Getty Images

Microsoft Visual Studio, Azure updates target AI developers

Microsoft updates this week aim to address the struggles of early enterprise generative AI adopters that have stalled efforts to put pilot projects into real-world production.

A broad package of updates to Microsoft products for AI developers this week looks to assist enterprises stuck in an experimental phase with generative AI.

After 18 months of intense hype from major cloud providers, generative AI specialists and startups such as OpenAI and Anthropic as well as virtually every enterprise IT vendor, there are still only lukewarm signs of real-world enterprise use of the technology among mainstream companies. In fact, one recent research firm survey suggests the so-called "trough of disillusionment" might be already arriving.

More than half -- 52% -- of AI projects so far have failed to make it into production, according to Gartner's 2023 "AI in the Enterprise" survey of 703 respondents in corporate leadership roles in the U.S., U.K. and Germany.

"There's still a high degree of either failure or finding that it's not as game changing as organizations think, so they don't proceed with it," said Jason Wong, an analyst at Gartner. "It's either finding out that they don't really have the quality of the data that's needed in order to make it a full production product or they tried it out and said, 'Is going to be worth the cost?'"

As in the initial wave of promotion for AIOps tools based on other forms of AI and machine learning, many enterprises have discovered that the effectiveness of generative AI tools depends heavily on the quality of IT automation practices and data management before those tools are applied, Wong said.

"People say, 'Well, our employees can't find anything. Our search is broken. … So we're going to turn it over to [Microsoft] Copilot and it'll just make sense of the mess,'" Wong said. "But you have to label your data, categorize your data, classify your data -- and if you had done that in the first place, you would have better search."

Successful generative AI examples.
Early examples of successful generative AI applications tend to be relatively simple projects that take an interative approach while the technology is still rapidly changing.

Microsoft boosts AI tie-ins and guidance for devs

Microsoft introduced Fabric, a fresh take on a SaaS data management platform, to address data integration and analytics, at its Ignite conference in November. Copilot chatbot within Fabric, meant to help users build data pipelines more easily, became generally available during Microsoft Build on Tuesday.

A preview feature called AI skills in Fabric is meant to help "analysts, creators, developers and even those with minimal technical expertise … to build intuitive AI experiences with data to unlock insights," according to a Microsoft press release.

Updates to AI developer tools followed suit in adding ease of use features and guardrails for software engineers creating AI apps. For example, version 17.10 of Microsoft's commercial enterprise IDE, Visual Studio, officially builds in a generally available, production-supported integration with the GitHub Copilot code assistant.

That integration has been in preview for multiple releases of Visual Studio, and Microsoft's free, lighter-weight Visual Studio Code editor has long supported GitHub Copilot. But as of Preview 3 of Visual Studio, "GitHub Copilot and Copilot Chat are available to install as a single extension that combines both Copilot and Copilot Chat into one package … built in and recommended by default in all workloads," according to Microsoft documentation.

Microsoft has also integrated GitHub Copilot with its existing Visual Studio code-completion tool, IntelliSense. While some overlaps between the two are initially confusing, one IT pro said that combination could be especially compelling for enterprises.

"IntelliSense is able to find libraries that are related but they are now using Copilot to add the variables into the library to complete the call," said Nick Cassidy, lead innovation engineer at Blue Shield of California, who emphasized that his opinions do not reflect those of his employer. "If I was writing new code, I would probably be using Copilot without IntelliSense. But [for] making changes to existing code, I could see this being useful."

Integrations of newer tech into familiar tools could help AI developers more quickly understand how to use it, said Larry Carvalho, an independent analyst at RobustCloud, a cloud advisory firm.

"Developers will pick the IDE they are most familiar with, and Microsoft tools are popular with a large segment of developers," Carvalho said.

Context awareness in Visual Studio's GitHub Copilot integration "will keep Microsoft ahead of the competition in assisting developers," he said.

Similarly, a bundle of updates to Azure AI Studio unveiled this week could help one Microsoft customer overcome hurdles to broader production use of GenAI. These updates include updated reference architectures, landing zone accelerators and service guides for the Azure OpenAI service; a preview of AI models as a service; and new monitoring support for LLM app performance.

"There are still big challenges to adopting AI in many enterprises, from hiring AI skills, defining an architecture on top of technologies that evolve or are replaced every quarter, to finding the hardware resources and partners to support critical service levels," said Nuno Guedes, cloud compute lead at Millennium BCP, Portugal's largest privately owned bank, headquartered in Lisbon. "This year's announcements show a comprehensive effort to lower these entry barriers."

Microsoft 365 users will also get a look at preview versions of third-party extensions for tools such as Atlassian's Jira in 365's Copilot Studio and Teams Toolkit for Visual Studio this week. This feature, Wong said, is notable for non-technical users and citizen developers that want to design custom chatbots.

"Based on our conversations with clients, there's really strong interest and follow-up for Copilot Studio among organizations that have started implementing Microsoft 365 Copilots. But often the question is, 'How do we get more role-based data from other systems, other data sources, applications?' Some clients have dabbled with that, but the full plugin approach is still a work in progress."

IT pros sort through a glut of tools, ROI confusion

In addition to enterprise hesitancy with generative AI due to data quality issues and concerns about security, data privacy and governance, at this point, there are simply too many generative AI developer and infrastructure automation tools to choose from, Cassidy said.

There are still big challenges to adopting AI in many enterprises, from hiring AI skills, defining an architecture on top of technologies that evolve or are replaced every quarter, to finding the hardware resources and partners to support critical service levels.
Nuno GuedesCloud compute lead, Millenium BCP

"What I'm dealing with right now -- and I think a lot of the industry may be starting to realize this -- is how do I even compare Watsonx to GitHub Copilot now that there's actual competition in this market?" he said. "Some of these announcements are so confusing that it doesn't even give me information enough [to] pay attention to what's coming out. It's just background noise."

Gartner clients piloting GenAI products find in-depth proof of concept comparisons between products confusing as well, especially in evaluating their return on investment, Wong said.

"We do see companies looking at this and saying, 'We need ROI here. We just can't be paying for all this based on faith of return,'" he said.

AI developer code assistants likely have caught on earliest for enterprises in part because it's relatively easy for some companies to evaluate whether they boost developer team productivity, Wong said. They can compare a team that uses them to a team that doesn't.

For Cassidy, however, this is not necessarily straightforward.

"There are so many variants and factors that can come into the performance of developer that what I'm finding is that it may be easier just to ask about the developer experience – 'Are you happier using a code assistant?'" he said. "But I can easily see junior developers lean on the code assistant a little too much and just assuming that it's right."

At Guedes' company, a few generative AI projects other than code assistants have made it into production despite these challenges so far, he said. He attributes this to an approach that focuses on relatively small, short-term iterations to address specific user stories with relatively well-defined ROI. So far, successful projects using that approach include improving internal information access across product and customer support teams and multimodal conversions between content formats, such as speech to text, sentiment analysis and quality assessment.

"Accepting that the current pace does not support long-term stable deployments, more than trying to figure out what will be the next best thing, we focus on being iterative and testing it out in the real world and making decisions based on that," he said.

Beth Pariseau, senior news writer for TechTarget Editorial, is an award-winning veteran of IT journalism covering DevOps. Have a tip? Email her or reach out @PariseauTT.

Dig Deeper on Software design and development

Cloud Computing
App Architecture