AI agents are becoming a key enabler for businesses looking to streamline processes, automate repetitive tasks, and empower their employees to work more efficiently. In Microsoft 365 Copilot, we’ve already seen a lot of solutions that focus on improving productivity for individuals. Yet, the potential for AI-driven automation goes much further when you can connect intelligent , natural language, agents directly to your own business data and processes—enter Azure AI Agent Service in Azure AI Foundry.
In this post, you read about why AI Agents are valuable, and how Azure AI Agent Service makes it easy to build and customize these agents. I also did some testing and share those experiences, including the Code Interpreter feature for data analysis. Finally, in the end you can read through some quick tips on how you can get started.
What Are AI Agents and Why Should Businesses Care?Stories of Transforming Business ProcessesAI AgentsAzure AI Agent Service is now in Public PreviewMy Testing Experience with Code Interpreter and Knowledge SourcesCode InterpreterExternal Knowledge & RAG TestingAgent Configuration Made SimpleModel Tuning Options and Limitations Getting Started: The Quickstart PrerequisitesCreating and Testing Your AgentExpand with SDK or Additional ToolsConclusionRead more from these sources:About writing this article
At its simplest, an AI agent is a self-contained “microservice” powered by a large language model (LLM) or similar AI model. It’s designed to answer questions, perform actions, and ultimately automate or augment specific tasks. Let’s take a look at some examples first.
Stories of Transforming Business Processes
A Fictional Look at a Multi-Agent Sales EcosystemImagine a large electronics retailer that operates in dozens of countries. They have separate specialized AI agents for different aspects of the sales cycle. One AI agent handles lead qualification by scanning incoming inquiries and extracting key information about prospective clients. Simultaneously, another agent is responsible for product recommendations based on real-time pricing and inventory data. Once a lead is qualified, a scheduling agent sets up demos with sales reps and automatically books a meeting in Microsoft Teams, complete with relevant documents attached.
These agents can also collaborate, passing information to one another about lead status or the best product bundles for a particular region. Thanks to this multi-agent approach, a sales rep can jump in only when human interaction is truly needed, rather than juggling repetitive tasks like re-checking stock or manually coordinating meetings. The entire process is a well-orchestrated system that frees employees from repetitive admin and data entry. The result: more personalized interactions with customers, higher sales velocity, and a drastically reduced chance of human error.
Fujitsu: RAG and Sales Efficiency in the Real WorldWhile the above scenario is hypothetical, real-life businesses already leverage Azure AI Agent Service to revolutionize their sales processes. For example, Fujitsu leveraged Azure AI Agent Service and Semantic Kernel to build an AI-powered automation solution to streamline proposal creation, enabling sales teams to focus on high-value customer engagement. The AI agent dynamically retrieves and synthesizes data from dispersed sources, ensuring accuracy and relevance while integrating seamlessly into Fujitsu’s existing Microsoft ecosystem. “We are using Microsoft’s Semantic Kernel and Azure AI Agent Service to orchestrate multiple specialized AI agents and an orchestrator AI to coordinate them to answer questions as a team,”
Cineplex: Transforming Customer Service Through AutomationCineplex, a leading Canadian media and entertainment company, has transformed its customer service operations using AI-powered automation. One of its biggest challenges was handling refund requests, a time-consuming process that took 5–15 minutes per request. To solve this, Cineplex implemented an AI Copilot agent using Microsoft Power Platform and Azure AI. Now, guest service agents simply input a booking ID and date, and the AI handles the rest—retrieving data, validating the request, and completing the refund in about 30 seconds This highlights how AI agents can optimize routine tasks, boost efficiency, and enhance customer service—without replacing human interaction. For businesses looking to scale support operations, AI-powered automation can be a game-changer.
AI Agents
Unlike basic chatbots, AI agents can incorporate context from historical conversations and connect to external systems, allowing them to:
Search your company’s knowledge base or the web.
Process and interpret files or real-time data.
Make calculations, generate reports, even run code.
Perform complex tasks that save employees time and effort.
Integrate with other systems, internal and external
Engage other agents in the process: multi-agent systems
Cope a lot better with various situations better than traditional automation. This is due to LLM in their “core” giving understanding of the goal and what is needed. AI Agents are flexible and can adapt to situations – and also know when to ask help from an another agent or a real person.
In other words, AI agents can complement human workers by taking on repetitive or time-consuming jobs. That might mean a customer support agent that automatically retrieves answers from a knowledge base, an internal finance agent that crunches budget data from spreadsheets, or a sales agent that triggers email workflows.
From a business standpoint, AI agents have tangible benefits:
Accelerated decision-making: Dynamic, context-aware AI reduces manual research.
Scalability: Agents can work around the clock, handling tasks for multiple teams simultaneously.
Consistency: They apply knowledge and logic in a uniform way—fewer mistakes due to human error. Today is the time when we experiment with agents, and it needs to be realized agents can also make mistakes – sometimes even plenty.. The goal is in the consistency and coming up with new ideas where AI can transform the process, and these won’t be reached without experimenting, coming up with challenging use cases and courage to try out something new.
Azure AI Agent Service, now available as public preview in the Azure AI Foundry portal, provides a managed environment to build, debug, and deploy these AI agents. It’s designed so that developers and tech-savvy business users can quickly shape an agent’s capabilities without having to assemble all the underlying code or infrastructure themselves. This speeds up pro-code agent development and is yet an another example of fusion teams where business and developers work together.
Key capabilities include:
Ready-to-Go Tools & Integrations:
Code Interpreter: Allows agents to execute Python code within a secure sandbox—great for number-crunching, data analysis, or generating graphs.
Bing Search & Azure AI Search: Agents can pull in external knowledge from the web or your own data, adding relevant context to tasks.
Azure Functions Support (SDK-based): Developers can expose custom business logic or external APIs to the agent, letting it trigger real-world actions.
Conversational Memory:Agents can maintain a thread of conversation, remember details, and continue where you (or the agent itself) left off. This is handled securely on the server side.
Multiple Model Options:Although Microsoft’s GPT-4o is a popular choice, you can also deploy other partner models like Cohere or Mistral in the Azure AI Foundry. (Note: Mistral-large-2407 is becoming legacy and may not be available much longer.)
Basic vs. Standard Setup:
Basic Setup (supported in the Azure AI Foundry portal today): You rely on Microsoft-managed resources for storage and search. Quick to start, minimal overhead, but it offers less control.
Standard Setup (Bicep template–only): You bring your own resources (like Azure Storage and Azure AI Search) for complete visibility and cost management.
I’ve spent some time exploring the new Agents UI in Azure AI Foundry, putting these features to the test. Here’s what stood out for me:
Code Interpreter
A fun (though fictional) scenario was exploring the terminal velocity of a laptop falling from an airplane. With the agent’s Code Interpreter tool enabled, I could ask the agent to run physics-related calculations. It can generate quick math scripts in Python—and this is just a simple example about the Code Interpreter.
I uploaded an Excel file for the 2023 budget of the city of Vantaa (available as open data) to the Code Interpreter. The agent then read and interpreted the file, making it straightforward to analyze budget figures, gather insights, and visualize the data.
External Knowledge & RAG Testing
I also tested a scenario using basic RAG (retrieval augmented generation). By uploading some demo documents, the agent was able to pull targeted facts from my own content, weaving them into its answers. The RAG with AI isn’t anything new anymore, but the Assistants API working behind the hood isn’t an everyday tool yet. So it made sense to play around to see how it performs – and it was just like I expected.
Agent Configuration Made Simple
The visual flow in the Azure AI Foundry UI is deceptively simple: define your agent’s name, add Knowledge sources (files or indexes), and specify which Actions (tools) the agent may use.
Currently, the only action available from the UI is Code Interpreter. If you want to integrate your own, such as Azure Functions, you can do more via the SDK.
Model Tuning Options and Limitations
Basic tuning for Temperature and Top P is easily accessible in the UI, so you can adjust how creative or deterministic your agent’s answers should be.
In the UI, only Code Interpreter is displayed as an “action,” but the underlying Assistants API definition is flexible—new actions or custom tools can be added once they are enabled.
I experimented with GPT-4o, which worked seamlessly. The service also promises support for non-OpenAI models like Cohere and Mistral, though my free Azure subscription didn’t allow me to deploy them.
Overall, these tests highlight how quickly you can piece together a specialized AI agent that’s unique to your brand, team, or project. With a few lines of code or a few clicks in the UI, you can transform a simple chat model into a mini-assistant with real business value.
Want to try it yourself? Here’s a short guide based on the official quickstart.
Prerequisites
An Azure subscription (create a free trial if needed).
The Azure AI Developer role assigned. This gives you the right permissions to create and manage AI agents.
Basic Setup via the Azure AI Foundry Portal: Because the Foundry portal only supports the “basic setup,” you’ll be using Microsoft-managed storage and search behind the scenes. This gets you getting started fast.
Creating and Testing Your Agent
Navigate to Agents in the Azure AI Foundry portal and select “New agent.”
Provide a name and add instructions (e.g., “You are a business analyst specializing in forecasting.”).A tip: use Chat Playground’s Generate prompt feature to build instructions for the agent.
Under “Knowledge & Action,” add Code Interpreter if you want the agent to handle data analysis or code execution. You can also attach up to 20 files that your agent can read and use for generating outputs.
After configuring your agent, switch to the “Playground” to begin chatting.
You can revise instructions, tweak model parameters (Temperature, Top P), or add new knowledge files and tools.
Confirm that your agent is responding as expected and refine your instructions or data sources if needed.
To incorporate your own Azure Functions or external APIs, you’ll need to define them as tools via the Azure AI Foundry SDK or the Azure OpenAI SDK. This is particularly helpful for more complex automations where the agent might, for instance, update a CRM record or send an email on your behalf.
Azure AI Agent Service is a promising step forward in automating diverse business processes—from data analysis and RAG queries to more action-oriented tasks like connecting to external APIs. The combination of large language models, integrated tools, and simple setup in the Azure AI Foundry UI makes it a compelling choice for trying out a variety of automation scenarios. In the future (near, I hope) we can also add multi-agent systems to this.
For business decision makers, one key factor is how quickly and securely it is possible achieve operational benefits (and ROI). Whether you’re in finance, manufacturing, retail, or beyond, AI agents offer a new way to tap into supercharging business processes. Think scaling processes that traditionally depend on human intervention, to agent-driven that improve productivity, reduce manual errors, and freeing tedious work (and precious) time from humans. When I talk with customers about Microsoft 365 Copilot, it already helps many to complete more tasks faster than before. For many of them, that means less long days turning evenings – or that pile of to do tasks stays in control.
If you’re curious, I recommend checking the quickstart, spinning up a basic agent, and giving Azure AI Service UI and especially the Code Interpreter with Assistants API a try. From data crunching to helping your sales or support teams, you can see how fast you can build an pro-code agent core capable of meaningful work.
Read more from these sources:
Yes, I used again the Azure OpenAI Service reasoning model o1 to help me out with this. I provided the model a long prompt, that included my goal, insights, information of what I did and what I wanted to express in the post. Along with the background information from Microsoft Learn and articles. After that I used some prompts to refine the result and added example use cases. Finally I coped the text to the blog and went through this – applying changes, deleting parts and adding new insights, and of course pictures. This speeded up the actual blog writing process quite a lot, but it still took a few of hours in total.
Perhaps for one blog post I will create a Teams meeting, that I record and transcribe, when I testing out new feature. Using that could provide quite an unique base for the post draft, that I generate with the help of o1. That would not be so structured as writing my selected insights, but would definitely be a different way. Will it be faster? That I can find out by testing it out.
