Microsoft’s new Copilot AI agents act like virtual employees to automate tasks

Microsoft will soon allow companies and developers to build AI-powered co-pilots that work like virtual employees and can perform tasks automatically. For example, instead of Copilot sitting idle and waiting for requests, it can monitor email inboxes and automate a number of tasks or data entry that employees would normally have to do manually.

It’s a big shift in Copilot’s behavior in what the industry commonly calls AI agents, or the ability of chatbots to intelligently perform complex tasks autonomously.

“We quickly realized that limiting Copilot to pure conversation severely limits what Copilot can do today,” explains Charles Lamanna, Corporate Vice President for Business Apps and Platforms at Microsoft, in an interview with The edge“Instead of having a copilot who waits for someone to chat with them, you could make your copilot more proactive and have them work on automated tasks in the background.”

Microsoft is rolling out this new feature to a very small group of early access testers today, ahead of a public preview in Copilot Studio later this year. Companies can create a copilot agent that can handle IT help desk service tasks, employee onboarding, and more. “Copilots are evolving from copilots who work with you to copilots who work for you,” Microsoft says in a blog post.

These Copilot agents are triggered by specific events and work with the company’s own data. Here’s how Microsoft describes a potential employee onboarding copilot:

Imagine you are a new employee. A proactive copilot greets you, discusses personnel information and answers your questions, introduces you to your buddy, gives you training and deadlines, helps you with the forms, and prepares for your first week of meetings. Now HR and employees can work on their regular tasks without the hassle of administration.

This type of automation will, of course, lead to questions about job losses and fears about where AI will go next. Lamanna argues that copilot agents can handle the repetitive and mundane tasks of jobs, such as: Other tasks, such as data entry, can be eliminated rather than replacing jobs entirely.

“What makes a job, what makes a role? It’s a bunch of different tasks and generally a very large number of very different and heterogeneous tasks. If someone was doing one thing over and over again, it would probably have been automated by current technology already,” says Lamanna. “We believe that with Copilot and Copilot Studio, some tasks will be fully automated… but the good news is that most of the things that will be automated are things that nobody really wants to do.”

Microsoft’s argument that it just wants to reduce the boring parts of your job sounds idealistic at the moment, but given the ongoing battle for AI dominance between tech companies, it seems like we’re increasingly on the cusp of more than just basic automation. Lamanna believes that human judgment and collaboration are still important parts of getting work done, and that not everything will be suitable for automation.

There are also still a lot of problems with generative AI right now, especially around hallucinations where it just makes things up on its own accord. Microsoft says it has built a number of controls into Copilot Studio for this AI agent push, so that Copilot doesn’t just go rogue and automate tasks at will. This is a big problem that we’ve already seen with Meta’s own AI advertising tools failing and wasting money.

You can build Microsoft’s Copilot agents with the ability to flag specific scenarios for human review, useful for more complex queries and data. This all means that Copilot should work within the boundaries of what has been defined and what instructions and actions are associated with these automated tasks.

Microsoft is also making it easier for companies to integrate their own data into their custom Copilot, with data connections to public sites, SharePoint, OneDrive and more. This is part of a broader effort within Microsoft to make Copilot more than just a chatbot that generates things.

“Copilot in 2023 – and Microsoft – was very focused on mining your data, aggregating your content and generating new content. We believe Copilot will have a strong focus on customization in 2024,” says Lamanna. New Copilot extensions will enable some of this customization, allowing developers to create connectors that extend Copilot to cross-industry systems.

Microsoft also wants Copilot to work more with groups of people, rather than the one-on-one experiences that existed last year. A new Team Copilot feature allows the assistant to manage meeting agendas and notes, moderate longer team chats, or help assign tasks and track deadlines in Microsoft Planner. Microsoft plans to preview Team Copilot later this year.

At Google I/O last week, the search giant also showed off some early concepts for its own AI agents that automate tasks for you, and demonstrated how Gmail users could use an AI agent to automatically fill out a return form for some shoes Someone collect them.

The big question remains how all these AI agents will work in reality. We constantly see AI fail at simple text prompts, give incorrect answers to questions, or add extra fingers to images. So do companies and consumers really trust it enough to automate tasks in the background? I guess we’re about to find out.