Listen to this article:
For a long time, we've been used to interacting with AI in the form of dialogue: we ask a question, we get an answer. However, the era of mere chatbots is just ending. At the Build 2026 conference, Microsoft presented a strategy that puts autonomous agents front and center – systems capable not only of talking, but also of acting, delegating tasks, and collaborating with each other.
Microsoft IQ: The Brain That Gives Agents Context
The biggest challenge for enterprise AI is not a lack of powerful models, but a lack of relevant information. An agent without context is like a very smart student without textbooks. To solve this problem, Microsoft introduced the concept of Microsoft IQ. It's not a single product, but a set of "headless" (interface-free) layers that provide agents with access to various types of data.
This new architecture consists of four pillars:
- Foundry IQ: Used for working with unstructured knowledge (documents, notes).
- Fabric IQ: Connects agents with business data in the Microsoft Fabric and Power BI environment. An agent can thus analyze sales without the need for manual report exports.
- Work IQ: The "agent-facing" side of the Microsoft 365 ecosystem. It enables agents to interact directly with Outlook, Teams, Word, or SharePoint.
- Web IQ: A new layer for web search, optimized specifically for agents, not humans. It's extremely fast and can automatically browse web pages and videos.
For developers, this architecture is crucial because, thanks to support for the MCP (Model Context Protocol) protocol, they can easily connect these IQ services to their own systems.
Multi-Agent Systems: When AI Collaborates
One of the most significant contributions is the ability to create multi-agent systems through the Copilot Studio platform. Instead of one giant model trying to solve everything, Microsoft now enables task delegation among specialized agents.
In practice, it looks something like this: One agent pulls customer data from a CRM system, a second agent prepares a contract draft in Word based on that data, and a third agent then schedules a meeting in Outlook. This process happens with minimal human intervention. As Ray Smith, VP of AI Agents at Microsoft, notes, splitting tasks among multiple smaller agents significantly increases the reliability and maintainability of the entire system.
New MAI Models: Microsoft's Own Intelligence
Microsoft has also moved away from dependence on a single provider. While OpenAI previously dominated, it now offers a wide range of models. Alongside its partnership with Anthropic (including the latest Claude 4.8 on Azure), Microsoft is massively developing its own MAI model family.
These models, including the new MAI-Thinking-1, are designed for specific enterprise needs: high token efficiency, cost optimization, and easy custom fine-tuning on company data. Compared to general models like GPT-4o or Gemini 1.5 Pro, MAI models focus on being "lighter" and faster for specific automation tasks where infinite creative intuition isn't needed, but precise execution is.
Comparison of AI Model Approaches
| Feature | Microsoft (MAI + Partners) | OpenAI (GPT Series) | Google (Gemini) |
|---|---|---|---|
| Model Selection | High (OpenAI, Anthropic, MAI) | Low (OpenAI only) | Medium (within Google ecosystem) |
| Focus | Agent workflows and context | General-purpose intelligence | Multimodality and Google Workspace integration |
Practical Impact for Czech Companies and the EU
What does this mean for a Czech entrepreneur or IT manager? First and foremost, it should be emphasized that Microsoft 365 Copilot and its new agent capabilities are fully available on the Czech market with Czech language support. For Czech companies already using the Microsoft ecosystem, this means AI implementation won't require building new, isolated systems, but rather integration into existing processes.
From the perspective of EU regulation (AI Act), the key topic here is "governance." Microsoft places great emphasis on monitoring, auditability, and security of agents within Copilot Studio. This is critical for European companies – they need to know why an agent decided what it did and ensure that data doesn't leak outside the company's secure boundaries. New tools for evaluating agents (Agent Optimizer) enable companies to check whether AI behavior is deviating from set rules.
Pricing Policy
For entrepreneurs, it's important to track costs. Microsoft 365 Copilot for businesses typically runs around $30 per user per month (approximately 700–750 CZK). Implementation of advanced agent features within Copilot Studio may have a different pricing structure based on token consumption and task complexity, which is standard for enterprise solutions.
What are the main security risks when using autonomous agents?
The main risk is so-called "uncontrolled execution," where an agent might perform an action (e.g., delete a file or send an email) that wasn't intended. Microsoft addresses this through Work IQ and robust governance frameworks that enable human oversight and transparent tracking of every agent step.
Can these new systems be connected to Czech software (e.g., accounting systems)?
Yes, thanks to support for standards like MCP and the ability to create custom agents in Copilot Studio, Microsoft's agents can be connected via API to local or specialized software, provided those systems have a public or internal interface.
Is special technical education required to use agents?
For a regular user (end-user), interaction remains intuitive – it works on the principle of task assignment. However, for setting up and managing complex multi-agent systems within a company, businesses will need experts in AI orchestration and Microsoft 365 administrators.
