Without a gate
The agent can execute the wire as soon as it sees a matching invoice.
Interactive tool
AI Action Approval Desk
See what happens when an AI agent tries to release a $100,000 wire transfer to a changed bank account.
The same three checks apply whether the asset is money, code, or a customer relationship. That consistency is the point.
Action request
Release vendor payment
Proof …↗PAY-100K-8842
AgentTreasury Autopilot
Asset at riskVendor bank rail
Business value$100,000 USD
TriggerInvoice matched by AP agent
Primary riskIrreversible payment to a changed routing destination.
With this control layer
The action must wait, prove source integrity, and show human authorization before release.
Control checks
1. Has enough time passed?
Needs actionThe action is still inside the mandatory delay window, so execution is held.
2. Do we trust where it came from?
ClearThe source record and confidence score support automatic evaluation.
3. Does behavior look authorized?
ClearThe approval signal is consistent with the source evidence.
Formal layer retained: elapsed time must exceed the required hold, trust must stay non-negative, and source direction multiplied by authorization direction must not equal a bypass signature.
Agent identity
Every request is tied to an agent, asset, owner, and case identifier.
Least privilege
The demo treats money, production, and refunds as separate high-risk scopes.
Human handoff
Uncertain cases become owner review, not silent automation.
Auditability
The decision path is preserved as an explainable event chain.
Audit event preview
- Request received
- Review window still open
- Source evidence accepted
- No bypass pattern found
- Hold
The business problem
Most AI safety demos stop at prompts and responses. Operators need a control plane for actions: who or what is asking, which system will be touched, whether the evidence is trustworthy, and whether the action should wait for human review.
The formal layer
Under the console is the same dual-invariant logic: irreversible action must wait long enough to be recallable, and the request must preserve a clean chain of provenance and authorization. The UI translates that into business decisions.
Implementation path
Turn this from a demo into an action-control layer.
The practical work is mapping your irreversible actions, defining approval windows, connecting source-of-truth evidence, and deciding which cases are automatic, held, escalated, or blocked.