Y Combinator and self-improving companies

There is a useful AI lesson in what Y Combinator is telling startups right now.

YC argues that startups should build companies in a way that allows them to improve themselves with AI.

That sounds like Silicon Valley. And of course, startups without grown hierarchies, legacy systems, and governance are not the benchmark for every larger company. Still, the idea is interesting because it asks a very concrete question: what does a company need to prepare so that AI can do more than make individual tasks faster?

The first step is not the agent. The first step is an AI-readable company memory.

Meetings, emails, decisions, documents, customer feedback, tickets, code, policies. As long as this knowledge sits only in people's heads, inboxes, chats, and slide decks, AI can do very little with it.

The point is that knowledge has to be searchable, citable, versioned, and broken into useful units. Even the knowledge of a startup does not fit into a context window. It needs structure before it becomes useful.

Only then does the second part become interesting: self-improvement cycles.

Customer service makes this tangible. A customer writes for the third time about the same problem. An agent answers the request. At the same time, an analytical process in the background detects a pattern: some of these tickets are not genuine one-off cases, but arise because an instruction is unclear, a form is labelled ambiguously, the agent gives wrong or incomplete guidance, or a process step regularly gets stuck.

Then AI does not stop at the answer. Similar tickets are grouped, the likely cause is investigated, and a change is proposed: a better help text, a clearer form, a process adjustment, or a small piece of software. A quality check, by an agent or a human, reviews the proposal. After that, the improvement flows back into the system.

In service, AI then does not only make responses faster. It helps reduce the reasons why recurring requests happen in the first place.

Y Combinator is speaking primarily to startups. There, it is much easier to build a company around data, context, and agents. For established companies, the thought is still relevant, not as a template, but as a strategic test.

Which parts of the organization are already documented, structured, and verifiable enough for AI to do more than make individual tasks faster?

Source: Y Combinator: How to Build a Self-Improving Company with AI https://youtu.be/X_JsIHUfUjc