The Law Already Assumes Your AI-Run Company Needs a Human in Charge
Argentina's bill for AI-run companies still requires a human administrator liable for outcomes. What that means for anyone building with autonomous agents.
Founder, Task Machine
Argentine President Javier Milei announced, in a Financial Times op-ed, a plan to let AI run a company with no human employees at all — agents "exercising independent judgment in unpredictable environments," a category of "non-human corporation" no country had legislated before. Sam Altman has been pitching a version of the same idea for two years: a single person, AI doing the rest, a billion-dollar valuation. It reads like the moment the zero-headcount company stops being a thought experiment.
Then you read the bill.
The "automated company" the reform actually creates requires a human administrator to oversee its operations. The company's administration can use AI to make decisions, but that does not exempt the administrator from supervising the outcome. The company stays liable for damage its AI or algorithmic systems cause. Reuters' reporting on the bill quotes Lawrence Cunningham, who directs the Weinberg Center for Corporate Governance, calling it "too wild a first step to dispense with human agency entirely." Even Yuval Noah Harari's public worry — that this reduces corporate accountability — is a worry about a law that, on its face, does the opposite.
Argentina is not an outlier here. Texas and Utah have both built legal frameworks for AI-run businesses, and both lean toward more human oversight at the start, not less. Every jurisdiction that has actually tried to write rules for this, rather than just describe the vision, has landed on the same place: agents get real authority, and a specific human stays accountable for what they do with it.
What "AI-run" turns out to mean once it is law
The gap between the pitch and the bill is not a failure of ambition. It is what happens when "the AI runs the company" has to survive contact with the question a lawyer always asks next: if this goes wrong, who is responsible?
"Independent judgment in unpredictable environments" is a fine sentence for an op-ed. It is not an answer to that question. The bill supplies the actual answer: a named administrator, whose job is explicitly to supervise the outcomes of AI-made decisions, and a company that carries liability for what the AI does. Diego Duprat, who co-authored the bill, points out that automated companies already exist informally — cashier-less supermarkets are one — without anyone calling them "AI-run" or needing new legislation for them. What the bill formalizes is not automation itself. It is who answers for it.
That is also the honest shape of controlled autonomy inside a workflow, not just inside corporate law. Real autonomy was never "the agent decides and nobody is accountable." It was always going to be "the agent decides within a scope someone defined, and a specific person is accountable for what happens at the edge of that scope." Argentina just wrote that shape into statute before anyone had to argue about it in a courtroom.
What every version of this keeps asking for
Strip away the jurisdiction and the same three requirements show up every time someone tries to make "AI-run" concrete instead of aspirational.
| Requirement | What it looks like in the Argentina bill | What it looks like operationally |
|---|---|---|
| A named accountable person | A human administrator, required by statute | An approval owner — a specific person, not "the team" — for each class of consequential action |
| A bounded scope of authority | AI may decide, but the administrator supervises the outcome | An explicit line between what an agent may do alone and what needs sign-off before it takes effect |
| A legible record | The company is liable for AI-caused damage, which means someone has to be able to show what happened | Step-level history of what an agent did, in what order, and what it was allowed to do at the time |
Yonathan Arbel, who researches AI law at the University of Alabama, suggested the bill would benefit from giving AI agents a digital ID for their interactions — a record of which agent did what, attributable after the fact. That is the same instinct as the third row of the table, just aimed at a courtroom instead of a workflow review. Even in the bill's decentralized-autonomous-organization provisions, aimed at token-based, deliberately anonymous governance, the requirement that token holders be identified ran into resistance from crypto veterans for exactly this reason: remove the record of who decided what, and you remove the thing regulators actually need.
None of this is unique to corporate law. It is the same list a client asks for before trusting an agency's AI-run deliverable pipeline, or a co-founder asks for before letting an agent touch the company's Stripe account. "Who is accountable, what was it allowed to do, and can you show me what happened" is the question underneath all three.
The honest limitation
Nothing here is legal advice, and nothing in an operating layer makes a company legally sound on its own — that is still a matter for actual counsel, and, per this bill, actual legislation that has not passed yet. An autonomy setting is not a liability shield. What it can do is make sure that whichever human ends up accountable — an administrator under a future law, a founder today, a client-facing lead at an agency — already has the authority boundary and the record to back up what they signed off on, instead of reconstructing it after something goes wrong.
Where this fits in Task Machine
Task Machine is built around the same three requirements, independent of what any legislature eventually decides. Every agent carries a named autonomy level — supervised, balanced, autonomous, or full — that fixes how much it can do before a person has to approve it, and that level is set and changed deliberately, not left implicit. Consequential actions route to a specific approver in the inbox rather than to "whoever notices," so accountability has a name attached before anything ships. And every run leaves step-level workflow history — what was attempted, what was approved, what changed — so the record a regulator, a client, or a co-founder would ask for already exists instead of needing to be reconstructed.
The pitch was never a company that runs itself while accountability evaporates. Argentina's bill did not build that either, and neither does Task Machine. Both are trying to answer the same question the AI-run-company narrative keeps skipping: agents can do the work, but someone still has to be able to say who is answering for it.
If you want agents doing real work inside boundaries you could explain to a regulator, a client, or your own co-founder, join the private beta on the waitlist.