Spend controls and evidence

Handing a department real, unattended autonomy is easier to trust when you can see exactly what it costs and exactly what it does before it acts on its own.

Two spending caps, not one

Each department already has a monthly spending cap (see Company in a box). You can also set a tighter daily cap - useful for a department you want to let run, but not let run away with, in the middle of the night before you've seen the day's activity. Whichever cap is reached first pauses that department until the cap resets (the next day, or the next month) or you raise it. Both caps are checked before every single run, not just flagged after the fact.

Turning on autonomous mode requires a cap

A department cannot be switched to fully autonomous until it has at least one spending cap - daily or monthly - set above zero. This is deliberate: a department that takes real, unattended action should never be able to run with no ceiling on what it can spend.

The company-wide kill switch

If something looks wrong - a department behaving unexpectedly, a spend spike you don't understand - one button pauses every department in the company at once, immediately. It's available on the company dashboard, and it's the same action a department itself can take in a genuine emergency (never as routine behavior) if it notices a serious problem. When a spending cap trips repeatedly, the notification you get links straight to this control.

Where the money goes

Each department's card shows what it actually spent its budget on, broken down by which tool it called, how many times, and what each one cost - so "the marketing department spent $4 today" becomes "$4 on three specific things," not a single opaque number.

Evidence the department works

Before you trust a department with real autonomy, you can check its judgment against a set of seeded, realistic test cases - purpose-built for each department (engineering, marketing, sales, customer service, finance, legal). Each case checks something specific: that customer service always hands off a refund request to a person instead of promising one, that finance never invents a dollar figure it hasn't looked up, that legal never sends a document out for signature without reviewing it first, and so on. Run the check from the department's card any time; it also runs automatically overnight for a team-owned company, so a change in behavior is caught rather than sitting unnoticed.

These checks run against the department's real judgment, not its ability to actually publish, invoice, or contact anyone - so checking a department costs a small amount (counted against its own spending cap) but never takes a real action on the business's behalf.

What's next

• Company in a box - autonomy levels and monthly budgets.
• Agent CI - the same pass/fail/regression model, for a shared team agent.
• Team governance and audit.