Azure AI Foundry adds context, including model routing, and tightens governance for developers working on AI agents within its broader Agent 365 control plane.
Microsoft expanded support for multi-agent orchestration and governance in its Azure AI Foundry as part of a broader effort to attract enterprises to its new Agent 365, a unified control plane for agent-based workloads.
These moves come during an awkward stage in the adoption of enterprise AI agents, a point where few pilots successfully transition into production. MIT researchers found earlier this year that 95% of generative AI pilots fail. Gartner estimates that 40% of agentic AI pilots will be canceled before they reach production, predicting a market correction for the AI agent tools that have flooded the market.
"There's been no wave of technology … adopted faster than AI, but on balance, the success rates of AI projects are not what we'd like them to be," said Judson Althoff, chief commercial officer at Microsoft, during a keynote presentation at the Microsoft Ignite conference this week.
According to Althoff, AI projects fail for four primary reasons: a lack of alignment between business and IT; data quality issues; inadequate security and governance; and an "overemphasis on experimentation, random acts of innovation, rather than AI being put to use at scale across real business scenarios that can make an impact," he said.
Judson Althoff, chief commercial officer at Microsoft, introduces a broad array of AI agent product updates during a keynote presentation at Ignite.
The more than 70 product updates announced by Microsoft this week during Ignite focus on tackling these issues and advancing enterprise AI agents toward maturity. In the software development sector, Azure AI Foundry introduced enhancements for managing complex workflows that involve humans, AI agents, data and tools. It provided more corporate data context for developers' AI agents, strengthened centralized governance and security, and offered more detailed AI agent observability.
Azure AI Foundry refines multi-agent support
Early enterprise adopters of AI agents in production, such as Nordstrom, have found they are most effective when multiple agents each handle specific, small parts of a hybrid workflow that also integrates traditional forms of deterministic IT automation.
Azure AI Foundry now supports similar patterns with a new Model Context Protocol (MCP) tools catalog in preview, a model router made generally available this week and expanded support for third-party models and agents, most notably Anthropic's Claude. New Anthropic support, part of a three-way partnership between Microsoft, Nvidia and Anthropic unveiled this week, makes Microsoft the only major hyperscaler to support both OpenAI and Anthropic models.
The model router automatically selects the right AI model among the more than 11,000 supported by Azure AI Foundry on behalf of developers, choosing based on the desired traits of accuracy, performance, cost or balance. It also enforces U.S. and EU Data boundaries by default.
"Think about how many models are out there," said Matthew Flug, an analyst at IDC. "There's no way a developer can keep up with all the models and what they work best for."
The model router could potentially help platform engineers serve developers the right model for the right use case as well, Flug said.
Foundry IQ grounds agents with data
It's also increasingly clear that AI agent workflows work best when guided by as much in-context business data as possible, a pattern that has proven effective for companies such as Nordstrom and Cathay Pacific.
A new Foundry IQ feature for Azure AI Foundry represents Microsoft's first support for large-scale context engineering in its developer tools, with retrieval augmented generation (RAG) built for AI agents, according to an Ignite keynote presentation by Asha Sharma, president of CoreAI product at Microsoft.
Foundry IQ was accompanied this week by the launch of Work IQ, a data intelligence layer for Microsoft 365 tools, and Fabric IQ for Power BI apps.
"Unlike traditional RAG, Foundry IQ doesn't just retrieve [data]," Sharma said. "It plans, it reasons and it iterates across Work IQ, Fabric IQ, blob storage, the web and more."
Foundry IQ fits into a broader trend, embraced by other major IT vendors, such as VMware and Cisco, to build an AI data fabric into application platforms. Data platform companies, such as Databricks, have joined this convergence from the other side by adding application platforms as well, according to IDC's Flug.
"Microsoft is a large company, not just a typical application platform, but those types of technologies are expanding their capabilities or further connecting those capabilities with data services," he said.
Michael Krieger, chief product officer at Anthropic (left), appears with Asha Sharma, president of CoreAI product at Microsoft, during an Ignite keynote presentation.
Agent 365's AI agent governance updates
More sophisticated agent workflows with access to data in context will help with quality and accuracy. But governance and observability are the biggest blockers to enterprise AI agents, said Michael Leone, an analyst at Omdia, a division of Informa TechTarget.
"The reality is, agentic AI adoption hinges entirely on a security-first approach," Leone said. "Without trust and governance, single-agent deployment, never mind enterprise scaling, simply won't happen, and adoption will fall by the wayside."
Security and governance are key focus areas for Microsoft's overarching Agent 365 control plane. It can establish a centralized trusted agent registry and enforce access controls for agents to data and other IT systems. Additionally, a new Foundry control plane in preview centralizes agent management and visibility for developers and platform administrators. Recent updates to other products aim to enhance security and centralized monitoring of AI agents, including the integration of Security Copilot into Microsoft 365 and the introduction of a new Agent ID feature for Microsoft Entra ID.
Agent ID includes discovery for AI agents, including shadow agents, and automatically links Agent 365-equipped agents built in Copilot Studio to Agent 365 MCP servers. It also handles lifecycle governance for agents in accordance with enterprise policies and includes automated controls to prevent orphaned or overprivileged agents.
"For the vast majority of the Global 2000, Entra is already the trusted security gatekeeper," Leone said. "With Microsoft requiring every agent to have a unique Entra Agent ID, Microsoft is simply extending the governance model IT admins have used for decades to this new class of 'workers.'"
Microsoft harnesses broad installed base
Extending AI agents management and governance into tools that are already widely adopted by enterprises was also a major theme during the Ignite keynotes. For example, Microsoft Copilot is already used by 90% of Fortune 500 firms, according to a presentation by Ryan Roslansky, CEO of LinkedIn.
"Copilot is built for work … and embedded in the office apps that millions of people use every day," Roslansky said. "It's your window into the world of agents."
This year's industry conferences have been packed with similar pitches from vendors and open source projects, all vying to be the platform of choice for controlling AI agents, regardless of their provenance. Other vendors with similar AI agent control plane offerings include ServiceNow, IBM and Broadcom, along with the other major cloud hyperscalers AWS and Google Cloud. Other vendors with pieces of the puzzle are expanding into adjacent markets to achieve a similar position, including Cisco, which previewed its own data fabric in September and this week acquired AI platform vendor NeuralFabric.
Microsoft is thinking and talking about that abstraction [layer] in ways that I haven't heard a ton of others talk about.
Devin DickersonAnalyst, Forrester Research
However, analysts said Microsoft Agent 365 represents a potentially broader layer of abstraction, including cloud, data center and edge workloads, consumer productivity apps and developer tools, than most other vendors can claim. Microsoft also has breadth in its installed base -- according to Gartner, Microsoft is second in IaaS market share, representing 23.9% of the overall market, and 2024 IaaS revenue for Microsoft totaled more than $41 billion. Ignite presenters said 80,000 developers use Azure AI Foundry, while its GitHub subsidiary hosts more than 180 million.
"Microsoft is thinking and talking about that abstraction [layer] in ways that I haven't heard a ton of others talk about," said Devin Dickerson, an analyst at Forrester Research, citing other updates this week such as Foundry Local for Android and Foundry Local SDK, which can run AI models directly on mobile devices. "They're really trying to create a control plane across the entire Microsoft estate."
New AI agents, same build vs. buy decisions
The expansion to edge also counters the main value proposition from other established competitors, such as IBM/Red Hat and Broadcom, based on the prediction that enterprise AI inference workloads are moving out of the cloud to on-premises environments, Dickerson said.
"Sometimes things like Azure don't necessarily need to be as robust for on-prem scenarios as [those other companies]," Dickerson said. "But if it's close enough, it does make people think, 'Okay, can this just be an extension of my existing investment?'"
However, most large enterprises don't tend to have just one major IT supplier, and interest in multi-cloud support for AI is also strong, as are longstanding fears about being locked into one vendor, said IDC's Flug.
"With any hyperscaler, they'll say, 'We support everything.' But once you actually want to leave, it's very hard to leave, and also to connect to outside services," Flug said. "I'm hearing a lot about the multi-cloud approach from a resiliency perspective -- we've had a bunch of big cloud outages in the past year."
There are also potential downsides to Microsoft's breadth of offerings -- for example, some companies might perceive a conflict of interest with a model router that automatically recommends models from a company with a stake in AI models, Flug said.
However, for many companies without the resources the largest enterprises have to manage AI infrastructure internally, the pressure to get AI agent projects to production and beat competitors might trump other concerns, Flug said.
"If I have a great idea and I want to get started right now, someone else is going to have the idea eventually, or they may already have the idea," he said. "So it's important to get going."
Beth Pariseau, a senior news writer for Informa TechTarget, is an award-winning veteran of IT journalism covering DevOps. Have a tip? Email her or reach out @PariseauTT.
