The Anatomy of an Azure AI Solution
Level Up Team
This post summarizes The Anatomy of an Azure AI Solution, module 1.2 of Level Up’s Azure AI Fundamentals series. The lesson is on YouTube as Azure AI Fundamentals AI-901 Exam: 1.2: Anatomy of an Azure AI Solution. Module 1.1 is What AI-901 Actually Tests.
A running example
The lesson uses a Contoso customer-support assistant. A customer can submit a typed question, a spoken description, a product image, or an invoice or document. The example request is: “Why was I charged twice, and can you correct it?” The customer attaches an invoice and a screenshot.
Seven parts of an AI solution
- Input modality — what the user provides
- AI workload — what the system must accomplish
- Model — the intelligence selected for the task
- Deployment — an accessible instance of the model
- Prompt or application — instructions and orchestration logic
- Output — what the solution returns or does
- Evaluation and controls — how the team measures quality and manages risk
Every AI-901 scenario can be unpacked with the same seven questions: What entered the system, what workload was required, which model handled it, where was it deployed, what directed it, what came out, and how was the result controlled?
The input does not determine the workload
Modality describes the form of the data. Workload describes what the solution must do with it.
| Customer input | Required workload |
|---|---|
| Typed complaint | Text analysis or generative AI |
| Spoken question | Speech recognition, followed by language processing |
| Product photograph | Computer vision or multimodal generation |
| Invoice document | Information extraction |
| Request for account action | Agentic AI |
A model becomes usable through a deployment
A model defines the available capabilities. It may support text, vision, audio, or multiple modalities, and models vary in quality, speed, context capacity, and cost.
A deployment makes a selected model available to an application. It provides an endpoint or named target, applies configuration and access controls, and can use parameters that affect model behavior.
The model is the design. The deployment is the running instance the application can call. The path is model catalog → deployed endpoint → application.
The application turns capability into a solution
A system prompt defines the assistant’s role, boundaries, and expected behavior: “You are a customer-support assistant. Use verified account information and explain any uncertainty.”
A user prompt contains the customer’s current request and relevant context.
Application logic collects and prepares the inputs, calls models or Foundry Tools, supplies retrieved or extracted information, invokes an agent when the task requires action, and presents the result to the customer.
The agent manages a multi-step request
The customer goal is: “Explain the duplicate charge and correct it if appropriate.”
Possible agent behavior:
- Read the customer’s request
- Use extracted invoice data
- Retrieve the relevant account transactions
- Compare the invoice with the account record
- Request an approved refund action when warranted
- Explain the result to the customer
The agent components in the lesson are instructions, a model, knowledge or context, tools, and conversation state. The example tools and context are invoice data, account records, and a refund tool.
A successful response must be useful and controlled
A useful reply in this scenario sounds like: “The invoice contains one order, but the account shows two completed charges. I prepared a refund request for the duplicate charge. Please confirm before submission.”
Evaluate the solution by asking whether it identified the duplicate correctly, used the supplied evidence, chose an authorized action, explained the result clearly, and how often it produces unsupported claims.
| Concern | Example control |
|---|---|
| Fairness | Test performance across supported users and languages |
| Reliability and safety | Validate extracted values before taking action |
| Privacy and security | Limit access to account and document data |
| Transparency | Tell the customer when AI generated the response |
| Accountability | Log decisions and require approval for refunds |
| Inclusiveness | Support accessible text and speech interactions |
Bring this to your team
Level Up can run a personalized Azure AI Fundamentals workshop for your team. Tell us a little about your group and we’ll reply with pricing.
Loading the form…