Most AI-in-the-enterprise advice still treats AI as a productivity tool: a copilot bolted onto whatever process already exists. YC Partner Diana Hu's recent Startup School talk, How to Build a Company with AI from the Ground Up, calls that framing the mistake.
AI is not the new Excel. It is the new operating system. For founders building from scratch right now, that distinction can lead to real advantage.
The talk is ten minutes. I would recommend watching it twice.
The core argument: companies have historically run on what Hu calls open-loop systems, where decisions get made, work gets done but not systematically measured, and outcomes are interpreted manually by humans pushing status updates uphill.
That model breaks the moment you have an intelligence layer capable of ingesting every artifact your company produces. The companies that win the next decade will likely be closed-loop by design: every meaningful action generates structured data, agents read that data continuously, and the system improves itself over time.
Three points that stuck with me
- Queryable beats hierarchical. Hu's argument that traditional middle management often exists to compress and route information for human bandwidth lands hard. When the routing layer is an agent needing full context, the organization has to be legible to AI, and parts of the org chart stop earning their keep as humans move to the edge of processes.
- Token spend is the new headcount. Lean teams that look underfunded on paper but run significant API bills are often optimizing the right input. CFOs, and a fair number of investors, will need to adjust quickly.
- The incumbent disadvantage is structural. Startups are already operating this way. Large incumbent SaaS companies trying to retrofit closed-loop models are dragging years of process debt behind them. That gives new entrants an edge to out-execute.
Why this matters for the companies we back
This connects directly to what we have been writing about with Vertical AI moats. The cornered data advantage is not just about customer data. It is also about your own operating data, captured and made queryable from day one.
Companies that build closed-loop from the start compound semantic depth and operational leverage in ways incumbents struggle to match. The ones that wait will be retrofitting in 2027 and wondering why a Series A competitor is shipping twice as fast on a tenth of the headcount.
Watch the video. Then audit your own ops against it.
How to Build a Company with AI from the Group Up - Diana Hu, Y Combinator