If you’re wondering how AI agents work, Microsoft’s new Copilot AI agents provide real-time answers on how data is being analyzed and sourced to reach results.
The Researcher and Analyst agents, announced on Tuesday, take a deeper look at data sources such as email, chat or databases within an organization to produce research reports, analyze strategies, or convert raw information into meaningful data.
In the process, the agents give users a bird’s eye-view on each step of how they’re thinking and analyzing data to formulate answers. The agents are integrated with Microsoft 365 Copilot.
The agents combine Microsoft tools with OpenAI’s newer models, which don’t answer questions right away, but can reason better. The models think deeper by generating additional tokens or drawing more information from outside sources before coming up with an answer.
The Researcher agent takes OpenAI’s reasoning models, checks the efficacy of the model, pokes around by pulling data from sources via Microsoft orchestrators and then builds up the level of confidence in the retrieval and results phases, according to information provided by Microsoft.
A demonstration video provided by Microsoft shows the Copilot chatbot interface publishing its “chain of thought” — for example, the step-by-step process of searching enterprise and domain data, identifying product lines, opportunities and more — with the ultimate output being the final document.
The approach is a major benefit for Microsoft since most models operate as a black box, said Jack Gold, principal analyst at J. Gold Associates.
Accountability and the ability to see how models are getting their results are important to assure users that the technology is safe, effective and not hallucinating, Gold said.
“Much of AI today is a ‘black hole’ when it comes to being able to figure out how it got to its results — most cite references, but not the logic on how they got to the end result,” Gold said. “Any transparency you can offer is about making users feel more comfortable.”
The Copilot Researcher agent can take a deeper look at internal data to develop business strategies or identify unexplored market opportunities — typical tasks by researchers. It provides highly technical research and strategy work that you’d expect to pay a highly skilled consultant, researcher, or analyst, a Microsoft spokeswoman said.
“Its ability to combine a user’s work data and web data means its responses are current, but also contextually relevant to every user’s personal needs,” the spokeswoman said.
For example, within the Researcher agent, a user can query the chatbot on exploring new business opportunities. In the process of analyzing data, the agent shares how the model is approaching the query. It will ask clarifying questions, publish a plan to reach an answer, show the data sources it is drawing information from, and explain how the data is collated, categorized, and analyzed.
The Analyst agent takes raw data and generates insights — typically the job of a data scientist. The tool is designed for workers using data to derive insights and make decisions without knowledge of advanced data analysis like Python coding, the spokeswoman said.
For example, the Analyst agent can take a spreadsheet with charts of unstructured data and share insights. Similar to the Researcher agent, the Analyst agent takes in a question via the Copilot interface, creates a plan to analyze the data, and determines the Python tools to generate insights. The agent shares its step-by-step process of how it is responding to the query and even shares the Python code used to generate the answer.
Microsoft has had a number of documented “misses” related to problematic generative AI (genAI) tools, such as Windows Recall, a Copilot feature that uses snapshots to log the history of activity on a PC, Gold said.
Giving users a sense of security is beneficial to getting users to try CoPilot, Gold said. “Think of it as having the safest car on the road when you go to select a new car for your family,” he said.
