Home | Blog | AI agents – what are they and how will they change the way we work?

AI agents – what are they and how will they change the way we work?

It’s Monday morning, the coffee hasn’t kicked in yet, and you have a busy day ahead of you. Perhaps, you have stacks of returns or invoices to review, you need to relay information to field staff, or you need to help employees with IT support. Of course, a computer and professional software are useful, but you can get help several levels higher – using artificial intelligence. AI agents will take care of all that, and you can reach for another cup of coffee and focus on long-term strategy, for example.

The AI agent can take on a number of tasks with you or for you – from taking on the role of virtual project manager to performing more complex tasks, such as reconciling financial statements to close the books. Microsoft 365 Copilot already acts as a personal assistant to help with both day-to-day duties and more complex, creative projects. Using it to interact with a variety of agents opens up a world of new possibilities for companies, assisting employees, driving business forward and allowing them to achieve even more.

Agents can work non-stop, 24 hours a day, such as reviewing and approving consumer returns or reviewing shipping invoices, to help the company avoid costly supply chain errors. They can infer based on product descriptions to give field technicians step-by-step instructions, or use context and memory to open and close tickets for the IT helpdesk.

Think of agents as new applications in an AI-driven world. Quickly add new capabilities to address the biggest pain points of people at work and achieve real business results – Says Jared Spataro, CMO of AI at Work, Microsoft.

Why agents (and non-agents) at all?

Simply put, agents are software and agents are people. An AI agent harnesses the power of generative artificial intelligence because instead of merely accompanying you, it can work alongside you or on your behalf. Agents can do all sorts of things – from answering questions to complex, multi-step tasks. What makes them different from a personal assistant is that they can be customized to have specific experience and knowledge.

For example, you can create an agent that knows everything about your company’s product catalog, so it can write detailed answers to customer questions or automatically compile product descriptions for an upcoming presentation. Other agents can do even more – like one that helps you process sales orders, freeing you from that duty and allowing you to focus on building new customer relationships, for example.

Having agents perform these routine tasks helps increase productivity in many industries – from manufacturing and research to finance and retail, ultimately helping companies save time and money. You can use such ready-to-use AI agents in Microsoft 365 i Dynamics 365 or build your own custom agents tailored to your needs in Copilot Studio.

Imagine you’re a merchant with a large quarterly target. Copilot acts as your personal AI assistant, it prepares emails, summarizes missed meetings and helps you polish your presentations. Meanwhile, the agent, whose specialty is generating sales leads, works autonomously in the background, searching for new prospects you can contact this week. Microsoft Copilot accompanies you in your daily tasks, while an agent with specially selected skills helps you achieve your quarterly goals.

AI agents are not only a way to get more value out of work, but will be a new paradigm shift in how we work.

Agents are not a complete novelty. Microsoft and other software developers have used the name in the past to describe various types of software, but in recent years the term has tended to mean AI applications with specific capabilities. The software has been gaining more attention recently thanks to advances in large language models (LLMs), which have made it possible for everyone – not just developers – to easily communicate with AI. It is this agent-LLM duo that makes AI tools so useful.

People expect artificial intelligence to do things for them, not just generate language, says Ece Kamar, managing director at Microsoft AI Frontiers Lab. If you want to have a system that can actually solve real-world problems and help people, these systems need to have a good understanding of the world we live in, and when something happens, such a system needs to see that change and react accordingly.

AI agents (AI agents) are like layers over artificial intelligence models that observe and collect information, provide information to the model and collectively generate an action plan and communicate this to the user – and even act on the user’s behalf if he or she agrees. Both agents and models are therefore equally important pieces of the puzzle.

Agents will be able to gain more autonomy and become more useful through innovations in 3 areas: memory, permissions and tools. With memory, agents can maintain continuity, so that when you ask for something, it’s not like starting from scratch.

To be autonomous, you have to carry the context through a lot of actions, but the models are very disconnected and don’t have the continuity that we have, so each prompt hits a void and can pull out the wrong memory. It’s like watching a stop-motion animation with each frame isolated from the rest, and your mind sets it in motion. The clay model doesn’t move on its own.
– Sam Schillace, CTO at Microsoft

To build a memory infrastructure that would solve the above problem, Schillace and his team are working on the process of chunking and chaining. This means exactly what it sounds like – the team is experimenting with dividing interactions into chunks that can be stored and chained according to their meaning, so that they can be accessed faster. This works in an area reminiscent of human memory and associations – for example, grouping conversations about a specific project allows an agent to recall details when you ask it for status updates, and it won’t have to search the entire database.

The work with permissions and tools, in turn, is to ensure that agents have secure access to the information they need to perform tasks for you with your permission. An example of such information might be who your boss is, for example. They also need permissions to use applications on your behalf, such as Teams or PowerPoint.

How to use and build AI agents to work?

You can now create and publish agents in Microsoft 365 Copilot to help you with your daily tasks. It’s as easy as creating a worksheet or presentation and requires no coding skills. You also don’t have to be a developer to build agents in Copilot Studio. Anyone can connect them to relevant business data, such as emails, reports and CRM systems, so they can perform tasks and provide insights.

Soon you will also be able to hire new agents in Microsoft 365 to help you with typical workflows and tasks. The Interpreter (Translator) in Teams will provide real-time speech-to-speech translation in team meetings, and can even simulate your voice. Employee Self-Service Agent, on the other hand, will simplify HR and IT support tasks, such as helping an employee solve a laptop problem or checking whether employees have already used certain benefits. It can also connect to company systems for further personalization in Microsoft Copilot Studio.

Microsoft Dynamics 365 will also receive agents, as well as a number of typical business workflows from the sales, supply chain, finance and customer service areas. All SharePoint sites will also be equipped with agents. They will be tailored to the organization’s content, allowing employees to quickly tap into this knowledge base and find exactly what they need within seconds – whether it’s details of a project buried somewhere deep, or a summary of the latest product memo.

Developers have even more options. Using the new Azure AI Agent Service they will be able to choose from small or large language models for orchestration, and develop and scale agent-driven applications to simplify and automate complex workflows such as order processing and customer data synchronization. The service provides a software development kit (SDK) with agent development tools, allowing developers to integrate agent capabilities with Visual Studio Code and GitHub.

One type of model, OpenAI o1, can provide agents with even more advanced inference/understanding capabilities, allowing them to perform more complex tasks by breaking them down into steps – such as obtaining information that someone in the IT helpdesk needs to solve a problem, incorporating that person’s tried-and-true solutions, and developing a plan.

The power of AI agents can also be used on LinkedIn, whose first agent helps recruiters with hiring.

Risk assessment with autonomous AI agent operation

There is a bit of a security concern when it comes to autonomous AI agents, and Microsoft is focusing on making sure that these agents only access what the user wants. This is especially important from the perspective of responsible AI. The same principles of responsible AI that apply to other applications also apply to agent risk estimation and prevention.

An example is the new Copilot Control System, which helps IT departments manage Copilot and agents, offering access and oversight of data, security and management controls, and providing measurement reports and tools to track deployment and business value.

Many agents, such as those developed for Microsoft 365 and Dynamics 365 suites, contain human-in-the-loop approvals, in which people are required to take a final step, such as reviewing and sending an email that a Sales Order Agent has written. For agents developed in Copilot Studio, authors can review the records to see what action the agent took and why.

Looking back, but also forward

People have been fascinated for years by the idea of autonomous systems working side by side with us and assisting us. Thanks to large-scale language models we finally got the missing ingredient that allows these systems to solve real problems. Artificial intelligence developers can now implement many of these ideas, developed over decades of research. What’s next. Microsoft’s Ece Kamar, mentioned earlier, imagines a new ecosystem or marketplace of agents that could assist humans the way smartphone apps do today.

Agents already have the basic building blocks they need to perform tasks, such as observing “I see your meeting is dragging, I should reschedule my next meeting,” Kamar says. She adds that AI agents are not only a way to get more value out of people’s work, but also a coming paradigm shift in how work is done in general.

And this is just the beginning. Copilot for Microsoft 365 is set to evolve with new capabilities from under the banner of Copilot Actions, designed to handle routine tasks that can ease the burden on employees, such as creating summaries of missed emails, creating schedules or generating monthly reports. More such capabilities will come next year.

Copilot will support every employee to do their best work in less time and be able to focus on more meaningful tasks. Agents created in Copilot Studio will transform every business process, helping companies simplify operations, strengthen collaboration and drive innovation at scale.
– Jared Spataro, CMO at AI at Work, Microsoft

You can read more about autonomous AI agents in our articles on CenterXP:

Source: Microsoft Source, own elaboration

Author: Krzysztof Sulikowski