Microsoft’s big focus at this year’s Build conference is generative AI. And to that end, the tech giant announced a series of updates to its platforms for building generative AI-powered apps and experiences: Azure AI Studio and Copilot Studio.
First, a quick refresher on Azure AI Studio and Copilot Studio. Azure AI Studio is a toolset within Microsoft’s Azure OpenAI Service that lets customers combine an AI model like OpenAI’s recently announced GPT-4o with their own data and build a chat assistant or another type of app that “reasons over” that data. Copilot Studio, meanwhile, provides tools to connect Copilot for Microsoft 365 — the AI-powered “copilot” in apps like Excel, Word and PowerPoint as well as Microsoft’s Edge browser and Windows — to third-party data.
Azure AI Studio, now generally available, will soon allow developers to build generative AI-powered apps using pay-as-you-go inference APIs — the APIs through which developers can access and fine-tune generative AI models hosted on Azure infrastructure. Microsoft calls this “model-as-a-service,” and it’s launching with models from Nixtla and Core42 to start, with models from additional providers including Cohere, Stability AI and AI21 Labs to come in the future.
Other new Azure AI Studio capabilities in preview let customers train and debug generative AI-powered apps by comparing different versions of them — and monitor apps in production for usage and quality. Users can visualize different trends and receive alerts based on custom-defined filters and settings.
Azure AI Studio also now integrates with Microsoft Purview (in preview), Microsoft’s service to prevent unauthorized access to data across apps and services, to discover potential “data risks” in AI apps, impose encryption on sensitive data and govern AI app usage. And Studio is shipping new tools to attempt to prevent “jailbreaks” of AI models — i.e. workarounds that disable a model’s safeguards — and detect hallucinations, or when a model invents facts from whole cloth.
On the Copilot Studio side, Microsoft is launching Copilot agents, which the company describes as AI bots that can “independently orchestrate tasks tailored to specific roles and functions.” Leveraging memory and knowledge of context, Copilot agents can navigate various types of business workflows, learning from user feedback and asking for help when they encounter situations they don’t know how to handle.
Here’s how Charles Lamanna, CVP of business applications and platforms at Microsoft, explains the concept in a press release: “Developers provide their copilot with a defined task, equip it with the necessary knowledge and actions and then Copilot Studio orchestrates dynamic workflows and acts behind the scenes to … integrate them to automate the task.”
Also new to Copilot Studio are extensions and connectors, both in preview for Copilot for Microsoft 365 and directly within Microsoft’s enterprise collaboration platform Teams. Extensions allow developers to customize AI-powered copilots with instructions, knowledge from databases and actions from plugins, for example to build copilots that handle tasks such as expense reporting and employee onboarding. Connectors, on the other hand, offer ways for developers to “ground” a copilot with organizational knowledge from a range of different sources.
“Extensions expand the actions Microsoft Copilot can take on the user’s behalf, customize grounding knowledge with relevant business data, and enable hand-off to other copilots,” Lamanna adds. “And Copilot connectors include … Power Platform connectors, Microsoft Graph connectors, and Power Query connectors — with Microsoft Fabric integrations coming soon. This makes it possible for copilots to use various data sources, including public websites, SharePoint, OneDrive, Dataverse tables, Microsoft Fabric OneLake and Microsoft Graph, as well as leading third-party apps.”