Every decade or so, a technology shift rewrites the economics of building a company. The public cloud was the last one. It meant startups no longer needed server racks, office space to house them, or the capital to buy hardware upfront. Everything got cheaper, and a wave of new ventures followed.
Amanda Silver, corporate vice president at Microsoft’s CoreAI division and a 24-year veteran of developer tools, believes agentic AI is the next shift of that magnitude. In a recent TechCrunch interview, she laid out a specific and testable claim: AI agents will reduce the operational costs of running a startup so significantly that we’ll see more ventures launching, and higher-valuation companies with fewer people at the helm.
It’s a bold comparison. And for UK businesses watching from across the Atlantic, it raises a pointed question: are you positioned to benefit, or will this be another American advantage?
The cloud comparison holds up — to a point
Silver’s analogy between cloud computing and agentic AI is more than rhetorical. Both shifts attack the same problem: the fixed costs of standing up a new business.
Strategic Reality: Cloud eliminated capital expenditure on hardware. Agentic AI targets operational expenditure on people and processes. Together, they compress the cost structure of a new venture from both directions.
The cloud removed infrastructure costs. AI agents target something different: the human labour involved in routine business operations. Silver specifically mentioned support functions, legal investigations, and software maintenance as areas where agents already reduce costs.
| Cost category | Cloud era impact | Agentic AI impact |
|---|---|---|
| Infrastructure | Eliminated hardware capex | Minimal additional change |
| Customer support | Enabled ticket systems | Agents handle routine cases |
| Legal and compliance | Cloud-based document management | Agents conduct initial investigations |
| Software maintenance | Continuous deployment | Agents update dependencies, resolve incidents |
| Staffing requirements | Reduced ops headcount | Enables higher valuation per employee |
The numbers Silver cited are specific: a 70-80% reduction in time for codebase maintenance tasks like updating library dependencies. That’s not a vague promise about productivity. It’s a measurable outcome from deployed multistep agents.
What agents actually do today
The most useful part of Silver’s interview was her willingness to get concrete. Two examples stood out.
First, codebase maintenance. Keeping dependencies current — updating to the latest version of .NET or a Java SDK across an entire codebase — is tedious work that developers typically postpone. Multistep agents can now reason over a complete codebase and bring it up to date. Silver estimated a 70-80% time reduction, and stressed that this genuinely requires agentic behaviour: the agent needs to understand the codebase structure, identify dependencies, test changes, and handle cascading effects.
Implementation Note: The 70-80% reduction Silver cited isn’t about simple find-and-replace. It requires agents that can reason about code structure, run tests, and handle dependency chains. This is multistep orchestration, not a single API call.
Second, live-site operations. Microsoft built an agentic system (Silver said “genetic system,” likely a transcription error) that diagnoses and often fully resolves live-site incidents without waking up the on-call engineer. Anyone who has been jolted awake at 3am to diagnose a minor production issue will appreciate the value here. The system doesn’t just alert — it investigates, diagnoses, and in many cases fixes the problem.
Both examples share a pattern worth noting: they’re boring, repetitive tasks that humans dislike doing. That matters. The most successful agent deployments aren’t replacing creative work. They’re handling the maintenance burden that slows teams down.
The real bottleneck isn’t technology
Here’s where Silver’s interview got genuinely interesting. Asked why agentic deployments haven’t happened as fast as expected, she didn’t blame model capabilities, hallucination rates, or safety concerns. She pointed to something more fundamental: organisations don’t know what they want agents to do.
Critical Context: The biggest stumbling block for enterprise AI agents isn’t technical uncertainty — it’s purpose. Organisations that can’t clearly articulate what success looks like for an agent won’t get value from deploying one.
“There’s a culture change that has to happen in how people build these systems,” Silver said. “What is the business use case that they are trying to solve for? What are they trying to achieve?”
This is a striking admission from someone at Microsoft. The technology works. The platform is ready. The barrier is on the customer side: they lack the clarity to define what an agent should accomplish, what data it needs to reason over, and what success looks like.
For UK organisations, this reframes the AI adoption challenge entirely. The question isn’t “should we deploy agents?” — it’s “do we have the organisational clarity to give an agent a well-defined job?”
SME Advantage: Smaller organisations may actually have an edge here. A 50-person company can identify its most painful manual processes faster than a 5,000-person enterprise navigating committee approvals and legacy systems.
The human-in-the-loop isn’t going away
Silver’s most nuanced point concerned trust and automation boundaries. She rejected the binary framing of “fully automated” versus “human controlled” and instead described a spectrum.
Her package return example was well chosen. Today’s process is roughly 90% automated, 10% human intervention — someone physically inspects the package to assess damage. Computer vision models are now good enough to handle most of those inspections. But borderline cases still get escalated. As Silver put it: “how often do you need to call in the manager?”
This is a useful mental model for any organisation designing agent workflows:
- Fully automatable: routine decisions with clear criteria and low stakes
- Mostly automatable: occasional edge cases escalated to humans
- Human-required: contractual obligations, production code deployment, safety-critical decisions
Strategic Insight: The question isn’t whether to have human oversight — it’s where to place the boundary. Silver’s framework suggests starting by automating decisions where computer vision or language models already match human accuracy, then gradually moving the boundary as capabilities improve.
Silver explicitly named two categories that will “always need some kind of human oversight”: incurring contractual legal obligations and deploying code into production systems that could affect reliability. These are useful guardrails for any organisation building automation policies.
What UK businesses should take from this
Silver’s perspective carries weight because she sits at the intersection of platform provider and enterprise customer. She sees what companies are actually building, and more importantly, where they get stuck.
Three practical steps emerge from her analysis:
Start with your most loathed tasks. Both of Silver’s concrete examples — dependency updates and live-site incident response — were tasks that engineers actively dislike. Agent adoption is fastest where the human motivation to hand off work is highest.
Define success before choosing tools. The biggest deployment failures Silver described came from organisations that jumped to building agents without first answering basic questions: what is this agent for? What data does it need? How will we know it’s working?
Take Action: Before evaluating any AI agent platform, document three specific workflows where you can define clear inputs, expected outputs, and measurable success criteria. If you can’t articulate these, you’re not ready for agent deployment.
Design your escalation model. The package return example isn’t just an anecdote — it’s a design pattern. Every agent workflow needs a clear answer to “when does this call in the manager?” Build the escalation path before you automate the happy path.
The challenges nobody is talking about
Beyond Silver’s analysis, four issues deserve attention from UK businesses evaluating this shift.
Cost opacity. Cloud computing was supposed to reduce costs too, and for many organisations it did. But cloud bills have a habit of growing faster than expected. Agentic AI will follow the same pattern: initial savings on routine tasks, followed by creeping costs as usage scales and edge cases multiply.
Hidden Cost: The per-token economics of AI agents are still maturing. A dependency-updating agent that makes hundreds of API calls per repository could generate meaningful compute costs at scale. Factor this into ROI calculations before committing to enterprise-wide deployment.
Skills displacement. If agents handle live-site incidents, junior engineers lose one of their best learning opportunities. The on-call rotation is miserable, but it teaches systems thinking, debugging under pressure, and production awareness. Removing that training ground has second-order effects on talent development.
Vendor concentration. Silver’s examples centre on the Microsoft ecosystem — Azure, Foundry, GitHub Copilot. The startup cost savings she describes partly depend on staying within that ecosystem. For UK businesses concerned about vendor lock-in, this is worth weighing against the productivity gains.
Regulatory asymmetry. The UK’s AI regulatory approach differs from the EU and US. Agent deployments that cross jurisdictions — a UK startup using US-hosted agents to serve EU customers — face compliance complexity that Silver’s startup-economics argument doesn’t account for.
The bottom line
Silver’s core claim — that agentic AI will transform startup economics as profoundly as the public cloud — is plausible but conditional. The technology works. The cost savings for well-defined tasks are real. But the “well-defined” part is doing a lot of heavy lifting in that sentence.
The organisations that benefit most won’t be those with the best models or the biggest compute budgets. They’ll be the ones with the clearest understanding of which problems to solve, the discipline to design proper escalation paths, and the patience to expand automation boundaries gradually.
Strategic Reality: This isn’t a technology race. It’s an organisational clarity race. The winners will be companies that can answer “what should this agent do?” with a specific, measurable answer — not “improve productivity” or “reduce costs,” but “process dependency updates across our three main repositories and flag breaking changes for human review.”
For UK startups, the opportunity is real. A two-person team with well-deployed agents genuinely could match the operational capacity of a 10-person team from five years ago. But only if they do the hard work of defining exactly what those agents should be doing.
Source: “How AI changes the math for startups, according to a Microsoft VP” by Russell Brandom, TechCrunch, 11 February 2026. Based on an interview with Amanda Silver, Corporate Vice President at Microsoft’s CoreAI division. Read the original article
Analysis by Resultsense — making sense of AI in the UK.