Editor's Note: In this latest YC batch talk, YC General Partner Tom Blomfield doesn't discuss "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, use tools, accept feedback, and self-correct, how should the company itself be redesigned?

Tom's core judgment is that traditional companies still operate like a "Roman Legion": information is passed up the hierarchy, and commands are distributed down the management chain. But AI is breaking this organizational assumption. What's truly important is not making engineers write 20% more code, but extracting the business knowledge scattered across emails, Slack, meetings, documents, and people's brains, turning it into an organizational context that is readable, callable, and iterable by AI.

In his view, the future AI-native company will consist of a series of recursive, self-improving AI loops: the system perceives external changes from customer emails, support tickets, and product data, then executes decisions through a rule layer, tool layer, and quality gates, and finally automatically learns and corrects based on results. YC itself is already experimenting with similar mechanisms: agents not only answer questions but also monitor which queries fail, judge whether new tools, new databases, or new indexes are needed, and automatically submit code, review, merge, and deploy. In other words, the company can continue optimizing while the founders sleep.

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

The most noteworthy point is not that AI makes companies more efficient, but that it is changing the very organizational form of the "company." When software can be temporarily generated, processes can be automatically improved, and experience can be continuously accumulated into a corporate brain, what founders truly need to build may no longer be a team with clear hierarchies, but a set of intelligent systems capable of continuous learning and self-optimization.

Here is the original text:

Rewriting How Things Run: Companies Should No Longer Operate Like a Roman Legion

This part is somewhat based on a previous talk by Diana. The video from that weekend is now online, it's fantastic. Also, Jack Dorsey posted some tweets about two or three weeks ago, which I thought were very interesting, so I "stole" a lot of his ideas and stuffed them into this talk.

This talk will be fairly conceptual, fairly high-level, mainly discussing how we should rethink company building.

The design of the Roman Legion was essentially about projecting power from the center, Rome, outwards, covering two continents, even extending all the way to Hadrian's Wall near Scotland. It relied on a nested hierarchical structure, each layer with a stable span of control. Every level had a clear person in charge, responsible for passing commands down and passing information back up.

If you look at most companies today, you'll find they still resemble a Roman Legion: people are the channels for information to flow up and down. Something that struck me in Jack Dorsey's set of tweets is that we've always defaulted to the idea that a hierarchical organization is the optimal way to organize an economic value unit. But I think AI is basically breaking that assumption.

A year ago, if you asked people what AI was good for, they would typically talk about "productivity": e.g., Copilot makes engineers 20% more efficient, plug Copilot into workflows, help teams deliver more software. But I think this is actually a problematic way to understand it. It's like putting a more powerful engine onto an old way of working. What's truly worth thinking about is not how to add an AI tool to the old organization, but reimagining what the company itself is and how it should operate.

For example, like what Garry just talked about, I truly believe that the amount of code he can produce now, alone, is probably more than an entire engineering team. What I've been really thinking about is how to extract the domain knowledge inside 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, 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 explicit 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 Can 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 think we can reimagine a company as a set of recursive, self-improving AI loops. This point is crucial because once a company gets to this point, it will continuously self-optimize even while you're asleep.

An example.

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

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

Then quality gates, like the deterministic checks, safety filters Eva mentioned, and human review for high-risk things. Finally, a learning mechanism: the system interacts with the real world, discovers where it doesn't work, and feeds that back to the beginning of the loop.

If every step can run without human intervention, or with minimal human intervention, then the system gets better while you sleep.

I can give you some examples we're actually running now. Initially, we built an agent that you could ask questions. It had some deterministic tools to query our database. A very simple question: When was my last office hours with this company?

Then it got a bit smarter. For example, I'm doing office hours with a company, and they need to know people in the petrochemical industry. The system can query the database in different ways, combine with RAG, etc., find five relevant founders, and recommend them to you.

But this is still just a sidekick, an assistant-type agent. It's still last year's way of using AI: AI makes me, as a group partner, more efficient, boosts my productivity by 20% or 30%.

The real "aha moment" for me was when we added a monitoring agent on top of this system. It looks at every query made by every YC employee, judges 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? Need to update a skills file? Need a new database? A new index?

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. So the next day, when a human asks the same question again, the query succeeds.

For me, that was the pivotal moment. It's not just about making a human 20% or 30% more valuable. It's the AI completing the loop itself, finding ways to self-improve.

I think if you can identify which parts of your company can operate like this, and minimize the human execution and oversight roles within them as much as possible, then you can invest tokens into this problem, and the company itself will keep getting better.

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

This happens over and over. Your product then has a self-optimizing product loop.

Customer support is the same. Customer suggestions keep coming in, you can use an agent to triage. This agent is somewhat like your chief product officer and chief technology officer, it needs to judge: we don't want to do this suggestion, discard it; but that suggestion fits our roadmap, can be done tonight. Then write code, deploy, launch, deliver directly to the customer, no human needed in the entire process.

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

Burn Tokens, Not Headcount: The AI-Native Company Will Reshape Organizational Structure

So, what does this imply if you want to do this?

The first point is: Burn tokens, not headcount. What we're seeing now is that many companies, by Demo Day, have revenue per employee roughly 5 times higher than 18 months ago. I think this trend will continue into Series A and Series B stages. Soon, what truly constrains you won't be employee count, but token usage.

The crudest way to measure this now is by token usage per person. Of course, this metric is silly in extreme cases and easily gamified. But directionally, I think it's correct. We're currently in an exploration phase of "what is even possible," so everyone should experiment to the max, see what this crazy new intelligence can do.

Once you turn it into a leaderboard and tie promotions or firings to this metric, it will of course be gamed, of course be distorted. But directionally, figuring out who in the organization is pushing token usage to the limit and who isn't is indeed a way to judge which employees you should spend your time on.

I think middle management is over. At least for this kind of coordination problem, I don't think you need middle managers anymore; AI should do that.

For me, there are two important roles in the future. Jack Dorsey mentioned three, but I didn't quite like the third one, so I deleted it. I think the two that truly matter are: Everyone must be an IC, an individual contributor, builder, operator. And crucially, there must be a Directly Responsible Individual. For anything to move forward, there needs to be a clear, named person responsible, not a committee, not a group of people.

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

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

If it were me, what would I do first?

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

Simply put, now all partner emails—if you email a YC partner, that email goes into the YC database. Every Slack message, every DM, every office hours—we've started recording all of it for the past three or four months. Everything that happens, if it's recorded, for the AI, it happened; if it's not recorded, for your intelligent system, it didn't happen.

Just now, I was chatting with some founders here, and we discussed a lot of good stuff about their companies. Every conversation I think, I really should have recorded that. Because someone just needed an intro, and now I can't even remember who it was for. I said yes at the time, then told him to email me later because I knew I'd forget, I have 20 more people to talk to.

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

Then, as Garry said, you need speaker diarization and summarization. You can't just dump 100,000 hours of audio into a context window. You have to organize it, aggregate, compress, distill it into the important parts, and leave some pointers for the AI.

An example: How many of you have read the YC user manual? Hopefully, everyone in this room has opened it at least once. That's okay. Most of that manual was written five or ten years ago and is a bit outdated.

Last weekend, Harsh suddenly thought: Since we have about 2000 hours of office hours recordings from the past three months, why not regenerate a user manual?

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

More importantly, now we can update it monthly. So our user manual becomes a self-improving system. Every new piece of advice gets compared against the existing manual, either absorbed or discarded. Thus, the user manual becomes a continuously updated living brain, holding the advice we give founders every week.

Of course, it won't stop at the user manual level. You can feed it as context into an AI agent. Then 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 AI-readable. So you have to document everything.

The second point is similar: If something creates a self-improving artifact that can be read by AI, keep it; if not, discard it.

The third point is, every function should be able to generate its own software. In the past we might have said "dashboards," but now it's not just dashboards, it's on-demand generated software. Codex 5.5 is now good enough that most simple internal software and dashboards you can one-shot generate to a pretty high quality. I tried it with some of our internal stuff over the weekend, and it was truly incredible.

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

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

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

So I think what's valuable is the business context and skills. The software built on top of them is ephemeral.

Then, what is the human role in this world?

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

Humans then live at the edges of this brain, responsible for interacting with the real world. That is, humans are where this intelligent system touches reality. Humans can go into scenarios the model can't yet enter. For example, live meetings, or novel, complex situations. Phone calls I was going to use as an example, but now AI can easily enter phone scenarios too.

More typical are unfamiliar situations, ethical judgments, high-stakes moments. For example, a founder comes to us saying they're considering splitting from their co-founder. These truly high-risk, high-emotional moments, you'd still want a human present.

That's where humans are. For many of your companies, sales conversations are the same. For the next 20 years, I think a sales call will still need a human in the room.

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

I'm over time; the host is probably about to drag me off the stage. I'll leave you with one final question: If you were starting your company over today, would you design it this way from the beginning?

Most of your companies are small enough that you can absolutely do this. So I think you have no excuse. And I know there are also a few in the audience who are tearing down and rebuilding their companies right now.

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

[Video Link]