← Decision Lab
03Autonomy

Blast Radius

What is the worst thing one wrong step can do?

  • Permissions
  • Autonomy
  • Human approval
  • Reversibility
  • AI safety

Interactive prototype

Open in a new tab →
https://blast-radius-sim.lovable.app

If the prototype doesn't load here, open it in a new tab.

The situation

An AI executive assistant starts read-only over a calendar. It's useful, so it gets email. Then files. Then the CRM. Then the ability to spend.

Each grant is individually reasonable. The combined surface is not.

The obvious answer

Give the agent broad access so it can be genuinely useful, then trust the model to behave.

The question underneath it

Every permission is a design decision about consequences, not convenience. Authority should be granted per action, and the cost of a mistake should stay bounded no matter how capable the model gets.

What I'm paying attention to

  • Reversibility

    Drafting is undoable. Sending, paying, and deleting are not.

  • Audience

    Internal mistakes are embarrassing; external ones are permanent.

  • Approval gates

    The right question is which actions cross a human, not whether any do.

  • Scope of credentials

    Read access to a folder is not read access to a drive.

  • Detection

    Autonomy without logging is autonomy without accountability.

My take

More autonomy is not more sophisticated. The interesting systems live between "AI suggests" and "AI runs the company," and staying there is a deliberate engineering choice.

I prefer bounded autonomy: wide read, narrow write, explicit approval on anything irreversible.

Related writing

How Much Power Should You Give an AI Agent?

Substack essay — Coming Soon

Office Hours with Chanel

Working through a version of this in your own company?

Private AI office hours for leaders who need to turn messy AI ideas into clear strategy, useful workflows, or safe internal agents — with senior-engineer judgment in the room.