Microsoft Ignite AI's Governance Question
Generative AI has a security and compliance problem that few vendors are willing to address. Now Microsoft is giving enterprises the tools they need.
It's fair to say that Microsoft's big IT conference didn't satisfy everyone. The head of Microsoft AI took to social media following a backlash against the focus on Agentic AI. There certainly was plenty of AI for the sake of it at Microsoft Ignite, including a myriad of new office agents and a brand new Ask Copilot on the taskbar. Yet, for those attendees willing to invest in AI, there was a notable difference in tone and content compared to other big tech firms.
Google and OpenAI have spent plenty of time recently hyping the capabilities of the technology, or tweaking how chatbots communicate with consumers. Microsoft are one of the few AI vendors talking about the biggest issue facing AI in the enterprise. Namely, how to secure and monitor the technology. Along with all the hype about Agents in Windows and Office, there were also plenty of discussion about AI governance and agent control.
Agent 365
Take Agent 365, one of the biggest product reveals from the event. Agent 365 is a governance layer for AI agents, designed to allow IT Pros to see all the agents in the enterprise, as well as control which Microsoft resources each agent has access to. It's not just for Office 365 agents either. Competing AI vendors such as ServiceNow and Salesforce integrate with it too. Microsoft already have an agent builder for Copilot, now they're giving customers the ability to monitor those agents as they’re piloted across the enterprise.
As agentic AI proliferates within a business, it’s important to have a view of where the technology is used. You can’t secure an app if you don’t have visibility of it. An agent inventory allows users and departments to see which processes need human-in-the-loop steps embedded within them. IT teams can also use Agent 365 to restrict agents from accessing confidential data. This is more than just a security issue. It prevents an agent from being confused by irrelevant information to which it doesn’t need access. Controlling background and context has emerged as the best way of improving AI output given the current state of the technology.
Agent Workspace
The same principle extends to Windows. Microsoft announced a new agentic layer for Windows last week. This included an MCP server within the OS, as well as an on-device Agent Registry similar to the one the firm is introducing into Microsoft 365. Perhaps the most interesting development is the concept of Agent Workspaces. These workspaces segregate agent activity from human activity, allowing agents to run in the background within their own user session. In doing so, agents can execute their tasks locally without interrupting the workflow of the end user. It also enables agents to operate on device with their own permissions and app access, rather than inheriting the permissions of the user.
More importantly, this principle of separating agent activity from human activity improves security monitoring. Through Agent Workspaces, every agent action is tracked and logged as being carried out by an AI, ensuring that automated activity can be monitored in real-time. That is a critical step towards building confidence in agentic workflows, both in terms of security and accuracy. It gives IT teams the confidence that agents can’t go rogue. It gives business teams the confidence they can review and correct agent outputs before they’re used in production. Above all, it puts the user in charge of the agents they’re managing.
Management Toolkits
Microsoft’s big pitch is still that of an AI teammate. They’ve spent most of the year promoting frontier firms that are already experimenting with AI. However, for all the executive mandates and vendor thought leadership, widespread adoption of AI can only happen if technical teams feel in control of the technology. We’ve already seen the same trend play out with machine learning and cloud computing over recent decades. Eventually, business teams will probably adopt AI automation is part of the day-to-day operations. But first they need the tools necessary to develop it and manage it. Finally, we’re starting to see that management toolkit take shape. Hopefully, it’s the start of another AI trend.