Level 1: Initial¶
Last updated: 2026-07-27
List of references
At Level 1, teams explore model capabilities without mature operating processes. The goal is to make experiments intentional: define an approved use case, test the model and prompt behavior, and collect enough telemetry to understand cost, latency, failures, and output quality.
Source: Azure AI Foundry documentation.
Establish the foundation¶
| Activity | Minimum practice |
|---|---|
| Use case | Document the user need, intended use, known limitations, and human escalation path. |
| Model access | Deploy approved models through Azure AI Foundry or Azure OpenAI Service with controlled access. |
| Prompt design | Keep prompts in source control and test representative input, failure, and safety cases. |
| Data | Do not send confidential data without an approved classification and handling design. |
| Telemetry | Capture request rate, latency, failures, token use, and user feedback. |
Explore and compare models¶
Use the Azure AI Foundry model catalog to identify models that meet the use case, then compare quality, latency, cost, input limits, and regional availability. Maintain a small test dataset that reflects expected user requests and difficult cases rather than selecting a model only from generic benchmarks.
Source: Model catalog in Azure AI Foundry.
Basic operational measures¶
- Quality: human review or task-specific scoring of generated responses.
- Reliability: request success rate, timeout rate, and dependency failures.
- Experience: end-to-end latency and user feedback.
- Cost: token consumption, model usage, and supporting service cost.
- Safety: blocked or flagged requests, unsafe output findings, and escalation outcomes.
Continue to Level 2: Defined when experiments need a repeatable development and deployment process.