How AI changes the math for startups, according to a Microsoft VP
Source: TechCrunch
Interview with Amanda Silver – Microsoft Core AI
For 24 years, Microsoft’s Amanda Silver has been working to help developers — and in the last few years, that’s meant building tools for AI. After a long stretch on GitHub Copilot, Silver is now a corporate vice‑president at Microsoft’s Core AI division, where she works on tools for deploying apps and agentic systems within enterprises. Her work is focused on the Foundry system (a unified AI portal for Azure) and gives her a close view of how companies are actually using these systems and where deployments fall short.
I spoke with Silver about the current capabilities of enterprise agents, and why she believes this is the biggest opportunity for startups since the public cloud.
This interview was edited for length and clarity.
Q: Your work focuses on Microsoft products for outside developers – often startups that aren’t otherwise focused on AI. How do you see AI impacting those companies?
A: I see this as a watershed moment for startups, as profound as the move to the public cloud. The cloud let startups avoid the cost and logistics of owning rack space and hardware, making everything cheaper. Agentic AI will similarly reduce the overall cost of software operations. Many of the jobs involved in standing up a new venture—support, legal investigations, etc.—can be done faster and cheaper with AI agents. This should lead to more ventures and higher‑valuation startups with fewer people at the helm. It’s an exciting world.
Q: What does that look like in practice?
A: We’re already seeing multi‑step agents used across many coding tasks.
Example: Maintaining a codebase often requires keeping dependencies up‑to‑date (e.g., .NET runtime, Java SDK). An agentic system can reason over the entire codebase and update it, cutting the required time by 70‑80 %. This requires a deployed, multi‑step agent.
Live‑site operations is another area. When a website or service fails, someone on call is woken up to respond. Historically this was a dreaded, frequent interruption. We’ve built a generative system that can diagnose—and often fully mitigate—issues automatically, so humans aren’t woken up in the middle of the night. This also dramatically reduces the average incident‑resolution time.
Q: Agentic deployments haven’t happened as quickly as expected six months ago. Why do you think that is?
A: The biggest stumbling block is often a lack of clarity about the agent’s purpose. Teams need a cultural shift to define a clear business use case and success criteria. They also must consider what data the agent needs to reason effectively. These issues are more limiting than the general uncertainty around deploying agents; once the ROI is visible, adoption accelerates.
Q: You mention “general uncertainty” as a blocker. Why do you see it as less of a problem in practice?
A: Many agentic systems will operate with a human‑in‑the‑loop model.
Example: Package returns used to be 90 % automated with 10 % human inspection to judge damage. Modern computer‑vision models are now accurate enough that most returns can be fully automated, with only borderline cases escalated to a human.
Some domains will always need human oversight—e.g., entering a contractual legal obligation or deploying production code that could affect system reliability. Even there, the question becomes how far we can push automation before human intervention is required.
Event Details
| Location | Date |
|---|---|
| Boston, MA | June 23, 2026 |
Russell Brandom has been covering the tech industry since 2012, with a focus on platform policy and emerging technologies.
Russell Brandom previously worked at The Verge and Rest of World, and has written for Wired, The Awl, and MIT’s Technology Review.
He can be reached at russell.brandom@techcrunch.com or on Signal at 412‑401‑5489.