Scope Overreach — Azure AI Engineer (AI-300)
You solved a broader problem than what was asked. The scenario had specific constraints — you addressed requirements that weren't there.
Azure OpenAI Is Not Always the Right Tool
When a question describes a batch document classification task with a fine-tuned BERT model already in production, reaching for Azure OpenAI Service feels modern and capable. But the scenario has no requirement for generative output, semantic flexibility, or foundation model behavior. Deploying to Azure Machine Learning endpoints satisfies the requirement with less cost and less complexity. The exam penalizes answers that upgrade the solution beyond what the problem demands.
The Scenario
A team needs to restrict access to a Storage Account so only resources in their virtual network can reach it. You design Azure Private Link for private IP connectivity, Azure Firewall for traffic inspection, and Network Security Groups for subnet-level filtering. The correct answer is enabling a Service Endpoint on the VNet subnet and adding a network rule on the Storage Account to allow only that VNet. The scenario said "restrict to VNet" — not "implement zero-trust network architecture." Service Endpoints are free, require no DNS changes, and take 30 seconds to configure.
How to Spot It
- •Azure Service Endpoints vs. Private Endpoints is the most common scope overreach test. Service Endpoints (free, VNet-scoped access restriction) are correct when you just need to limit network access. Private Endpoints ($0.01/hour, private IP in your VNet) are correct when you need private IP addressing, on-premises access via VPN/ExpressRoute, or cross-region private connectivity.
- •Azure Firewall costs ~$912/month base. If the scenario does not mention traffic inspection, threat intelligence, or centralized network security policy, adding Azure Firewall is scope overreach.
- •The exam tests whether you solve exactly the stated problem. If your answer addresses security threats, compliance requirements, or scaling concerns the scenario never mentioned, you have expanded scope beyond what was asked.
Decision Rules
Whether to assign a narrow data-plane storage role scoped to the workspace's storage account, or to satisfy access by elevating scope to the resource group or subscription with a broader management role.
Which combination of configurations—network isolation control and GitHub Actions authentication mechanism—satisfies both constraints at the minimum necessary scope, without extending RBAC beyond the workspace resource group or adding managed VNet infrastructure not required by the stated threat model?
Assign the AzureML Data Scientist built-in role scoped to the workspace plus Storage Blob Data Reader scoped to the specific container, rather than assigning a single broad role (e.g., Contributor) at subscription or resource group scope.
Whether to enforce Foundry inference-endpoint network isolation using a scoped Private Endpoint paired with public-network-access disabled on the Foundry resource (least complexity, constraint satisfied) versus a hub-spoke topology with Azure Firewall and custom route tables that overreaches the single-project boundary requirement.
Should the RBAC role assignment be scoped to the Foundry project resource group with a read-only role, or to the subscription with a broader role — and which pairing is the only one that satisfies least-privilege without granting cross-project write access?
Whether to assign a Private Endpoint scoped to the Foundry project's resource group VNet (satisfying network-isolation at the narrowest viable scope) versus deploying a broader network security appliance or hub-spoke topology that exceeds the constraint and introduces unjustified operational complexity.
Domain Coverage
Difficulty Breakdown
Related Patterns