AI capabilities are continuing to ascend. Coding models crossed a clear threshold at the end of 2025, some programmers no longer write code manually at all. The same transition will play out for many other skills in the next few years. There is wide disagreement about how this will propagate through the economy. Some imagine an instant discontinuity, others expect no meaningful human substitution at all.
To think through how the transition might go, it helps to get very concrete. For most of the last ten years I ran a startup with around 100 employees. What would it take for every role in that company to be replaced by AI?
Capabilities
First, let’s briefly discuss the AI capability ramp for anyone who remains skeptical. AI has been improving broadly, but the most spectacular results have been in coding and mathematics. This is largely because AI labs have chosen to invest reinforcement learning effort in these areas. They make good initial targets because they are economically valuable, and the RL data is easy to collect. It is mostly a matter of turning the handle to do the same in many other domains, and the data collection efforts to do that are well underway.
Programmers and mathematicians are loud about how far things have moved in the last few months. Any discrete task solvable with ‘known techniques’, or even deep non-obvious combinations of known techniques, is now usually tractable for AI. This encompasses almost everything people do in a typical project. Reliability isn’t perfect, but it’s good enough to make you completely rethink how work is done.
Even with no further AI progress, the recent experience of coders and mathematicians will be replicated across the economy over the next few years. This is the baseline, and it will feel radical to most people. On top of that, AI will continue to improve. Datacenter investments that have already been made will flow through to capabilities with a lag. There will also be algorithmic improvements, maybe even an unlock around continual learning. By the time most people experience all of this in their own jobs, it’s going to feel even more surprising than what coders just went through.
This is not AGI. One of the clearest things you can see in AI mathematics is a lack of true conceptual novelty. An algorithmic breakthrough will likely be needed to change that1. But AI is nearly superhuman at anything involving known techniques
Running a company with AI
Let’s fast forward to 2028. AI has continued to ramp, and RL has been applied to most economically valuable tasks. What does it look like to run a company in this world? I’ll go through all the different roles in my old company (Pointy) and estimate what AI could replace.
Customer Support
This one is already happening. Intercom’s Fin agent has a resolution rate of 73%, rising at 1% per month2. Two years from now any text based customer support is going to be almost entirely automated.
However, once you get into the messy details of a real company, it’s not so cut and dried. Text based support was maybe 25% of our workload, the rest was mostly outbound phone calls, which we found the most efficient way to resolve user issues with our hardware device. There’s no technical reason an AI voice agent couldn’t make these calls, but would customers accept it? For a use case like this where they’re actively seeking assistance, I would tentatively guess they would. But even a small drop in our case resolution rate could easily justify the cost of a human employee. We also had a field service technician for physical troubleshooting, obviously that one remains AI-proof.
In total we had 8 people on customer support of various kinds. My guess is that with AI we could have done it with 2, but with high uncertainty on that.
Headcount: 8 → 2 Reduction: 75%
Data Operations
We had a substantial data-cleaning operation using remote contractors. This involved taking messy product catalog data and performing cleanup and QA. This is a straightforward fit for modern models.
Headcount: 10 → 0 Reduction: 100%
Engineering
Change here is ‘already happening’, but the headcount impact is hard to predict.
We had a small team of very strong engineers. Essentially every engineering team has more work than capacity to do it – we would have hired more if I could have found candidates who met our talent bar. I think with AI we would have had a team the same size and just shipped more, faster.
One exception: we had a dedicated data scientist. AI is particularly strong there, so I think we would have skipped that hire and absorbed the work into the broader team.
Headcount: 12 → 11 Reduction: 8%
Product
You could approximate an average product manager with AI. But the best product work depends on taste and creative insight, which is exactly where we see current AIs struggling in frontier mathematics applications. So I think this remains a human role for now. I can imagine that changing over time.
Headcount: 2 → 2 Reduction: 0%
Inside Sales
This was our largest headcount, and the area where I have the most uncertainty.
Our sales were mostly phone based, with some email. It was highly structured, with call scripts, clear funnel stages, detailed measurement. In principle, an AI agent would fit. But the most essential requirement is that the person picking up the phone is willing to talk to you. Will anyone want to take an AI sales call? And if they do, it seems like the volume would explode until they don’t?
Sales tactics are always a kind of Red Queen race. There could be a window for early adopters to drive growth with AI-only sales, but it doesn’t seem durable. I honestly don’t know what our sales model would look like in a world of AI agents.
Headcount: 50 → ? Reduction: 0 – 100%
Partnerships & Enterprise Sales
This is easier. Partnerships and enterprise sales rely heavily on in-person meetings. This is straightforwardly human-required work.
Headcount: 4 → 4 Reduction: 0%
Marketing
Some parts of marketing – ad campaign management, creative, copywriting – are already being automated. Others, like trade shows, remain AI-proof. The latter was an important channel for us. On net, I think AI would give us productivity gains with only modest headcount reduction.
Headcount: 5 → 4 Reduction: 20%
Finance
The basics of a finance function – tax filings, management accounts, simple forecasts – are already well on track to be automated. More advanced finance functions will surely follow.
We made our first dedicated finance hire shortly before Series B. With AI we would likely have pushed that out further.
Headcount: 1 → 0 Reduction: 100%
HR
Much junior HR work is process driven and could be automated. This might reduce HR headcount in large firms. At startup scale we only had one HR hire, and I would still make that hire. Otherwise the complicated people problems that inevitably crop up as your team grows end up being dealt with by the founder directly, and consume a large amount of time.
Headcount: 1 → 1 Reduction: 0%
Office Manager
At first glance this seems AI-proof, since it involves lots of interaction with the physical world. However, Andon Labs has an AI office manager that uses building security cameras for real-world context. This is clearly still at experimental stage, but at our scale the role was somewhat optional, so we might well have tried an AI alternative.
Headcount: 1 → 0 Reduction: 100%
Logistics & Operations
Our hardware device required physical assembly, testing and dispatch. Clearly AI-proof for now.
Headcount: 2 → 2 Reduction: 0%
Leadership
So long as a company has human employees, it will likely have human leadership. A lot of the job, both internal and external, is human-to-human. Other parts of the job could be offloaded to AI, so it’s possible we see fewer COO hires, but I don’t think it will be a large effect.
Headcount: 3 → 3 Reduction: 0%
Total
Headcount: 99 → 79 Reduction: 20%
So overall I’d guess the AI of two years from now, when fully digested, might result in a 20% reduction in our headcount. Most of this is in low level roles. The more meaningful impact is the productivity boost everywhere else. The big uncertainty is around sales – if that could be done effectively by an agent, the headcount reduction could be as much as 70% in our case.
Mechanics of a Phase Change
When water freezes it doesn’t happen everywhere at once. A few ice crystals form here and there where the local conditions are most favourable, then expand gradually. Most of the bulk can remain liquid for a long time after the temperature reaches freezing point.
Major technology transitions behave the same way. Retrofitting existing companies is slow work. AI-native startups can move faster by designing around agents from day one, but even they exist in the context of a broader economy that is built for humans, not agents.
What would it take to run a company with close to zero human employees? A few classes of work remain AI-proof in the short term, so a zero-employee company would need to avoid:
- human-to-human interaction
- physical labour
- context that AI can’t access (e.g. informal hallway chat)
A typical tech startup can mostly sidestep the latter two, but avoiding all human-to-human interaction is harder. Initially only very simple digital businesses will fit3.
What about a more complex company? You can control your internal surfaces – how agents are integrated, how decisions flow. You cannot control your external surfaces. Those are set by the environment. Sometimes you may get away with an external agent-to-human interaction, but how well would your enterprise sell or your VC fundraise go, if you send your agent to meet their human? This is the coordination problem that sets the speed limit for transition to a fully AI economy.
I can’t fully delegate to my agent until you fully delegate to yours
AI-only companies are clearly the destination. For these companies to exist, we need to reach a world where agent-to-agent interactions across company boundaries are the norm.
To imagine the path from here to there, let’s go back to the example of my sales team. Say we ran some tests and discovered that an AI agent could do this job for us. If it works for us, pretty soon everyone will be doing it and our retail clients will be drowning in AI calls and emails. This agent-to-human stage is awkward and unpleasant for everyone. So maybe the store owner gets an AI agent to screen their calls and emails. Now instead of a human-to-human interaction, we have an agent-to-agent one. In principle this could be a productive sales call. Our product had a clear ROI, a good buy-side agent could surface it as a recommendation to a human decision maker, or eventually even make an autonomous purchasing decision.
The same transformation can play out at every level. As CEO I spent a lot of time fundraising. Since VCs are human, this is not something an agent can do for me. But this can change gradually. Initially maybe all of that inbound from VC associates might be replaced by VC agents. Then I get my agent to handle the noise. Now my agent is pitching their agent. For a while this is just a layer between humans, but eventually their IC comprises Claude, ChatGPT, Gemini and Grok, and the fund is fully autonomous.
The phase change needs time to spread through the economy, even absent political reactions. The percentage of agent-to-agent interactions will rise slowly over many years.
But if I was starting a company right now, I would be building for agents as my customers. They are going to be the fastest growing, highest spending customer segment in history, and right now they are very poorly served. Some of this is just plumbing, some requires a full rethink of entire business categories. There are huge opportunities to build.
So hang out on Moltbook and see what the agents are frustrated by. Then build that.
Footnotes
- My hunch is that it may come for free with true continual learning, but that’s very speculative. ↩︎
- Sources. ↩︎
- Something roughly on the scale of a Pieter Levels project. ↩︎