AI-300 · Associate
AI-300 Cheat Sheet 2026: Quick Reference
Every MLOps and GenAIOps domain condensed to a table, plus a quality-metrics guide and a deployment/throughput decision table for the exam's most-confused pairs.
Exam Snapshot
Questions: ~40
Duration: 100 minutes
Passing score: 700 / 1000
Cost: $165 USD
Level: Associate
Last updated: 2026-08-04 (no dated skills-measured revision published)
1. MLOps Infrastructure (15-20%)
| Concept | Area | Quick note |
|---|---|---|
| Workspace, datastores, compute | ML workspace basics | Core resources every workspace needs |
| Identity & access for workspaces | RBAC scoping | Isolate teams within one shared workspace |
| Data assets, environments, components | Workspace assets | Reusable, versioned building blocks |
| Registries | Cross-workspace sharing | Share assets across multiple workspaces |
| GitHub integration | Secure access config | Connects source control to the workspace |
| Bicep / Azure CLI deployment | Infrastructure as code | Version-controlled, repeatable workspace setup |
| GitHub Actions | Automated provisioning | CI/CD for resource provisioning |
| Network restriction | Workspace network access | Limits who/what can reach the workspace |
| Git source control | ML project versioning | Standard version control for ML project code |
2. ML Model Lifecycle & Operations (25-30%)
| Concept | Area | Quick note |
|---|---|---|
| MLflow experiment tracking | Training orchestration | Log and compare metrics across training runs |
| AutoML | Training orchestration | Automatically explores candidate models |
| Hyperparameter tuning | Training orchestration | Automated search across parameter combinations |
| Distributed training | Training orchestration | For large or deep learning models |
| Training pipelines | Training orchestration | Repeatable, automated training runs |
| MLflow model registration | Registration & versioning | Registers a trained model with version history |
| Responsible AI evaluation | Registration & versioning | Checked before/at registration, not after |
| Model archiving | Registration & versioning | Manages the end of a model version's lifecycle |
| Real-time vs. batch endpoints | Production deployment | Real-time = single predictions; batch = bulk, latency-tolerant |
| Progressive rollout / safe rollback | Production deployment | Shift traffic gradually; revert quickly if needed |
| Data drift detection | Production monitoring | Catches model degradation with no code change |
| Retraining / alert triggers | Production monitoring | Automated response when thresholds are crossed |
3. GenAIOps Infrastructure (20-25%)
| Concept | Area | Quick note |
|---|---|---|
| Foundry resources & project environments | Platform setup | Base configuration for GenAIOps work |
| Managed identity + RBAC | Access control | No password-based access to Foundry resources |
| Private networking | Network security | Keeps traffic off the public internet |
| Bicep deployment | Infrastructure as code | Same IaC pattern as MLOps, applied to Foundry |
| Serverless API endpoints | Foundation model deployment | Best for variable, unpredictable load |
| Managed compute | Foundation model deployment | Dedicated compute for a foundation model |
| Provisioned throughput units | Foundation model deployment | Predictable latency at known, fixed volume |
| Model versioning & deployment strategy | Production rollout | Controlled promotion of foundation model versions |
| Prompt design & variants | Prompt engineering | Comparing prompt performance systematically |
| Git version control for prompts | Prompt engineering | Prompts are versioned like code |
4. GenAI Quality Assurance & Observability (10-15%)
| Concept | Area | Quick note |
|---|---|---|
| Test datasets & data mapping | Evaluation setup | Foundation for any quality evaluation run |
| Groundedness / relevance / coherence / fluency | Quality metrics | See the quick reference table below |
| Risk & safety evaluations | Evaluation setup | Detects harmful content, not just quality issues |
| Automated evaluation workflows | Evaluation setup | Built-in and custom metrics, run automatically |
| Continuous monitoring in Foundry | Observability | Ongoing, not a one-time check |
| Latency, throughput, response time | Observability | Performance metrics |
| Token consumption, resource usage | Observability | Cost metrics |
| Logging, tracing, debugging | Observability | Production troubleshooting capability |
5. GenAI Optimization (10-15%)
| Concept | Area | Quick note |
|---|---|---|
| Similarity thresholds & chunk sizes | RAG tuning | Direct levers for retrieval accuracy |
| Embedding model selection/fine-tuning | RAG tuning | Domain-specific accuracy improvement |
| Hybrid search | RAG tuning | Combines semantic and keyword-based retrieval |
| Relevance metrics & A/B testing | RAG tuning | How RAG performance is actually evaluated |
| Advanced fine-tuning methods | Model customization | Beyond basic fine-tuning |
| Synthetic data creation | Model customization | For fine-tuning when real data is limited |
| Fine-tuned model monitoring | Model customization | Ongoing, not just at deployment |
| Dev-to-production management | Model customization | A managed path, not a one-time promotion |
Quality Metrics Quick Reference
Four metrics that sound similar but test completely different failure modes:
| Metric | Measures | Fails when… |
|---|---|---|
| Groundedness | Factual support from source content | Response includes unsupported or fabricated details |
| Relevance | How on-topic the response is | Response drifts from what was actually asked |
| Coherence | Internal logical consistency | Response contradicts itself or reads disjointed |
| Fluency | Grammatical, natural language quality | Response is awkward, broken, or unnatural |
Deployment & Throughput Quick Decision
| Scenario | Use this |
|---|---|
| Unpredictable, variable request volume | Serverless API endpoint |
| Fixed, known high-volume workload needing predictable latency | Provisioned throughput units |
| Testing a new model version on a small % of traffic first | Progressive rollout |
| Need to undo a bad deployment fast | Safe rollback strategy |
| Real-time, single predictions | Real-time endpoint |
| Large batch of predictions, latency not critical | Batch endpoint |
Common Questions
Is AI-300 about building models or operating them in production?
Operating them. AI-300 tests MLOps and GenAIOps — deployment, monitoring, and optimization of systems that already exist, not model design or data science fundamentals.
What's the difference between groundedness and relevance?
Relevance is whether a response is on-topic. Groundedness is whether it's factually supported by the source content. A response can be relevant and still ungrounded if it invents unsupported details.
What's the fastest way to review right before the exam?
Scan the five domain tables plus the quality metrics and deployment decision tables — those two distinctions account for a large share of missed questions.
Can I print this AI-300 cheat sheet?
Yes — every table here is plain HTML, so a browser print or "print to PDF" renders cleanly.
MSCertQuiz sells practice-exam access for AI-300 and other Microsoft certifications; this cheat sheet is written by the same team that builds those questions.
Related Resources
Free AI-300 Practice Questions
10 scenario questions with full rationale, by domain.
AI-200 Study Guide
A related Azure AI developer certification.
DP-900 Study Guide
A data-fundamentals complement to this operations-focused exam.
AI-300 Certification Overview
Full exam details and the complete 500-question practice bank.
Reviewed the cheat sheet? Now drill it.
Start with 40 free AI-300 questions covering every domain above.
Start Free AI-300 Practice