Video Title: How to Build a Self-Improving Company with AI

Video Author: YC Root Access

Editor: Peggy

Editor's Note: In this latest YC batch talk, YC General Partner Tom Blomfield discusses not "how to use AI to improve employee efficiency," but a more fundamental question: When AI is no longer just a Copilot, but can perceive, make decisions, call tools, receive feedback, and self-correct, how should the company itself be redesigned?

Tom's core judgment is that traditional companies still operate like "Roman legions": information flows upward through layers of hierarchy, and commands flow downward through management chains. But AI is breaking this organizational assumption. What truly matters is not letting engineers write 20% more code, but extracting the business knowledge scattered across emails, Slack, meetings, documents, and human brains, and turning it into organizational context that AI can read, call upon, and iterate on.

In his view, the future AI-native company will be composed of a series of recursive, self-improving AI loops: the system senses external changes from customer emails, support tickets, and product data, executes decisions through a rule layer, tool layer, and quality gates, and then automatically learns and corrects based on results. YC is already experimenting with similar mechanisms: the agent doesn't just answer questions; it monitors which queries fail, determines whether new tools, databases, or indexes are needed, and automatically submits code, reviews, merges, and deploys. That is, the company can continue to optimize itself while the founders are asleep.

This also means AI's impact on companies will not stop at the tool layer but will further change organizational structure. Tom proposes "burn tokens, not headcount" — the bottleneck for future startups may no longer be employee count, but token usage, business context quality, and the readability of organizational knowledge. The coordination functions of middle management will be largely replaced by AI, while ICs, direct responsible individuals, and human roles capable of handling high-risk judgments in the real world will become more important.

The most noteworthy point isn't that AI makes companies more efficient, but that it is changing the organizational form of the "company" itself. When software can be generated on the fly, processes can improve automatically, and experience can be continuously deposited into the company's brain, what founders truly need to build may no longer be a team with clear hierarchies, but an intelligent system capable of continuous learning and self-optimization.

Below is the original text:

Reimagining Operations: Companies Should No Longer Function Like Roman Legions

This part is based somewhat on a previous talk by Diana. The video from the weekend is already online and is excellent. Additionally, Jack Dorsey posted some tweets about two or three weeks ago that I found very interesting, so I "borrowed" quite a few of his ideas and incorporated them into this talk.

This talk will be quite conceptual and high-level, mainly discussing how we should rethink building companies.

The Roman legion was designed, in essence, to project power outward from the center of Rome across two continents, all the way to Hadrian's Wall near Scotland. It relied on a nested hierarchy, with a stable span of control at each level. Each level had a clear person in charge, responsible for passing commands down and sending information back up.

If you look at most companies today, you'll see they still operate like a Roman legion: people act as the channels for information to flow up and down. What struck me from Jack Dorsey's tweet thread is that we have always defaulted to the assumption that hierarchical organization is the best way to organize economic value units. But I believe AI is essentially breaking this assumption.

A year ago, if you asked people what AI was for, they would usually talk about "productivity": things like Copilot making engineers 20% more efficient, integrating Copilot into workflows, and helping teams deliver more software. But I think this is a problematic way of understanding it. It's like putting a more powerful engine into an old way of working. What's truly worth thinking about isn't how to add an AI tool to an old organization, but reimagining what the company itself is and how it should operate.

For example, from what Garry was just talking about, I truly believe he can now produce more code by himself than an entire engineering team. What I constantly think about is how to extract the domain knowledge within a company and define it as context, a skill set, or whatever you want to call it.

So-called domain knowledge, business knowledge, know-how – it was originally scattered in people's brains, Slack messages, emails, Notion documents. This information collectively defines how your company operates. Once you can make this knowledge clear and readable, you can move from a hierarchical organization to an intelligent organization driven by AI-native software.

Making the Company Better While You Sleep: How AI Loops Automatically Discover, Fix, and Deploy

AI is not something attached to the side of a company. It's not just a tool for engineers to be more efficient. I believe we can reimagine companies as a set of recursive, self-improving AI loops. This point is very important because once a company reaches this stage, it will continuously self-optimize even while you sleep.

For example.

Diana also mentioned this AI loop in her talk. It first has a "sensor layer." This term sounds fancy, but it can be quite simple: customer emails, support tickets, code changes, user cancellations, product telemetry data – these are all sensor data used to gather information from the outside world.

Then there's the strategy or decision layer, i.e., the rules: what the AI can do, what things must ask for human permission, what operations must be logged. Next is the tool layer, which is a bit like the skills and code Garry mentioned – essentially deterministic APIs, like querying a database or checking a calendar – a set of tools the AI can call upon.

Then comes the quality gate, like the deterministic checks, safety filters, and human review for high-risk items Eva mentioned. Finally, there's the learning mechanism: the system interacts with the real world, discovers where it isn't working, and feeds that feedback back to the start of the loop.

If each of these steps can run without human intervention, or with minimal human intervention, then the system will get better and better while you sleep.

I can give you some examples we are running now. Initially, we built an agent you could ask questions, and it had some deterministic tools to query our database. For instance, a simple question: When was the last time I had office hours with this company?

Later, it got smarter. For example, I'm doing office hours with a company, and they need to connect with someone in the petrochemical industry. This system could query the database in different ways, combine it with RAG methods, find five relevant founders, and recommend them to you.

But this was still just a sidekick, an assistant agent. It was still the way people used AI last year: AI makes me, as a group partner, more efficient, boosting my productivity by 20% or 30%.

My real "aha moment" came when we added a monitoring agent on top of this system. It looks at every query made by every YC employee, determines which queries succeeded and which failed. Then it asks: Why did it fail? What would make this query succeed? Do we need a new deterministic tool? Does a skills file need updating? Do we need a new database? New indexes?

These things now actually happen automatically at night. It writes code, submits a merge request to YC's codebase, has another agent review it, then merges and deploys it. So the next day, when a human asks the same question, the query succeeds.

For me, that was the critical moment. It wasn't just making a human 20% or 30% more valuable. The AI completed the loop itself, finding a way to self-improve.

I think if you can identify which parts of your company can operate like this and minimize the human role in execution and supervision, you can pour tokens into the problem, and the company itself will continuously get better.

There are many other examples. For instance, with product analytics data, you can have an agent analyze the data, find the friction points in the sales funnel. It can research best practices, set up an A/B test, run it for a week, pick the winning version, and deploy it.

This will happen over and over again. Your product will have a self-optimizing product loop.

Customer support is the same. Customer suggestions keep coming in. You can use an agent for triage. This agent acts a bit like your Chief Product Officer and Chief Technology Officer, making judgments: We don't want to do this suggestion, discard it; but this one aligns with our roadmap and can be completed tonight. So it writes the code, deploys it, and delivers it directly to the customer, all without human intervention.

So, if you can view every part of your company as a self-improving recursive AI loop, it becomes something entirely different from a "Roman legion" style hierarchical company.

Burn Tokens, Not Headcount: How AI-Native Companies Will Reshape Organizational Structure

So, what does it mean if you want to do this?

Point one: Burn tokens, not headcount. We are now seeing that many companies, by Demo Day, have about five times higher revenue per employee than 18 months ago. I think this trend will continue into Series A and B rounds. Soon, the real constraint won't be the number of employees, but token usage.

The crudest measurement now is everyone's token usage. Of course, this metric is stupid in extreme cases and easily gamified. But directionally, I think it's right. We are in a phase of exploring "what is possible," so everyone should be experimenting maximally to see what this crazy new intelligence can do.

Once you turn it into a leaderboard and tie promotions or firings to it, it will certainly be gamed and distorted. But directionally, figuring out who in the organization is leveraging tokens to the max and who isn't is indeed one way to determine which employees deserve your time.

I believe middle management is over. At least for these coordination problems, I don't think we need middle management anymore; AI should handle it.

For me, the future has two important roles. Jack Dorsey mentioned three, but I didn't like the third, so I took it out. I think the truly important roles are two: Everyone must become an IC, an individual contributor, a builder, an operator. And crucially, there needs to be a directly responsible individual. Any initiative needs a single named person responsible, not a committee, not a group.

I believe companies can be built entirely on ICs. Middle management is truly over. And building a self-improving company is this vision.

By the way, I think everyone is at the very forefront of this. I'm also really curious to know where you all are in your progress. It feels like everyone is still exploring the boundaries. I'm not sure if anyone has truly built a self-improving company across every function yet. Maybe I'm wrong, and you can prove me wrong.

What would I do first?

The first very important thing is to make the entire organization readable and understandable by AI. What does that mean? It means you have to record everything.

Simply put, all our partners' emails now – if you email a YC partner, that email goes into the YC database. Every Slack message, every DM, every office hours session – we've been recording them all for the past 3-4 months. Everything that happens, if it's recorded, it happened for the AI; if it's not recorded, it didn't happen for your intelligent system.

I was chatting with some founders here earlier, talking a lot of great content about their companies. Every time I had such a conversation, I thought I should really be recording this. Because someone needed an introduction I promised, and now I can't even remember who the introduction was for. I agreed to do it, then told them to email me later because I knew I would forget, given I had 20 more people to talk to.

So, this might require phones, recording devices, smart glasses, or microphones in every room. In short, everything needs to be recorded so that AI can read it.

Then, like Garry said, you need speaker diarization and summarization. You can't just stuff 100,000 hours of audio into a context window. You have to process it, aggregate, compress, distill it into important parts, and then give the AI some cues.

For example: Who here has read the YC user manual? I hope everyone in this room has opened it at least once. It's okay. Most of that manual was written 5-10 years ago; it's a bit outdated.

Last weekend, Harsh had this idea: Since we've accumulated about 2000 hours of office hours recordings over the past three months, why not regenerate the user manual?

So you give the system a set of instructions to first organize, compress, and synthesize the recordings, then categorize them by topic like fundraising, hiring, co-founder disputes, and then have it write a new version of the user manual. By the end of the weekend, he had generated a 150-page user manual that was clearly better than the existing version.

More importantly, now we can update it every month. So our user manual becomes a self-improving system. Every new piece of advice is compared against the existing manual, either absorbed or discarded. This way, the manual becomes a continuously updated living repository of the advice we give to founders weekly.

Of course, it doesn't stop at the user manual level. You can feed this as context into an AI agent. Suddenly, you can ask a super-intelligent AI a question and get the combined wisdom of 16 YC partners. But the prerequisite is that this knowledge must be readable by the AI. So you must record everything.

Point two is similar: If something can create a self-improving artifact that AI can read, keep it. If not, discard it.

Point three is that every function should be able to generate its own software. In the past, we might say "dashboards," but now it's not just dashboards; it's software generated on demand. Codex 5.5 is now good enough that for most simple internal software and dashboards, you can generate them to pretty high quality in one go. I tried it this weekend with some of our internal stuff, and the results were incredible.

So, all internal operations teams should sit above this layer: having an intelligent understanding of the business, and then generating their own dashboards and workflows.

And I would view this software as entirely disposable. What should be carefully preserved is the data. Like Garry said, he saves all emails as Markdown, never discarding anything. But the software itself is ephemeral, temporary. You can generate it and regenerate it.

What is truly valuable is the understanding of the business in people's minds: how this function operates, how we run a YC event, etc. As for the software actually used to execute the event, you can generate it for this event, use it, and then discard it. A month or two later, when the model gets smarter, you throw away the old software, give it the original instructions again, and generate a new version of the software.

So, I believe the valuable things are business context and skills. The software built upon them is temporary.

So what is the role of humans in this world?

I think we are discussing a "company brain." I know many people in this room are working on something similar. The middle part – all your data, emails, DMs, skills, know-how – is the company brain.

Humans are on the edge of this brain, responsible for interacting with the real world. That is, humans are where this intelligent system touches reality. Humans can enter scenarios that the model temporarily cannot access, like meeting someone in person, or novel, complex situations. I was going to say phone calls, but AI is easily entering phone call scenarios now.

More typical are unfamiliar situations, ethical judgments, high-risk moments. For example, a founder comes to us saying they are considering separating from their co-founder. In these truly high-risk, high-emotion moments, you will still want a human present.

This is the human's place. For many of your companies, this applies to sales conversations too. For the next 20 years, I think you will still need a human in the room for a live sales pitch.

So, I think humans will live at the edge of the company brain, responsible for bringing intelligence into the real world.

I've already run overtime; the host is probably about to pull me off stage. I'll leave you with one question: If you were starting your company again today, would you design it in this form from the start?

Most of you have companies small enough that you can absolutely do this. So I don't think you have any excuse. And I know there are several people in the room who are already in the process of tearing down and rebuilding their companies.

I'll stop here and hand it over to Pete. Thank you.

Video Link