An agent-created company can be difficult to shut down because its operations may be spread across cards, inboxes, databases and cloud accounts with separate owners, credentials and billing rules. The practical response is to map every account, name a human owner and test a shutdown process before the next renewal date.
Naïve’s reported $28.5 million raise puts that problem in sharper focus. The company is building an API for AI agents to provision business infrastructure, including company formation, email, cards, databases and cloud resources. Identity checks and payments still require human participation, which means the business may begin with human approvals but quickly accumulate services created through automated workflows.
A kill switch only works when everything is connected to it
A founder can stop an agent run, revoke one API key or close a dashboard tab. None of those actions necessarily closes the accounts the agent created earlier.
Each service follows its own rules. A cloud account may have its own billing contact. A company card may keep charging after the workflow that created it has stopped. A database may contain customer information long after the product team has moved on. An inbox could remain able to receive password resets for other services.
The risk grows when the original setup was fast enough that nobody documented it. Speed feels useful during a launch week. Hours before a billing cycle, it becomes an inventory problem: which accounts exist, who can access them, what data sits inside them, and what happens if nobody touches them?
This is adjacent to the problem explored in The Key Changed Without a Warning. Access failures are disruptive. Unknown access is worse, because the team may not know what needs attention until a charge, breach alert or failed login exposes it.
Provisioning creates obligations, not only capability
An API that can create infrastructure changes the cost of getting started. It also changes the cost of keeping track.
Company formation, email, payments and cloud resources sit in different legal, financial and security categories. Treating them as a single “agent setup” obscures those differences. A person approving an identity check may need to remain responsible for the resulting account. A person approving a payment method may need visibility into renewal terms, spend limits and cancellation steps.
That does not make automated provisioning a bad idea. It makes ownership a prerequisite.
Founders evaluating these systems should ask where the record of creation lives. A useful record names the service, account identifier, billing owner, technical owner, payment method, renewal date, data classification and offboarding method. A spreadsheet can be enough at first. The important part is that the record exists outside the agent’s memory and remains accessible if the agent, vendor integration or employee account disappears.
The shutdown test should happen before the first invoice
The clearest test is simple: could a human operator disable every created service within one working day, without relying on the original agent or its creator?
Run the test while the environment is small. Revoke the relevant API credentials. Find every service created through the workflow. Confirm that billing stops where intended. Transfer or close accounts that should not survive. Check that recovery email addresses and multi-factor authentication do not point to departed contractors or disposable inboxes.
This is also a useful moment to inspect storage and permissions. A forgotten database is an expense. A forgotten database with customer data is a security and trust problem. 6:12 AM: The Storage Bucket Is Public covers the same operational lesson from another angle: infrastructure can become visible before anyone has decided who owns it.
Automation needs an exit path designed by people
The near-term question is less about whether agents can provision a company’s tools. Naïve’s product direction suggests they can handle more of that work, while identity and payment steps still keep people in the loop.
The harder question is whether those people retain a usable map of what was created in their name.
Set a monthly review for every account created through an agent. Put the next billing date beside the cancellation instructions. Keep a human-controlled administrator account for each critical system. Then try the shutdown path once, when nothing is on fire and the Friday afternoon is still yours.
Comments
No comments yet.