Modes and safety

An assistant can do anything you can do, refunds included. This page is what you need to know before letting it.

Practise in your sandbox

Every brand has a sandbox: a parallel copy of your business with fake bookings, fake clients and fake payments. Nothing in it is real and no money can move in it.

Ask the assistant to work in test mode and it will:

Use my sandbox, then cancel tomorrow's squad and refund everyone.

This is the one protection that is enforced rather than requested. Arvald itself decides which dataset a request touches, and the separation is held by the database, not by the assistant's good behaviour. An assistant working in your sandbox cannot reach live bookings even if it tries.

If you are unsure which mode you are in, ask. The assistant can always report it, and it is worth asking before anything irreversible.

Confirmation is a convention, not a guarantee

Actions that move money or destroy something are marked so that assistants know to ask first, and Claude and ChatGPT both honour that today.

It is worth understanding what that marking actually is: a hint the assistant chooses to respect. Arvald asks the assistant to confirm; it cannot force it to. An assistant that ignored the hint, or a future version that handles it differently, would not be stopped by Arvald.

So:

• For anything reversible, the hint plus your own attention is fine.
• For refunds, cancellations and anything touching money, read what it proposes before approving.
• If you want a guarantee rather than a convention, work in the sandbox.

Disconnecting

Removing the connector in Claude or ChatGPT stops it making new requests.

It is not instant revocation. The access token the assistant already holds stays valid until it expires, which is up to an hour. In practice a disconnected assistant stops being asked to do anything, so this is a narrow window, but it is real and we would rather say so than imply a cleaner story.

If you need a connection dead immediately, contact us and we can end its sessions on our side.

What the assistant provider sees

Whatever you ask about goes to whoever runs the assistant: Anthropic for Claude, OpenAI for ChatGPT. That includes client names, contact details and payment status when those are part of what you asked.

Arvald does not send anything to them on its own, and they are not Arvald sub-processors. The connection is yours, and what happens to the data once it reaches the assistant is covered by their terms rather than ours. Our side of it is in the privacy policy.

A short checklist

• Connect to your sandbox first and get a feel for it.
• Ask what mode it is in before anything irreversible.
• Read refund and cancellation proposals before approving them.
• Disconnect when you are done experimenting.