I’ve spent most of my adult life building companies, investing in them, and trying to understand why some work and others die. These days I spend most of my time thinking about a bigger question: What does entrepreneurship look like when intelligence becomes almost free? AI is changing the cost of building, the size of teams, the value of labour, the infrastructure underneath companies and eventually, I think, the economic system around all of it. We’re testing a lot of these ideas in real time across Bifrost. I’ll share what we learn here. Including the things we get completely wrong.

There is going to be a hilarious period where companies have AI agents doing the work, humans checking the agents, managers checking the humans checking the agents, and consultants explaining the transformation. We’ll figure it out eventually.

Your AI agent doesn't need another meeting. Doesn't care about titles. Doesn't need information slowly passed through four layers of management. Once agents become a meaningful part of the workforce, I think we'll discover how much company structure existed to coordinate humans.

What happens when a company can increase its productive capacity without increasing headcount? We have spent decades using employees as a rough proxy for scale. AI is going to make that increasingly useless

Every AI demo looks incredible for five minutes. Then you put it inside a company. Now it needs context, permissions, reliable data, integrations, security and some understanding of what the business is actually trying to accomplish. The model is only one piece of the system.

A startup can survive a mediocre product for longer than people think. What it rarely survives is learning too slowly. Wrong customer. Wrong pricing. Wrong channel. Wrong hire. None of these are necessarily fatal if you find out quickly enough. The dangerous companies are the ones that can spend 18 months being wrong without anything forcing them to notice.

I think a lot of AI startups are accidentally features waiting for the next model release. If your entire advantage comes from making the model do something the model itself is rapidly getting better at doing, I would be nervous. Build where the underlying models getting 10x better makes your company 10x better too.

AI should make company building more experimental. At Bifrost, ideas move through concept, product and GTM validation before spin-out. Bifrost Studios. As the cost of testing falls, founders should be willing to run more experiments and throw away more answers. Being wrong is becoming cheaper. Staying wrong is still expensive.

Most companies still treat AI like a software purchase. Buy the licenses. Run the pilots. Add it to existing workflows. Show the board that you have an AI strategy. I think that misses the interesting part. If AI actually works, you redesign the workflow around what it can do. You change where decisions happen, what humans spend their time on and eventually how the company itself is structured. That is a much bigger project than buying ChatGPT Enterprise.

One thing I think VC gets wrong is treating every great company like it should eventually become a venture case. Some companies have enormous returns to scale and should take capital aggressively. Others can build fantastic businesses, generate cash and create serious wealth without ever needing that trajectory. The financing should fit the company. The company shouldn't have to contort itself to fit the financing.

The next big AI opportunity might be removing software from the workflow. A lot of software still requires a human to open it, understand it, navigate it and tell it what to do. Agents change that assumption. The best software may increasingly be the software you barely interact with.

VC is very good at pricing companies after they exist. I’m more interested in what happens before there is anything to price. Finding the market. Finding the founder with the unfair advantage. Killing weak ideas. Getting the first customer. Building the operating system around the company. There is a lot of value created before a term sheet ever appears. That's the part of venture I really like.

I’m increasingly convinced the biggest mistake with AI inside companies is starting with the tool. “Where can we use AI?” is how you end up with 14 pilots and a very impressive internal presentation. Start with the work. What is expensive? Slow? Repetitive? Bottlenecked by expertise? Then decide whether AI belongs there.

AI makes small markets interesting again. A problem can be very real and still not support a traditional software company once you add engineering, sales, support and overhead. Change the cost structure and suddenly the TAM calculation changes too. I think we’re going to see a lot of companies built for markets VCs previously considered too small.

One of the more interesting models we’re building around at Bifrost Forge is service → software. Do the work for the customer first. See every ugly edge case. Learn the workflow. Get paid while you learn. Then automate what you now understand better than almost anyone. AI makes that progression much more powerful.

I think “human in the loop” will age badly as a blanket AI principle. Sometimes the human is essential. Sometimes the human is the bottleneck. The interesting work is figuring out which is which.

One thing AI should kill is the startup back office. A founder should not spend the first months of a company rebuilding finance, recruiting, GTM systems, websites and basic infrastructure from scratch. At Bifrost, those functions are shared and every company starts with infrastructure already running. There are much harder problems founders should be spending their time on.

Over 1,000 followers here, so probably time to reintroduce myself. I’m Thorbjørn Rønje, GP at Bifrost Studios, where we build venture studios around high-conviction industries and use them to create companies from scratch. We’re currently at 7 studios and 33 companies built. I’m also involved with Granite, where we’re building sovereign cloud infrastructure for Europe, and Hyperion, a quantitative hedge fund built around models rather than analysts. Most of what I write about here sits somewhere between company building, AI, infrastructure, capital and what the next economy looks like. Plenty to figure out.

I think AI will make domain expertise more valuable in company building. Building gets cheaper, which means more people can attack the same problem. The founder who has spent ten years inside the industry knows where the bodies are buried. That knowledge doesn't come with the model.

We may be building the first generation of companies where headcount tells you almost nothing about capability. For most of modern business history, there has been a fairly reliable relationship between what a company could do and how many people worked there. A larger customer base required more support. More products required more engineers. More transactions required more operations. Growth eventually produced organisation charts because human coordination was part of the cost of scale. AI starts pulling those relationships apart. If agents can perform meaningful parts of research, coding, administration, analysis and eventually much more operational work, capability can grow without headcount growing at the same rate. The company does not simply become more efficient. Its basic organisational architecture starts changing. This is something we think about a lot at Bifrost because we are already designing companies around shared capabilities, small teams and AI rather than assuming every business needs to recreate the same functions internally. I suspect we are still very early in understanding what this does to company building.

One thing we deliberately push at Bifrost is decision-making at the edge. The operator closest to the problem should have the authority to act. AI gives small teams more information and more capability. Adding three layers of approval on top of that feels like bringing a fax machine to a Formula 1 race.