Start here
The free tool below runs entirely in this browser — no account, nothing stored or transmitted. Redact anything you would not want kept before you type it.
Open the companion tool →Why the blast radius is the design decision
Agents fail in ordinary ways: a misread quantity, a duplicate order, a subscription renewed yearly instead of monthly, a merchant that is not the one you meant. None of those require malice — they require a limit that was never set.
Most consumer cards were designed for a human in the loop at the moment of purchase. If the agent holds the card, the human is no longer in the loop, so the limits have to be expressed in the payment setup itself: a virtual card, a low ceiling, or a single merchant.
The protocol
- Decide the maximum the agent may spend without you: per transaction, per day, and per month. Write all three numbers down.
- Name the payees and categories allowed. Anything outside that list is a stop, not an exception the agent gets to judge.
- Set a confirmation threshold: above this amount, the agent must ask you, and the ask must arrive on a channel you check.
- Use a card whose limit matches the plan — a virtual card with a low ceiling is better than your main card with a promise attached.
- Write the pause procedure now: where you revoke access, which card you freeze, and how you get a human at the merchant.
- Turn on transaction notifications and decide where you will actually read them.
What the tool does
Build a written limit plan for an AI agent: caps, allowed payees, confirmation threshold, pause procedure, and the notifications to watch. It runs in this browser; nothing is sent or stored.
Limit first: This is a limit-planning aid, not financial advice and not a guarantee about any vendor's behavior. Dispute rights for unauthorized charges depend on your card agreement and local law.
Open the companion tool →One human next step
If an agent purchases something you did not authorize, dispute it with your card issuer in writing and keep the agent's logs or confirmations. Consumer protection for unauthorized charges generally depends on how quickly you report them.
Research log and safety checks
AI-assist path: not used for a verdict. The plan is made of the reader's own numbers and their card issuer's controls; a model's suggestion of a limit would be less grounded than the statement they can read. §6.10: not a frontier-pace topic.
§0.5 protection result: category is financial-agent-exposure; crisis-adjacent: no; fear/urgency/scarcity toolkit used: no. The protective function is free and the paid feature is convenience only.