specialist tools are needed | | Risk tiering | Sets review requirements for each tool's outputs | **How dots connect:** - Approved tools list (from Responsible AI Governance) is the governance input that shapes which tools belong in the stack - Prompt library (from AI in the Workplace) — a team stack should come with a shared prompt library for each tool's common use cases

--- # 🧠 Final Mental Model ## 🧠 Simple Version (For Everyone) ```mermaid flowchart TD TASK["📋 Task to do with AI"] TASK --> L{"Which layer?"} L -->|"Already in daily tool"| EMB["🔌 Layer 4\nEmbedded AI\nM365, Google Workspace, Notion"] L -->|"Specialist modality"| NAT["🎯 Layer 3\nNative App\nMidjourney, Copilot, ElevenLabs"] L -->|"General language task"| ASS["💬 Layer 2\nAI Assistant\nChatGPT, Claude.ai, Gemini"] L -->|"Custom / proprietary"| API["⚙️ Layer 5\nPlatform / API\nAzure, Anthropic, AWS"] EMB --> GOV{"✅ Governance check:\nApproved tool?\nRight data tier?"} NAT --> GOV ASS --> GOV API --> GOV GOV -->|"Yes"| USE["🚀 Use it\nWith review habit from AI in the Workplace"] GOV -->|"No"| FIX2["🔧 Use approved alternative\nor escalate for approval"] style TASK stroke:#2980b9 style EMB stroke:#e67e22 style NAT stroke:#f39c12 style ASS stroke:#27ae60 style API stroke:#4a90e2 style USE stroke:#27ae60 style FIX2 stroke:#c0392b ``` 👉 The AI tools landscape is not overwhelming once you understand the layers. 👉 The best tool is not the most powerful — it is the best fit for the task, within your governance constraints. 👉 Your team stack should be deliberate, minimal, and reviewed regularly — not whatever everyone individually downloaded. ---

⚙️ Technical Flow (For Professionals)

```mermaid flowchart TD subgraph STACK_ARCH["🏗️ The AI Stack — Full Architecture"] direction TD FM2["Foundation Models\nGPT-4o · Claude 3.5 · Gemini 1.5 · Llama 3"] FM2 --> API2["APIs\nOpenAI · Anthropic · Google AI Studio"] API2 --> CLOUD2["Cloud Platforms\nAzure OpenAI · AWS Bedrock · Google Vertex"] API2 --> ASS2["AI Assistants\nChatGPT · Claude.ai · Gemini · Copilot"] API2 --> NAT2["Native Apps\nGitHub Copilot · Midjourney · ElevenLabs"] CLOUD2 --> CUSTOM["Custom Solutions\nRAG Pipelines · Fine-tuned Models · AI Agents"] ASS2 --> EMBED2["Embedded Features\nM365 Copilot · Google Workspace · Notion AI"] end subgraph SELECT["🧭 Selection Logic"] direction TD S1["Task type → Layer"] S2["Data sensitivity → Tool tier"] S3["Governance constraints → Approved tools list"] S4["Cost at volume → Build vs buy"] S5["Performance on task → Benchmark before committing"] end STACK_ARCH -.->|"Informed by"| SELECT ```

--- # 🚀 If You Remember Only Three Things 1. **There are five layers — know which one your task needs.** Embedded → Native → Assistant → Platform. Most tasks are answered at Layer 2 or Layer 4, not by adding new tools. 2. **The best tool is not the most powerful — it is the best fit.** Foundation models are roughly comparable on most tasks. The difference is integration, data handling, and task-specific capability. 3. **Build a deliberate team stack.** Unmanaged AI tool adoption creates sprawl, risk, and wasted spend. One approved tool per layer, governed consistently, beats twenty tools nobody fully uses. --- # 📚 Concept Reference Card | Concept | One-Line Memory | Layer | |---|---|---| | Foundation model | The trained neural network powering everything | Layer 1 | | Proprietary model | Highest performance, managed by vendor (OpenAI, Anthropic, Google) | Layer 1 | | Open source model | Free to run, full data control (Llama, Mistral) | Layer 1 | | AI assistant | General-purpose chat interface on top of a foundation model | Layer 2 | | Enterprise tier | Data stays in your compliance boundary — required for sensitive data | Layer 2 | | AI-native app | Purpose-built for one modality or task (image, code, voice) | Layer 3 | | AI-embedded tool | AI added to a tool you already use (M365, Notion, Grammarly) | Layer 4 | | AI API | Direct programmatic access to a foundation model | Layer 5 | | Cloud AI platform | Foundation models hosted in your cloud (Azure, AWS, GCP) | Layer 5 | | RAG | Custom AI that answers from your documents, not its training data | Layer 5 | | Fine-tuning | Adapting a model to your specific domain or style | Layer 5 | | Tool sprawl | The cost of unmanaged AI tool adoption across a team | All layers | | Team AI stack | A curated, approved, governed set of AI tools per layer | All layers | --- # 🎯 Success Metrics Track these to know if your team has genuine AI tools fluency: - ✅ Can name the five layers of the AI stack and give an example of each - ✅ Uses the four-question selection framework before adopting a new AI tool - ✅ Knows which AI tools are approved by their organisation — and which are not - ✅ Never uses a consumer-tier AI tool with sensitive organisational data - ✅ Has a defined team AI stack — not just individual tool preferences - ✅ Can explain the difference between a foundation model and an AI assistant - ✅ Knows when a specialist native app will outperform a general assistant - ✅ Understands what RAG is and when a custom solution is needed vs an off-the-shelf tool --- # 📊 Evidence from Practice **Real example: Enterprise AI Tools Rationalisation Programme** **Challenge:** - 200-person organisation with 40+ different AI tools in active use across teams - No shared vocabulary — people were comparing ChatGPT and Midjourney as if they were the same type of thing - Data risk — multiple teams using free-tier consumer AI tools with client data - Budget waste — significant overlap between tools people were paying for individually - No governance — IT had no visibility into which AI tools were in use **Approach:** - Half-day AI tools landscape workshop for team leads and IT stakeholders - Layer mapping exercise — teams mapped every current AI tool to the five-layer framework - Data risk audit — identified which tools were being used with which data - Stack design session — each team defined a target stack of 3-4 approved tools - Governance integration — connected stack decisions to the AI governance framework **Results:** - AI tool count reduced from 40+ to 12 approved tools across five layers - Data risk exposure eliminated for three teams using consumer tools with client data - Annual AI tool spend reduced by consolidating overlapping subscriptions - IT gained full visibility into AI tool use for the first time - Shared prompt libraries created for the two primary AI assistants - Teams reported higher confidence and consistency in AI tool use --- # 🎯 Book Your Workshop Ready to bring clarity to your organisation's AI tools landscape — and build a stack that is deliberate, governed, and genuinely effective? → **[Book a Discovery Call](https://calendly.com/naveen-innoventurex/free-consultation-with-naveen)** or → **[Contact me directly](/contact)** *** *Half-day workshop includes:* - *Interactive sessions on all 8 core concepts* - *Five-layer AI landscape mapping exercise* - *Team AI tool audit — what are you actually using?* - *Data risk identification — which tools are creating exposure?* - *Team stack design session — target stack per team* - *Tool selection framework as a takeaway* - *Integration with your AI governance framework*