For 24 years, Microsoft’s Amanda Silver has worked to help developers. In recent years, that focus has shifted to building tools for AI. Following a long period working on GitHub Copilot, Silver now serves as a corporate vice president in Microsoft’s CoreAI division. There, she focuses on tools for deploying applications and agentic systems within enterprises. Her work centers on the Foundry system inside Azure, designed as a unified AI portal for businesses. This gives her a close view of how companies are actually using these systems and where deployments often fall short.
In a conversation, Silver discussed the current capabilities of enterprise agents and why she believes this represents the biggest opportunity for startups since the advent of the public cloud.
Silver’s work focuses on Microsoft products for outside developers, often startups not otherwise focused on AI. She sees AI impacting these companies as a watershed moment as profound as the move to the public cloud. The cloud had a huge impact because startups no longer needed physical space for server racks or large capital investments in hardware. Everything became cheaper. Now, agentic AI is set to further reduce the overall cost of software operations. Many jobs involved in launching a new venture, such as support or legal investigations, can be done faster and cheaper with AI agents. This will likely lead to more ventures launching, resulting in higher-valuation startups with fewer people at the helm.
In practice, this looks like multi-step agents becoming broadly used across various coding tasks. For example, developers must maintain a codebase by staying current with the latest versions of dependent libraries. Agentic systems can reason over an entire codebase and update it much more easily, potentially reducing the time required by 70 or 80 percent. This requires a deployed multi-step agent.
Live-site operations is another area. Maintaining a website or service often involves someone being on call to respond to incidents. While people still remain on call, agentic systems can now successfully diagnose and, in many cases, fully mitigate issues that arise. This means humans don’t have to be woken up in the middle of the night to groggily diagnose problems, and it dramatically reduces the average time to resolve an incident.
One puzzle of the current moment is that agentic deployments haven’t happened as fast as expected. Silver believes a primary reason is that the people building agents often don’t clearly know the agent’s purpose. A culture change is needed in how these systems are built. Teams must be clear-eyed about the business use case and the definition of success. They also need to consider what data they provide the agent so it can reason through the task. These are seen as bigger stumbling blocks than general uncertainty about deploying agents, as anyone who examines these systems can see the return on investment.
Regarding the general uncertainty, which can seem like a major blocker, Silver sees it as less of a problem in practice. First, it will be very common for agentic systems to have human-in-the-loop scenarios. Consider a package return process. Previously, a workflow might be 90% automated with 10% human intervention for judgment calls, like inspecting damage. Now, computer vision models are becoming so good that less human oversight is needed. There will still be borderline cases requiring escalation, but the need for human intervention is shrinking.
Some operations will always require human oversight, such as incurring a contractual legal obligation or deploying code into a production system that could affect reliability. Even then, the question remains how much of the surrounding process can be automated.