How AI changes the math for startups, according to a Microsoft VP

For more than two decades, Amanda Silver has focused on empowering developers. Over the past several years, that mission has increasingly focused on artificial intelligence. After spending a significant period working on GitHub Copilot, Silver now serves as a corporate vice president within Microsoft’s CoreAI division. In this role, she concentrates on tools that enable enterprises to deploy applications and agent-based AI systems at scale.

Her work focuses on the Foundry platform in Microsoft Azure, a unified AI portal for enterprise customers. That vantage point gives her insight into how organisations are implementing AI agents in real-world settings—and where those deployments fall short.

In a recent conversation, Silver shared her perspective on the current state of enterprise AI agents and why she believes the moment represents the most significant opportunity for startups since the advent of public cloud computing.

A Shift Comparable to the Rise of the Public Cloud

Silver views today’s AI landscape as a transformative inflexion point for emerging companies. In her assessment, the impact could rival the shift from on-premise infrastructure to cloud services.

She explained that the move to the cloud dramatically reduced the barriers to launching a startup. Entrepreneurs no longer needed to invest in physical server space or significant hardware infrastructure. Operating costs dropped, and access to scalable computing became widely available.

Now, she argues, agentic AI is poised to reduce operational costs further. Many of the foundational tasks required to build and sustain a company — including customer support, certain legal reviews, and other administrative processes — can increasingly be handled by AI-driven agents. By accelerating and reducing the cost of these functions, AI may enable more startups to launch and scale with leaner teams.

Silver believes this dynamic could lead startups to achieve higher valuations while operating with smaller core teams. In her view, that represents a compelling evolution in how companies are built.

Practical Applications of Agentic Systems

In everyday development workflows, Silver notes that multistep AI agents are already being widely adopted. For example, developers frequently need to update dependencies within their codebases, such as older versions of the .NET runtime or the Java SDK. Traditionally, bringing these systems up to date can be time-consuming and error-prone.

Agentic systems can now analyse entire codebases, reason through dependencies, and modernise them, delivering substantial time savings—sometimes cutting effort by 70% to 80%. Achieving this level of automation requires deployed multistep agents capable of performing structured reasoning across complex environments.

Another area seeing meaningful progress is live-site operations. Maintaining websites and digital services often requires teams to remain on call around the clock in case incidents arise. Historically, engineers might be called in for relatively minor issues that require manual diagnosis.

Silver described how AI systems are increasingly capable of diagnosing and, in many cases, fully resolving operational incidents autonomously. By reducing the frequency of human intervention, these tools not only improve system uptime but also shorten the average time required to resolve disruptions.

Why Agent Adoption Has Been Slower Than Expected

Despite strong interest, the deployment of agentic systems across enterprises has not progressed as rapidly as some predicted just months ago. Silver attributes much of this slowdown to organisational challenges rather than technical limitations.

According to her, one of the most common obstacles is a lack of clarity around an agent’s intended purpose. Successful implementation requires clearly defining the business problem and establishing measurable success criteria. Organisations must also carefully determine what data is provided to the agent so it can reason effectively about a given task.

In her experience, these design considerations pose greater hurdles than the general apprehension about deploying AI agents. Once organisations see tangible returns on investment, scepticism tends to diminish.

Human Oversight Remains Central

From an external perspective, uncertainty about handing control to AI systems can appear to be a major barrier. Silver acknowledges the concern but believes hybrid approaches will remain the standard.

She pointed to product return workflows as an example. Traditionally, a largely automated system might still require human review for certain judgment calls, such as assessing the condition of returned merchandise. As computer vision models improve, more of these evaluations can be automated. However, edge cases may still require human escalation — similar to calling in a manager for complex decisions.

Silver emphasised that certain critical processes will continue to require human oversight. Decisions involving legal obligations or production code deployments that could affect system reliability are areas where humans will likely remain involved. The key question, she suggested, is how much of the surrounding workflow can be automated safely and effectively.

The Next Platform Shift

In Silver’s view, the current moment signals an early phase of a broader transformation. Just as the public cloud reshaped startup economics by reducing infrastructure costs, agentic AI has the potential to compress operational overhead further.

As enterprises refine their understanding of business use cases and develop clearer frameworks for defining success, adoption may accelerate. If that happens, startups could find themselves operating in an environment where AI agents handle much of the foundational work, allowing founders and small teams to focus on strategy, creativity, and growth.

For Silver, the opportunity lies not only in technological advancement but in rethinking how companies are structured and scaled in an AI-driven era.