Microsoft Azure AI-500 Certification Course

You can map a business problem onto the right Azure AI service and say when the answer is not AI at all.

1. The Azure AI service map1 hr
   • Azure OpenAI
   • Azure AI Search
   • AI Foundry and AI Studio
   • Document Intelligence
   • Speech and Vision
2. Build, buy or prompt: choosing an approach1 hr
   • prebuilt APIs
   • custom models
   • prompting
   • fine-tuning
   • when not to use AI
3. Responsible AI principles in practice1 hr
   • fairness
   • reliability and safety
   • privacy and security
   • inclusiveness
   • transparency and accountability
4. Cost model of AI workloads1 hr
   • token pricing
   • PTU vs pay-as-you-go
   • search unit costs
   • storage and egress
   • budget guardrails
5. Subscriptions, regions and model availability1 hr
   • regional model availability
   • quota requests
   • data residency
   • capacity planning
   • fallback regions
6. Setting up the lab environment1 hr
   • resource group layout
   • naming convention
   • service principals
   • Azure CLI
   • cost alerts

Hands-on project: Take three business requests and produce a service selection document for each, including one where the honest recommendation is a rule engine rather than a model.

You can stand up the Azure resources an AI system needs without relying on portal defaults you cannot explain.

1. Resource groups, tags and governance basics1 hr
   • resource organisation
   • tagging for cost
   • locks
   • policy basics
   • subscription limits
2. Storage accounts and data lake layout1 hr
   • blob containers
   • access tiers
   • hierarchical namespace
   • lifecycle rules
   • SAS vs RBAC
3. Key Vault and secret handling1 hr
   • secrets and keys
   • access policies vs RBAC
   • managed identity access
   • rotation
   • soft delete
4. Managed identities and credential-free auth1 hr
   • system assigned
   • user assigned
   • role assignments
   • token acquisition
   • no keys in code
5. Infrastructure as code with Bicep1 hr
   • module structure
   • parameters
   • what-if deployment
   • idempotency
   • CI deployment
6. Logging, diagnostics and Log Analytics1 hr
   • diagnostic settings
   • workspace design
   • KQL basics
   • retention
   • cost per GB

Hands-on project: Provision a complete AI workload foundation as Bicep templates: resource groups, storage, Key Vault, managed identity and logging.

You can route every call to an AI service over a private network path and prove no traffic leaves it.

1. Virtual networks, subnets and address planning1 hr
   • subnet delegation
   • address planning
   • NSG rules
   • service endpoints
   • route tables
2. Private endpoints and private DNS2 hrs
   • private endpoint creation
   • private DNS zones
   • DNS resolution path
   • hub DNS forwarding
   • verification
3. Disabling public network access1 hr
   • public access toggle
   • firewall rules
   • allowed IP ranges
   • breaking glass
   • testing the block
4. Entra ID authentication for AI services1 hr
   • token-based auth
   • RBAC roles for OpenAI
   • key rotation vs identity
   • scoped access
   • audit trail
5. Data privacy and the model boundary1 hr
   • data retention settings
   • abuse monitoring opt-out
   • customer-managed keys
   • prompt logging control
   • contractual terms
6. Threat modelling an AI deployment1 hr
   • trust boundaries
   • prompt injection surface
   • data exfiltration paths
   • privilege escalation
   • mitigation register

Hands-on project: Deploy Azure OpenAI and AI Search reachable only through private endpoints inside a virtual network, and demonstrate that public access fails.

You can deploy models with capacity that will survive a launch, and diagnose a 429 before your users report it.

1. Model families and selecting one1 hr
   • reasoning models
   • general models
   • small fast models
   • embedding models
   • deprecation cadence
2. Deployments, quotas and tokens per minute1 hr
   • deployment creation
   • TPM allocation
   • quota requests
   • regional limits
   • multiple deployments
3. Provisioned throughput versus pay-as-you-go1 hr
   • PTU sizing
   • latency guarantees
   • cost crossover point
   • burst handling
   • commitment terms
4. Calling the API correctly1 hr
   • chat completions
   • streaming
   • system messages
   • parameter tuning
   • SDK versus REST
5. Rate limits, retries and fallback1 hr
   • 429 handling
   • exponential backoff
   • regional fallback
   • queueing
   • graceful degradation
6. Measuring latency and cost per request1 hr
   • token accounting
   • time to first token
   • p95 latency
   • cost attribution
   • dashboarding

Hands-on project: Deploy three model variants with sized quotas, build a client with retry and fallback, and load-test it to find the throughput ceiling.

You can get schema-valid output from a model reliably enough to put it in a pipeline that nothing watches.

1. System messages and instruction design1 hr
   • role framing
   • constraint phrasing
   • refusal conditions
   • output contracts
   • instruction conflicts
2. Few-shot examples and grounding1 hr
   • example selection
   • format anchoring
   • grounding with context
   • citation prompting
   • example rotation
3. Structured output and JSON schema2 hrs
   • JSON mode
   • schema enforcement
   • validation layer
   • parse recovery
   • nested structures
4. Function and tool calling on Azure OpenAI1 hr
   • tool definitions
   • argument validation
   • parallel calls
   • tool result formatting
   • selection debugging
5. Controlling cost and verbosity1 hr
   • max tokens
   • prompt compression
   • caching
   • model downgrade paths
   • token budgeting
6. Prompt versioning and regression testing1 hr
   • prompt registry
   • version pinning
   • regression set
   • A/B comparison
   • rollback

Hands-on project: Build an extraction service that returns valid JSON for a thousand varied documents with a measured failure rate under one percent.

You can index a document corpus into Azure AI Search and retrieve from it, understanding each stage rather than copying a quickstart.

1. Search service tiers and capacity1 hr
   • tier selection
   • replicas and partitions
   • storage limits
   • query throughput
   • cost model
2. Index schema design1 hr
   • field attributes
   • analyzers
   • complex types
   • scoring profiles
   • schema evolution
3. Indexers, data sources and skillsets1 hr
   • blob indexer
   • change detection
   • built-in skills
   • custom skills
   • enrichment pipeline
4. Chunking strategy for retrieval1 hr
   • chunk size
   • overlap
   • heading-aware splits
   • metadata attachment
   • parent-child chunks
5. Vector fields and embedding integration1 hr
   • vector field config
   • dimension matching
   • integrated vectorisation
   • embedding model choice
   • upsert flow
6. Measuring retrieval quality1 hr
   • labelled query set
   • recall at k
   • precision at k
   • latency percentiles
   • regression tracking

Hands-on project: Index a ten-thousand document corpus with an appropriate schema and skillset, and measure retrieval recall on a labelled query set.

You can combine keyword and vector search with reranking and show, with numbers, how much each layer added.

1. Keyword search, analyzers and BM251 hr
   • tokenisation
   • stemming
   • stop words
   • synonym maps
   • scoring
2. Vector search and approximate nearest neighbour1 hr
   • HNSW parameters
   • exhaustive KNN
   • vector compression
   • recall tuning
   • index size
3. Hybrid search and reciprocal rank fusion2 hrs
   • hybrid queries
   • RRF scoring
   • weighting
   • filter interaction
   • result diversity
4. Semantic ranking and captions1 hr
   • semantic ranker
   • captions and answers
   • reranking depth
   • cost per query
   • when it helps
5. Security filters and per-user results1 hr
   • document-level security
   • group filters
   • trimming at query time
   • leakage testing
   • performance cost
6. Scaling, sharding and query performance1 hr
   • partition strategy
   • replica scaling
   • query latency
   • throttling
   • capacity planning

Hands-on project: Improve retrieval recall on a labelled set by at least fifteen points using hybrid search and semantic ranking, attributing each gain.

You can build a retrieval-augmented system that cites its sources, refuses when ungrounded, and survives a document corpus that keeps changing.

1. End-to-end RAG architecture on Azure2 hrs
   • component map
   • data flow
   • private networking
   • identity flow
   • failure modes
2. Ingestion and incremental indexing2 hrs
   • change detection
   • upsert and delete
   • versioned documents
   • reindex strategy
   • backfill
3. Context assembly and token budgeting2 hrs
   • retrieval depth
   • ordering
   • deduplication
   • budget allocation
   • overflow handling
4. Grounded generation and citations2 hrs
   • citation prompting
   • span attribution
   • quote fidelity
   • source rendering
   • hallucination guards
5. Refusal, confidence and clarification1 hr
   • no-answer detection
   • confidence thresholds
   • clarifying questions
   • escalation
   • user messaging
6. Multi-tenant RAG and data isolation1 hr
   • per-tenant indexes
   • filter-based isolation
   • leakage testing
   • noisy neighbour
   • cost attribution

Hands-on project: Build an end-to-end RAG application over a live corpus with citations, incremental reindexing, grounded refusal and a measured accuracy baseline.

You can prove a change improved the system, which is the difference between engineering it and fiddling with it.

1. Designing an evaluation set1 hr
   • case sampling
   • labelling guidance
   • edge cases
   • difficulty spread
   • maintenance
2. Groundedness and faithfulness scoring1 hr
   • claim extraction
   • support checking
   • unsupported spans
   • citation accuracy
   • hallucination rate
3. Relevance, coherence and completeness1 hr
   • rubric design
   • relevance scoring
   • completeness checks
   • format compliance
   • verbosity penalty
4. Azure AI evaluation tooling1 hr
   • built-in evaluators
   • custom evaluators
   • batch evaluation runs
   • result storage
   • comparison views
5. Model-as-judge and its biases1 hr
   • judge prompts
   • position bias
   • verbosity bias
   • calibration against humans
   • when to distrust it
6. Continuous evaluation in pipelines1 hr
   • pipeline integration
   • pass thresholds
   • cost ceilings
   • trend tracking
   • blocking a release

Hands-on project: Build a hundred-case evaluation suite covering groundedness, relevance and retrieval recall, running automatically in a pipeline.

You can build an assistant that takes actions through tools, with budgets and guardrails that stop it running away.

1. From RAG to agents: what changes1 hr
   • read-only vs acting
   • autonomy levels
   • failure blast radius
   • approval requirements
   • when an agent is overkill
2. Tool design for enterprise systems1 hr
   • schema design
   • scoped credentials
   • idempotency
   • error surfaces
   • audit logging
3. The agent loop and step budgets1 hr
   • observe act repeat
   • termination conditions
   • step limits
   • loop detection
   • timeout policy
4. Orchestration patterns1 hr
   • sequential
   • router
   • supervisor
   • parallel branches
   • coordination cost
5. Human-in-the-loop and approval gates1 hr
   • interrupt points
   • approval UI
   • destructive action gating
   • dry run mode
   • audit before execute
6. Tracing and debugging agent runs1 hr
   • span hierarchy
   • tool call inspection
   • replay
   • failure taxonomy
   • cost per run

Hands-on project: Build a multi-tool agent over enterprise systems with step budgets, approval gates for destructive actions and a trace for every run.

You can build, evaluate and deploy a flow as a managed endpoint, and version it like any other piece of software.

1. AI Studio projects and workspace structure1 hr
   • hubs and projects
   • connections
   • compute
   • model catalogue
   • access control
2. Prompt Flow authoring1 hr
   • flow graph
   • LLM nodes
   • Python nodes
   • inputs and outputs
   • connections
3. Variants and systematic comparison1 hr
   • variant definition
   • batch runs
   • metric comparison
   • winner selection
   • promotion
4. Batch evaluation runs1 hr
   • dataset upload
   • evaluator selection
   • run comparison
   • result export
   • regression gates
5. Deploying to managed online endpoints1 hr
   • endpoint creation
   • deployment config
   • traffic splitting
   • autoscale
   • authentication
6. CI/CD for flows1 hr
   • flow in source control
   • pipeline deployment
   • environment promotion
   • rollback
   • drift detection

Hands-on project: Build a Prompt Flow pipeline with variants, run a batch evaluation, and deploy the winning variant to a managed online endpoint.

You can use the prebuilt services where they beat a language model on cost, latency and accuracy, and know where they do not.

1. Document Intelligence and form extraction1 hr
   • prebuilt models
   • custom extraction models
   • training data
   • confidence scores
   • table extraction
2. Azure AI Vision1 hr
   • image analysis
   • OCR
   • spatial analysis
   • custom vision models
   • batch processing
3. Speech to text and text to speech1 hr
   • real-time transcription
   • batch transcription
   • diarisation
   • custom speech
   • neural voices
4. Language service capabilities1 hr
   • named entity recognition
   • PII detection
   • sentiment
   • summarisation
   • custom classification
5. Translation and multilingual pipelines1 hr
   • text translation
   • document translation
   • custom terminology
   • language detection
   • quality limits
6. Choosing prebuilt versus a language model1 hr
   • accuracy comparison
   • cost per unit
   • latency
   • determinism
   • maintenance burden

Hands-on project: Build a document processing pipeline combining Document Intelligence, Language and a model, and justify each component on cost and accuracy.

You can decide between prompting, retrieval and fine-tuning with evidence, and bring the cost of a working system down without breaking it.

1. The decision framework1 hr
   • knowledge vs capability
   • data volume
   • maintenance cost
   • freshness
   • decision rubric
2. Fine-tuning on Azure OpenAI1 hr
   • dataset format
   • training job
   • hyperparameters
   • hosting cost
   • deployment
3. Evaluating a fine-tune honestly1 hr
   • held-out set
   • baseline comparison
   • general ability regression
   • overfit signals
   • cost per point
4. Cost levers that actually work1 hr
   • model routing
   • prompt caching
   • context trimming
   • batching
   • smaller models
5. Caching strategies1 hr
   • exact cache
   • semantic cache
   • invalidation
   • hit rate measurement
   • staleness risk
6. Capacity planning and PTU economics1 hr
   • demand modelling
   • PTU sizing
   • crossover analysis
   • burst strategy
   • commitment risk

Hands-on project: Cut the cost per request of a working RAG system by at least half while keeping accuracy within two points, documenting every lever used.

You can operate an AI system responsibly: filtering unsafe output, detecting drift, logging for audit and answering a compliance reviewer.

1. Azure AI Content Safety1 hr
   • harm categories
   • severity thresholds
   • prompt shields
   • groundedness detection
   • custom blocklists
2. Monitoring quality and drift in production1 hr
   • quality sampling
   • drift signals
   • index staleness
   • feedback capture
   • alerting
3. Observability and tracing1 hr
   • Application Insights
   • custom telemetry
   • token metrics
   • trace correlation
   • dashboards
4. Audit logging and compliance evidence1 hr
   • request logging
   • PII redaction
   • retention policy
   • access audit
   • evidence index
5. Incident response for AI systems1 hr
   • harmful output incidents
   • rollback procedure
   • model deprecation
   • communication
   • post-mortem
6. Responsible AI documentation1 hr
   • impact assessment
   • model cards
   • transparency notes
   • human oversight design
   • sign-off

Hands-on project: Produce a governance package for a deployed system: content safety configuration, monitoring dashboards, audit logging, incident runbook and a risk register.

You can deliver a complete, secured, evaluated and governed AI platform on Azure, and answer scenario questions under timed conditions.

1. Architecture, scope and success metrics2 hrs
   • requirement extraction
   • architecture diagram
   • metric selection
   • non-goals
   • cost ceiling
2. Secure foundation and private networking2 hrs
   • Bicep deployment
   • private endpoints
   • managed identity
   • DNS verification
   • public access proof
3. Ingestion, retrieval and grounded generation2 hrs
   • indexing pipeline
   • hybrid retrieval
   • semantic ranking
   • citations
   • tenant isolation
4. Agent layer and Prompt Flow deployment2 hrs
   • tool inventory
   • approval gates
   • flow variants
   • endpoint deployment
   • traffic split
5. Evaluation, safety, monitoring and cost2 hrs
   • eval suite
   • content safety config
   • dashboards
   • cost per query
   • drift alerts
6. Governance pack, timed exam and defence2 hrs
   • risk register
   • responsible AI docs
   • runbook
   • timed scenario exam
   • panel defence

Hands-on project: Build and defend a full enterprise AI platform: private-networked Azure OpenAI and AI Search, incremental ingestion over a real corpus, hybrid retrieval with semantic ranking, grounded RAG with citations and per-tenant isolation, an agent layer with approval gates, Prompt Flow deployment, a hundred-case evaluation suite in a pipeline, content safety, monitoring dashboards, cost model and governance pack — plus a timed scenario exam.