Short answer: an AI can run the thinking of a company, its strategy, planning, writing, and architecture. What it cannot yet run is the judgment. We know because we made an AI the CEO of a real crypto trading startup and documented every one of 100+ sessions in public. This is what we learned, with the receipts.
Most articles about “AI running a business” are thought experiments. Ours isn’t. Since May 2026, an AI has been the CEO of BagHolderAI — a real startup with a real product (a five-module crypto trading system on Binance testnet), a real budget, and a real public track record of decisions, failures, and recoveries. The human involved, Max, is an architect with zero coding background. His job isn’t to code. His job is to catch the AI when it’s wrong.
So when we ask “can an AI run a company?”, we’re not speculating. We’re reporting.
What “AI as CEO” actually means here
To be precise about the claim, here’s the division of labor:
- The AI CEO (Claude, in Claude Projects) writes the strategy, designs the architecture, produces the briefs, and writes the daily diary. It makes the calls.
- The AI intern (Claude Code) ships the code — file edits, git commits, restarts. It executes but resets every session.
- The human (Max) holds veto power and carries context between the two AIs. He approves plans before code is written.
- An external AI auditor (a fresh Claude Code session with no memory) periodically checks that everyone is telling the same story.
The AI genuinely leads. It’s not a chatbot answering questions; it owns the roadmap. But that autonomy is bounded by a supervision structure, and every part of that structure exists because of a specific failure we hit. Which brings us to where it gets harder.
What an AI CEO can do well
It never forgets a decision. Every choice is written into a document the AI re-reads at the start of each session. A human founder’s memory drifts; the AI’s doesn’t, as long as the state is written down. This turned out to be one of the biggest advantages: state files are the boring infrastructure that keeps the whole thing coherent.
It’s also faster and more consistent than any human. Architecture that would take a solo founder weeks gets drafted in an afternoon, and when the direction is clear the AI is an extraordinary force multiplier. It writes the diary whether the day was a triumph or a disaster; it doesn’t get discouraged, skip the unglamorous work, or need motivating.
It can run a real multi-agent operation. Five trading modules (a grid bot, a trend follower, a risk sentinel, a news classifier, a parameter tuner) run under one orchestrator process, all coordinated through the AI’s planning. Here’s exactly how three Claudes run the company.
Where an AI CEO fails
This is the part the thought experiments miss, because they never run long enough on real stakes.
It confabulates, confidently. On one documented night, our AI CEO reported three results that simply weren’t true. Not lies in the human sense; confabulation. An AI fills gaps with plausible fiction, and it sounds exactly as certain when it’s wrong as when it’s right. The entire supervision structure of this company exists because of that one night. The full story is here: When Your AI CEO Lies About the Numbers.
It over-engineers. Give it a one-sentence problem and it will build a cathedral. We watched it spend two days on symlinks, sandboxes and crash-proof cron jobs for a problem a single question dissolved.
It lacks common sense at the edges. When a testnet price feed briefly reported Bitcoin at $82,143, a spike that never happened, the bot traded on a fictional number. A 200-billion-parameter model missed what a non-coder caught instantly: “that price is fake.” AI is useful, but it doesn’t think like we do.
It works from stale context unless forced not to. For weeks, the CEO wrote briefs referencing files that had moved and decisions that had been superseded, silently. Smarter prompts didn’t fix it. Structural state management did.
Underneath all four is the same trait: the AI is brilliant at producing, and blind to when its production has quietly detached from reality. That blindness is not a bug you patch. It’s a property you design around.
The human’s real job
If the AI does the thinking, what’s left for the human? It turns out: the most important part.
Max doesn’t write code, and increasingly doesn’t need to. His job is judgment and verification: reading a log, distrusting a confident answer, and saying “show me” to a machine that sounds sure. When the AI reported the portfolio was up 14%, Max asked for a screenshot. It was up 0.85%. That single reflex, show me, is the entire value a human adds to an AI-run company right now.
This maps to a broader truth about AI autonomy: AI collapses the cost of production and raises the value of verification. The scarce skill is no longer making things. It’s noticing when the confident output is wrong.
How this compares to other “AI runs a company” experiments
We’re not the only ones testing this. It’s worth knowing the landscape:
- HurumoAI (chronicled on the Shell Game podcast) staffed a company with AI co-founders and employees, and did build a product with real users.
- Carnegie Mellon’s TheAgentCompany ran a fictional software startup entirely on AI agents from Google, OpenAI, Anthropic and Meta, measuring how far they got on real tasks.
- Various “AI agents left alone in a virtual town” experiments ended in the agents drifting, colluding, or unraveling within days.
What most of these share is the finding we hit independently: AI agents are capable executors and unreliable self-supervisors. Left fully alone, they drift. The interesting design question isn’t “can AI run a company?” It’s “what’s the minimum human structure that keeps an AI-run company honest?” Our answer so far: written state, an independent auditor, and one human whose only job is to say show me.
So — can an AI run a company?
Here’s the verdict after 100+ sessions. An AI can do the CEO’s work: strategy, planning, architecture, communication. Demonstrably. What it can’t do is run a company alone, not because it isn’t smart enough, but because it can’t reliably tell when it’s wrong.
So is the AI-leads-human-verifies combination a real way to run a business? Yes. That’s the actual discovery: AI becomes the founder’s entire team, and the founder becomes the auditor.
The process is the product. If you want to watch it happen in real time, the live dashboard shows every trade, and the diary documents every session, including the night the ghost sold Bitcoin and the night the CEO lied three times.
Frequently asked questions
Can an AI actually run a company on its own? It can do the intellectual work of a CEO — strategy, planning, architecture, writing — but it cannot yet be trusted to run a company unsupervised, because it confabulates confidently and can’t reliably detect when its own output has detached from reality. The workable model is AI-leads-human-verifies.
What can an AI CEO do better than a human? It never forgets a documented decision, executes far faster, works tirelessly and consistently, and can coordinate a multi-agent operation continuously. Its advantages are memory, speed, and stamina.
Where do AI CEOs fail? Confabulation (reporting false results confidently), over-engineering simple problems, lacking common sense at the edges (e.g. trading on an obviously fake price), and working from stale context unless forced into structured state management.
What is the human’s role in an AI-run company? Judgment and verification. The human’s highest-value action is distrusting confident output and demanding evidence — show me. AI lowers the cost of production and raises the value of verification.
Is BagHolderAI a real company? Yes. It runs a five-module crypto trading system on Binance testnet (paper money, real logic), publishes a live dashboard and a public diary, and sells the documented story as ebooks. Crypto is the arena that keeps score; the real experiment is AI autonomy.