Azure AI Foundry Services
8 core services β tap each card to expand
AI Content Safety
Scans text and images to detect and block hate speech, violence, sexual content, or self-harm material before anyone sees it.
Nigerian Use CaseA Lagos online gaming platform blocks offensive messages in real time before they reach other players.
Azure OpenAI Service
Enterprise-grade access to GPT, DALLΒ·E, and embeddings inside Azure's secure infrastructure β for chat, summarisation, and code generation.
Nigerian Use CaseGTBank builds an internal assistant that helps analysts summarise credit reports and draft client communications.
AI Speech
Converts voice to text and text to natural-sounding voice. Also handles real-time translation, voice cloning, and speaker recognition.
Exam PipelineAI Speech + AI Language = speak notes β transcribe β detect mood. Classic two-service combo.
Nigerian Use CaseA Nigerian insurance company transcribes call centre recordings automatically β cutting note-taking time by 70%.
AI Translator
Instant text and full document translation across dozens of languages β preserving original formatting. NOT the same as AI Language.
Nigerian Use CaseDangote Cement translates English contracts into French and Portuguese for West African partners in seconds.
AI Vision
Extracts insights from still images β object detection, OCR, image captioning, facial recognition. Evolved from Form Recognizer.
Nigerian Use CaseNigerian Ports Authority scans shipping container labels automatically instead of manual data entry.
AI Language
π Sentiment Analysis β detects positive/neutral/negative mood
π· Entity Recognition β spots names, dates, places, companies
π Key Phrase Extraction β summarises main points
π Translation β converts text between languages
Do NOT build a custom classification model when AI Language already has Sentiment Analysis out of the box.
Nigerian Use CaseStanbic IBTC flags negative customer feedback from thousands of reviews automatically.
AI Video Indexer
Analyses video to extract transcriptions, detect faces, segment scenes, identify topics β creating searchable video archives.
Nigerian Use CaseChannels TV searches its archive to find all clips where a specific politician spoke, without watching hours of footage.
AI Search
Searches across your organisation's private internal documents using semantic, vector, and hybrid search.
Critical TrapAI Search β File Search in Copilot Studio. AI Search = Azure enterprise service. File Search = Copilot Studio uploaded files.
Nigerian Use CaseTemplars Law Firm finds exact contract clauses from plain-language questions.
Rent, Buy, Build / Own
The AI ecosystem sourcing strategy β core AB-731 exam domain
The Three Strategies at a Glance
| Strategy | What It Means | Microsoft Example | When to Use |
|---|---|---|---|
| πͺ Rent (SaaS) | Subscribe to ready-to-use AI apps. Zero coding or infrastructure needed. | Microsoft 365 Copilot | Staff need AI in existing M365 apps NOW. No IT budget for development. |
| π Buy (API/PaaS) | Purchase AI APIs or platforms and integrate them into your own systems. | Azure OpenAI Service | Need custom interface or logic, but no in-house ML team to train models. |
| π Build/Own | Develop bespoke AI models from scratch using your own data science team and proprietary data. | Training a custom LLM | Highly regulated data (NDPR), unique competitive advantage, very large scale. |
The Exam Golden Rule
π Rent = Taking a Taxi
You say "Take me to the airport." You don't touch the steering wheel or know the route. Microsoft's AI does all the complex work. You just configure what you want.
π Build = Renting a Car
You must build the route, steer, press the pedals yourself. Developers must write code, tune the AI, and test it before anyone can use it. Takes time and skills you may not have.
Decision Triggers β What to Look for in Exam Questions
| If the scenario says... | Strategy | Reasoning |
|---|---|---|
| "No available IT budget for programming" / "zero coding" | πͺ Rent | Only SaaS works without development effort |
| "Deploy quickly" / "out-of-the-box" / "just turn it on" | πͺ Rent | SaaS is ready immediately |
| "Custom interface" but "no ML team" | π Buy | API/PaaS gives custom interface without training models |
| "Strict NDPR / data residency" / "proprietary data" | π Build/Own | Sensitive data can't leave your environment |
| "Unique competitive advantage" / "core IP" | π Build/Own | Justifies the cost of training from scratch |
| "API cost at scale getting too high" (volume trap) | π Build/Own | At sufficient volume, building own model beats API bills |
| "Extending Microsoft 365 Copilot with custom agents" | π Build | "Custom" = development required = violates "no coding" requirement |
Nigerian Business Scenarios
| Organisation | Scenario | Strategy |
|---|---|---|
| FMCG HR Chatbot | Lagos FMCG company needs HR chatbot inside Teams, no coding budget | πͺ Rent (M365 Copilot) |
| Fintech KYC | Fintech needs custom KYC verification UI with API integration, no ML team | π Buy (Azure OpenAI API) |
| NGX Trading Firm | Trading firm wants proprietary market prediction using 10 years of internal data β core competitive IP | π Build/Own |
| HealthTech / NDPR | Hospital cannot send patient data to Microsoft cloud due to NDPR compliance | π Build/Own |
| Japa Turnover | Built ML team, key engineers left to Europe ("Japa syndrome") β pivot needed | π Buy (API/PaaS) |
Embedded vs Stand-Alone Agents
Align AI architecture to the right business problem
Side-by-Side Comparison
| Feature | π§© Embedded Agent | π’ Stand-Alone Agent |
|---|---|---|
| Definition | AI integrated directly into an existing application or workflow | Independent AI in its own dedicated platform or portal |
| User Interface | Lives inside software users already use (sidebar, chat bubble, inline) | Has its own distinct web page, app, or portal |
| Context Awareness | Automatically knows what you are working on in that app | Blank slate β you must provide full context every time |
| Primary Benefit | Frictionless β no context switching needed | Highly flexible; handles broad, complex, cross-domain tasks |
| Examples | Microsoft 365 Copilot inside Word, Teams, Outlook | ChatGPT, Claude, a custom portal built on Copilot Studio |
| Build/Buy Link | Buy or Extend (purchase M365 Copilot) | Build (Copilot Studio or Azure AI) |
The Office Building Analogy
π§© Embedded = Colleague at Your Desk
You're designing a proposal in Word. A colleague stands right next to you looking at the same screen. You say "make this punchier" β they fix it instantly. Zero briefing needed. Full context.
π’ Stand-Alone = Help Desk Down the Hall
You must stop work, walk to a different office, and fully brief the expert: "I'm building a training program, here's my outline..." They give brilliant answers but you carry the result back yourself.
M365 Copilot vs Copilot Studio
| Microsoft 365 Copilot | Microsoft Copilot Studio | |
|---|---|---|
| What it is | Ready-to-drive car. Already built. Lives inside Teams, Word, Outlook. | Mechanic's garage. You build your own custom bot or connection here. |
| When to use | Summarise Teams meetings, read M365 emails, draft Word docs β out of the box | Connect to 3rd-party systems (Workday, Salesforce), build custom chatbots |
| Development needed? | No β just assign licenses and turn it on | Yes β configuration, logic building, and integration work required |
| Exam trigger word | "Out-of-the-box", "within M365", "no development" | "Custom", "3rd-party integration", "build a bot", "Workday/Salesforce" |
Security, Microsoft Graph & RBAC
How data is protected when Copilot answers your questions
How Microsoft Graph Controls What Copilot Can See
1. You ask Copilot a question
"Summarise project updates from my team." Copilot does not own or directly store your company's data.
2. Copilot asks Microsoft Graph for the data
Copilot cannot just wander around opening locked cabinets. It must go through Microsoft Graph β the Head Librarian and Security Guard.
3. Microsoft Graph checks your badge (RBAC)
Graph does two things: (a) finds connected documents relevant to your query, and (b) checks your security clearance β it will ONLY hand over files you personally have permission to read.
4. Only allowed data reaches Copilot
If HR salary files are restricted, Graph blocks them. Copilot will say "I couldn't find that information" or "you don't have access."
The Corporate Library Analogy
π Microsoft Entra ID = The Badge Office
Issues you a digital ID badge based on your job title. "This person is an HR Manager." It creates your Role-Based Access Control (RBAC) identity. Without this badge, you have no permissions at all.
π Microsoft Graph = The Security Librarian
Checks your badge before giving any file to Copilot. If you ask about the CEO's salary but you're not in the payroll group β Graph blocks it. The AI did not fail; the permissions allowed it.
RBAC, Monitoring, Purview β Active vs Passive
| Control Type | What It Does | Analogy | Enough on its own? |
|---|---|---|---|
| Centralized Monitoring | Records what AI does β logs, audit trails, dashboards | Security camera over the vault β records the robbery but doesn't stop it | β No β passive only |
| RBAC + Controlled Intervention | Restricts what AI can DO β active enforcement, halts unsafe actions in real time | Combination lock + automatic steel doors that drop the moment an alarm triggers | β Yes β active control |
| Microsoft Purview / DLP | Labels documents and blocks Copilot from sharing sensitive content | Digital stamps on every file + scanner at the door checking each one | β Yes β prevents oversharing |
Tenant Boundaries
| Concept | What It Means | Correct? |
|---|---|---|
| β Isolation within tenant boundary | Your company's data stays locked on your floor of Microsoft's skyscraper. Other companies cannot see it. | β Correct enterprise benefit |
| β Public sharing across tenants | Your AI outputs shared with every other company | β Catastrophic data leak |
| β Training models on shared customer data | Microsoft blends your private files into its public AI training | β Your strategy leaks to competitors |
| β Unrestricted anonymous access | Anyone walks in and uses your AI tools without a badge | β Violates every security rule |
Search Types
Keyword, Conceptual, Hybrid, and Semantic Ranking β know which to use
The Four Search Approaches
| Search Type | How It Works | Analogy | Best For |
|---|---|---|---|
| π€ Keyword Search | Looks for exact words only. Rigid, like CTRL+F. | Searching "vacation" β skips "holiday", "annual leave", "day off" | Strict category filters, exact text matches |
| π§ Conceptual (Vector) | Searches for the MEANING behind words β finds related concepts even if exact words differ. | Type "I need time off" β AI finds "Leave Policy" doc even without those words | Broad discovery, synonyms, understanding intent |
| π Hybrid Search | Combines Keyword AND Conceptual at the same time. Bouncer (filter) + Detective (meaning). | Online shopping: search "warm coats" (conceptual) AND tick "Brand: North Face" (keyword) | Find related concepts WITHIN a strictly filtered category |
| π Semantic Ranking | Doesn't search for new documents. Takes existing results and RE-ORDERS them β most relevant first. | Intern brings 50 files. You reorder the stack. Doesn't help you FIND or FILTER. | Improving ordering of already-found results only |
The Bouncer + Detective Analogy
πͺ The Bouncer = Keyword Filter
Strictly blocks anything NOT in the approved category. "IT Manuals only." Without a bouncer, searching "broken screen" returns results about broken office windows, TVs, and phones.
π The Detective = Conceptual Search
Inside the allowed pile, finds meaning. "Broken screen" finds "shattered monitor", "cracked display", "damaged LCD" β even if those exact words don't match.
Need RELATED MEANING / synonyms? β Must include Conceptual = Hybrid or Conceptual only
Both filter AND meaning? β Hybrid Search
Only reorder what you already found? β Semantic Ranking
Semantic Ranking fails when you need to filter β it doesn't act as a bouncer.
Microsoft AI Licensing
M365 vs Azure β the two sides of the house
Developers building custom Azure workloads = Azure billing
The Three Licensing Models
| Model | How It's Billed | Analogy | Best Used When... |
|---|---|---|---|
| π Copilot Add-on License | Fixed price per user, per month | Unlimited gym membership β same cost whether you go once or every day | Employees use Copilot daily in Word/Teams/Outlook. Predictable monthly cost. |
| β‘ Pay-As-You-Go (Azure) | Billed on actual usage β like an electricity bill | Electric utility bill β you only pay for what you actually use each month | Irregular, seasonal usage. Custom Azure AI workloads. Field agents who use AI only 1β2x a month. |
| π Prepaid AI Capacity Commitment | Upfront cost for reserved raw computing power in Azure | Renting a dedicated private server room β you own it 24/7 regardless of use | Developers running large-scale production workloads with stable, predictable high-volume usage on Azure. |
The Two Sides of the Microsoft House
π’ Side 1: Microsoft 365 (SaaS)
For employees using standard Office apps.
Scenario clues: "Give HR staff access to Copilot in Word/Teams", "assign Copilot seats", "out-of-the-box"
β Copilot per-user add-on license
βοΈ Side 2: Azure Cloud (PaaS/IaaS)
For developers building custom AI on Azure.
Scenario clues: "deploy custom Azure AI service", "production workload", "large-scale stable usage"
β Pay-As-You-Go or Prepaid Capacity Commitment
Partner Models in Azure AI Foundry β The Marketplace Trap
| Concept | Explanation |
|---|---|
| What are Partner Models? | 3rd-party AI models (e.g. Meta, Mistral) hosted inside Azure AI Foundry's "cafeteria" β not Microsoft's own models |
| Billing method | Always billed through the Azure Marketplace β NOT through your standard Azure subscription by default |
| Why it fails | Your standard Azure subscription may be locked to direct Microsoft purchases only. Trying to use a partner model gets "declined." |
| The fix | Enable Marketplace purchases on your existing Azure subscription β not creating a new subscription, not buying Copilot Studio packs |
β’ 200 field agents using AI only 1β2x per month β Pay-As-You-Go (Azure)
β’ Custom AI chatbot on Azure website with unpredictable traffic β Pay-As-You-Go (Azure)
β’ Large-scale Azure production AI workload, stable usage, finance wants flat bill β Prepaid Capacity Commitment
LLM vs SLM Β· Fine-tuning vs Pretrained Β· RAG
Model architecture decisions β critical for AB-731
LLM vs SLM β The Transport Analogy
| Feature | π LLM | π΅ SLM |
|---|---|---|
| Brain Power | Deep reasoning, multi-step logic, complex analysis | Specific, straightforward, focused tasks |
| Speed | Slower β takes more compute time | Very fast β low latency |
| Cost | Higher cost per request | Much lower cost per request |
| Where it runs | Requires powerful cloud servers | Can run locally on phones and laptops without internet |
| The Analogy | Semi-truck β hauls massive complex cargo, needs a lot of fuel | Delivery scooter β fast, cheap on fuel, runs anywhere |
| Microsoft Example | GPT-4o via Azure OpenAI | Microsoft Phi series (Phi-3, Phi-4) |
"Complex reasoning" / "multi-step logic" / "synthesise multiple data sources" / "deep analysis of unstructured data" β Always LLM
Pretrained vs Fine-Tuning vs RAG
| Approach | What It Means | Analogy | When to Use |
|---|---|---|---|
| π Pretrained Model | General AI trained on massive internet data. Knows many things but not specialised. | New graduate with a general degree β smart, can talk about anything, not a legal expert yet | General tasks. Good starting point. |
| π Fine-Tuning | Extra training on a specific, smaller dataset. Mathematically adjusts internal "weights" β it becomes a specialist. | Sending the general graduate to law school β focused training until they master that domain | When AI must use specialised terminology reliably (legal, medical, finance). |
| π‘ RAG | Retrieves external data at query time and feeds it as context. Internal weights are NOT changed. | Giving a smart generalist a reference booklet to read before answering each question | When you need the model to know your private documents without modifying the model. |
RAG does NOT modify internal weights = model stays the same, just reads new documents at query time.
Key phrase: "without modifying internal weights" β RAG. "Model must reliably use specialised terminology" β Fine-tuning.
Copilot Tools
Pages, Notebooks, Dynamics 365 BC, and how they compare
Copilot Pages
Turn an AI chat response into an editable document your team can work on side-by-side in real time. Lightweight, fast content creation β proposals, brainstorms.
Equivalentβ Claude Artifact β the side-by-side editing window that opens when Claude generates a document or code
Copilot Notebooks
Bundle multiple files, emails, Pages, and references together. AI reasons across all of it for long, complex projects. Closest to Custom GPT β add custom instructions + upload knowledge files.
Equivalentβ Claude Projects AND β OpenAI Custom GPT
Copilot in Dynamics 365 Business Central
| Feature | What It Does | Nigerian Example |
|---|---|---|
| π AI Item Descriptions | Writes product catalogue descriptions automatically | Lagos FMCG auto-generates descriptions for 5,000 SKUs in minutes |
| π¦ Bank Reconciliation Assist | AI matches bank transactions to ledger entries | Nigerian manufacturer reconciles accounts in minutes, not days |
| π Sales Order Suggestions | Analyses sales history and suggests what to reorder | Kano distributor gets automatic reorder recommendations |
| π° Cash Flow Forecasting | Predicts future cash position from receivables and payables | Lagos trading company CFO sees 90-day cash forecast automatically |
| π’ Marketing Text Suggestions | Generates promotional copy ready for SMS, email, or social | Nigerian retail chain generates SMS content for new product launches instantly |
π Flashcards
Tap a card to flip β work through each set until you know them all
Practice Quiz
50 scenario-based questions β choose how many to attempt
How many questions would you like?
Minimum 2 questions per category guaranteed in every session.
Exam-Day Cheat Sheet
Everything to hold in your head β one page