Microsoft is strengthening its Copilot generative AI software as the centerpiece of the workspace by adding key features to help users be more productive. The company’s next wave of Copilot updates includes “digital labor” tools to automate work and merge document intelligence, with the aim of helping enterprise users to work smarter, not harder. For example, a new Copilot Search feature neatly summarizes information by going through wads of unstructured information stored in an organization’s apps and systems. Copilot can provide context and present information after searching for data in common Microsoft productivity tools and third-party apps that include Google Drive, Slack, ServiceNow, and software from Atlassian. The new AI features, which include other summarization, desktop personalization, and administrative tools, are possible because of “advanced models, adaptive memory, and reasoning agents,” Microsoft said in a blog post. Other new Copilot features include “memory and personalization” capabilities — something Microsoft touted at its recent 50th anniversary event — which can personalize the desktop experience by retaining and understanding user preferences over time. “Memory and personalization are private just to you, and if you’re ever working with sensitive information, you can control what Copilot remembers,” Microsoft wrote in its blog. The “Copilot Notebooks” feature allows users to create a mini-AI model based on their notes and documents. Users can create a “notebook” and load it with relevant documents based on the task at hand. The notebook will then present contextual answers by scanning only those documents. The notebooks can be updated with new documents, which the AI model indexes into the corpus of data. Copilot Notebooks is similar to Google’s NotebookLM app, which is included in the competitor’s Workspace productivity suite. Google also announced the addition of new genAI features to Workspace at this month’s Cloud Next conference. The new capabilities will begin rolling out to Microsoft 365 Copilot customers in late May, the company said. Microsoft’s paid enterprise genAI offering is available as an add-on to commercial M365 subscriptions for an additional $30 per user per month. In Microsoft’s new Agent Store, customers can find pre-built AI agents from Microsoft and partners including Jira, Monday.com, and Miro. Microsoft’s previously announced Researcher and Analyst reasoning agents are rolling out to customers in the new Agent Store starting today via the company’s Frontier program. Companies can also offer their own custom agents in the store. Microsoft also announced the Copilot Control System for system administrators to manage the creation and deployment of AI agents within enterprise systems — including the ability to enable, disable, or block agents for specific users or groups. Microsoft is at the very beginning of the journey toward agentic AI, said J.P. Gownder, vice president and principal analyst at Forrester Research. “Microsoft Copilot Agents are good for tapping into specific data sources — like a SharePoint site — to ensure that you are querying the right data source,” Gownder said. Forrester views some Copilot features that handle simple tasks and tap into specific sources as being more “agent-ish” as opposed to true agentic AI. “Agent-ish solutions can solve point problems, but they aren’t the ‘digital coworkers’ that vendors are talking about, at least not yet,” Gownder said. As Copilot and other AI systems further integrate with enterprise tools, system administrators will need to put in guardrails so unauthorized users aren’t able to access sensitive information from AI features, said Jack Gold, principal analyst at J. Gold Associates. “Do you let Jack, who is a low-level employee, ask anything he wants? Knowing roles and security and data access clearance of various users… is going to be very important for general-purpose AI systems,” Gold said. Users will also need to use common sense, as AI tends to wander if not given specific context or criteria in queries. Depending on the role and data complexity, the user training process “will probably be a trial-and-error issue,” Gold said. “The biggest challenge will be: can you trust the results that the agents just provided you?” Gold added.