Virtual labs give our employees at Microsoft on-demand access to the environments they need for development, testing, troubleshooting, and customer support.
But at our scale—we have hundreds of thousands of labs and virtual machines internally here at Microsoft—determining the right amount of cloud resources for each workload becomes a significant challenge.
Managed Cloud Labs (formerly MyWorkspace) is our internal, Azure-based cloud-labs-on-demand platform that supports more than 20,000 labs and 150,000 VMs across our customer support, cloud solution architecture, engineering, testing, and release teams.
With Managed Cloud Labs, our employees can model all different kinds of lab environments. Some environments might support a simple customer issue, while others re-create a complex enterprise infrastructure scenario.
Each lab runs with a mix of workloads, performance needs, storage demands, and usage patterns. Its flexibility makes Managed Cloud Labs beneficial; it’s also what makes cost optimization difficult.
For example, our business leads configure CPU, memory, and storage quota limits for their teams within Managed Cloud Labs, typically based on the largest environment used. Yet after these quotas have been set, there wasn’t an easily scalable way to determine if a given lab was provisioned for more capacity than it actually needed, resulting in unnecessary excess cost for these teams.
As the annual cost of Azure resources on Managed Cloud Labs climbed to eight figures at Microsoft, finding a better solution became urgent. This was especially true for storage performance—one of the platform’s primary cost drivers. Our team in Microsoft Digital, the company’s IT organization, went to work on a new approach that connected real usage telemetry with the infrastructure decisions that drive cost.
The result is an AI optimization service that can process performance telemetry, user behavior, infrastructure configuration, and cost signals across Managed Cloud Labs. The service learns how resources are used and can right size them for labs, matching the needs of each individual workload without disrupting the teams who depend on it every day.
Using AI to match cloud resources to workload needs
Managed Cloud Labs was designed as a cost-efficient lab platform, enabling us to quickly provision and deprovision resources on demand. However, due to the popularity of the service, usage and costs continued to rise. The obvious solution was cutting resources or imposing blanket quota, but this would risk damaging the very user experience that made the service valuable.
“It’s an incredible amount of data for a human to analyze. Even using our reporting and querying database capabilities, it would be very difficult to manage manually.”
Nathan Prentice, principal product manager, Microsoft Digital
To assess whether a lab is over-provisioned, our team would need to inspect every virtual machine, CPU, disk, performance tier, usage pattern, and cost signal across an enormous, constantly changing environment. Our team uses existing tools like Azure Advisor to help analyze resource configurations and usage telemetry and provide recommendations for optimizing resources, but storage costs remained a concern.
CPUs can be turned off when not in use, but operating systems and data disks preserve the state of each lab and keep incurring costs. That made storage the most promising place to reduce costs without breaking the experience.
Our team built an AI workflow that analyzes 30 days of disk usage data—including read/write operations and usage trends—then recommends whether a disk should stay on premium storage or move to a lower-cost SKU. This helps us individually evaluate each lab, VM, and disk, then recommend or apply the right performance tier, rather than forcing everyone into the same blanket policy.
“It’s an incredible amount of data for a human to analyze,” says Nathan Prentice, a principal product manager in Microsoft Digital. “Even using our reporting and querying database capabilities, it would be very difficult to manage manually.”
Building intelligence into the platform
We focused on extending the cost optimization capabilities through an intelligent, data-driven framework that evaluates real-world infrastructure usage at scale. The team developed a system that identifies opportunities to optimize storage performance tiers based on actual workload requirements. A disk may be appropriately sized for capacity, but workload telemetry often reveals that it does not require the performance characteristics of a higher-cost storage SKU.
Building on Azure’s native optimization capabilities and recommendations from Azure Advisor, the team created an AI-powered workflow that analyzes billions of telemetry signals, including disk utilization trends, throughput, IOPS, workload behaviors, and performance patterns. AI models hosted on Azure AI Foundry assess historical and real-time usage data to generate high-confidence recommendations while continuously evaluating potential risks to application performance and user experience.
“We saw an opportunity to combine Microsoft AI capabilities with operational telemetry to change how cloud cost optimization is performed. By using AI models hosted on Azure AI Foundry and enriching them with real-world workload and performance data, we built an intelligent decision framework that identifies optimization opportunities across Azure resources at scale.”
Nirag Tibdewal, senior software engineer, Microsoft Digital
We designed the solution with an emphasis on reliability, governance, and operational safety. Recommendations pass through multiple layers of validation and policy-driven guardrails. Managed services perform safety checks and assess workload impact, making sure recommendations meet predefined confidence thresholds before any action is considered.
To further minimize risk, the team adopted a phased deployment strategy. The system initially operated in observation mode, generating recommendations without taking action. Recommendations are validated in pre-production environments before being rolled out incrementally across increasingly larger virtual machine populations. This approach allowed the team to validate outcomes and refine recommendation quality to build confidence in the system and provide a seamless experience for engineers.
“We saw an opportunity to combine Microsoft AI capabilities with operational telemetry to change how cloud cost optimization is performed,” says Nirag Tibdewal, a senior software engineer in Microsoft Digital. “By using AI models hosted on Azure AI Foundry and enriching them with real-world workload and performance data, we built an intelligent decision framework that identifies optimization opportunities across Azure resources at scale.”
Tangible impacts without disruption
The Microsoft Customer and Partner Solutions (MCAPS) team has been working with the Managed Cloud Labs AI solution since January. MCAPS support engineers work with Microsoft customers and partners who face issues with our products. They use Managed Cloud Labs to set up labs that replicate customer environments and can be used to investigate these issues and test fixes.
Sometimes these lab environments are used to test performance and to scale related issues, so the business leads configure resource quotas at high levels. But often the labs are built to test configuration issues where large storage capacity or high IOPs aren’t required, resulting in mismatches between Azure resources the teams are paying for and what they actually need.
“With this AI-driven solution, we’ve reduced Azure Storage costs by 40%. We’re targeting underutilized disks that can be resized to less costly SKUs while still providing the required performance, with no impact to users. Cost savings is a significant priority for our leadership, so our ability to save this amount of money without impacting users has been a big success.”
Kevin Williamson, principal technical advisor, MCAPS
Our solution lowers costs for MCAPS and frees up constrained Azure capacity for other Microsoft teams. The storage-cost reduction is already saving the team a significant chunk of money without having to change business rules or impact support engineers.
“With this AI-driven solution, we’ve reduced Azure Storage costs by 40%,” says Kevin Williamson, a principal technical advisor in MCAPS. “We’re targeting underutilized disks that can be resized to less costly SKUs while still providing the required performance, with no impact to users. Cost savings is a significant priority for our leadership, so our ability to save this amount of money without impacting users has been a big success.”
Optimizing at scale
We can now focus on optimization at platform scale. We’ve reduced premium disk deployment from 95% to approximately 16% without reported disruptions. And we’re hosting the same scale as before, but doing it at 20% to 30% lower cost.
Our engineers are unaffected by lab performance, and they haven’t changed the way they work.
“We advanced infrastructure efficiency by building a time-series–driven pipeline to better understand real disk usage patterns across workloads. This supports AI-powered SKU recommendations, helping make Azure-based lab environments more data-driven and cost-efficient while intelligently optimized at scale.”
James Ferris, software engineer, Microsoft Digital
Key takeaways
Here are some tips to help your organization improve your cloud efficiency:
- Replace assumptions with telemetry before right-sizing cloud resources. Identify where telemetry exists today and use it to challenge resource assumptions.
- Use AI to scale what humans can’t. With tens of thousands of labs and VMs, AI analyzes usage patterns across disks, CPUs, and workloads to deliver recommendations no manual process could match.
- Target storage for the biggest savings. By focusing on disk performance tiers—rather than size—teams reduced one of the largest cost drivers without impacting lab functionality.
- Treat cost optimization as a continuous capability, not a one-time cleanup. By embedding optimization directly into the platform, cost management becomes continuous instead of reactive.
- Use guardrails to protect the employee experience. Pair AI-driven recommendations with validation steps and human-designed safeguards so optimization improves cost efficiency without disrupting the teams who rely on the platform.
