AI agents with human approval
An AI agent with human approval does the reading, searching, and drafting on its own and stops before anything that would be hard to undo: sending, posting, paying, or changing a record. On Krafty Bot the server enforces that gate. The model doesn't get to decide.
Why does an agent need human approval?
Because the costly mistakes are the committing ones. A wrong draft costs a minute. A wrong email to a customer, a wrong bid change, or a deleted row costs more than the task saved. Pausing before writes gives you a chance to catch mistakes, although an approval gate cannot remove every risk. It also lets you widen trust one action at a time instead of deciding up front whether to trust the whole thing.
What needs your approval on Krafty Bot?
- Any write to a connected app: sending an email, posting a message, creating or updating a record, saving a file, adding a row. The Bot's turn stops, you see the request in the chat, on the Activity page, and as a push notification on your phone, and the stored action runs once when you approve it.
- Committing steps on the cloud computer: a click, drag, Enter, or key that would submit, send, or change something, as judged by a separate model from a screenshot the server takes.
- Risky shell commands on the computer. Obvious ones are refused outright; others are held and run once, exactly as you saw them.
What doesn't need approval?
- Reading your connected apps: your inbox, sheets, calendar, CRM, tickets.
- Searching the web and reading pages, within your plan's allowance.
- Opening sites and navigating in the browser on the cloud computer, taking screenshots, and reading what is on screen.
- Drafting text in the chat needs no approval. Saving a draft to Gmail is a connected-app write and needs approval unless an eligible rule allows it; sending always needs approval.
- Saving notes to the Bot's own memory.
How do always-allow rules work, and what are their limits?
When you approve a request, you can choose to always allow that action. The rule is bound to that Bot, that action, and that connected account, and you can list and remove rules at any time. Three things keep it narrow.
- Sending, posting, replying, deleting, payments, and access changes are high-risk. They ask every time and cannot be always allowed.
- A rule covers one action on one account for one Bot. A second Bot, or the same Bot on a second account, asks again.
- The approved action runs from the stored call the Bot made, never from arguments the model supplies later.
How does approval work on the cloud computer?
Every click, drag, Enter, and non-navigation key is sent to a separate vision model along with a screenshot the server takes, and the model decides whether the step is routine or would commit something. Routine steps run. Committing steps run only inside an approval you have given for that Bot and that chat, which is time-boxed. Navigation keys are exempt from the screen check. The check fails closed: if the model cannot decide, the step does not run. Account-security changes and commands typed into a desktop terminal are refused. Eligible shell-tool commands require separate approval when held.
What approval can and can't guarantee
Enforced by the server: app writes run only from the stored call; high-risk actions never get a standing rule; rules are bound to Bot, account, and action, and revocable; the computer guard fails closed; a decline blocks every new request for the rest of that turn, and an app call declined unchanged is refused for a day; nothing in a web page, document, or tool result can grant an approval. Reduced, not removed: the computer guard is a model reading pixels and can be wrong or misled by an overlay or a misleading label; typing without Enter can still commit on some sites; and the shell on the computer is a convenience, not a security boundary. We would rather you know that than find out.
Where do you approve?
In the chat, where the request appears as a card with Approve and Decline, plus Always allow only for eligible actions. On the Activity page, which lists every pending approval, routine, and recent unattended event. And on your phone, where a push notification can bring you back to the request. Always allow never creates a rule for high-risk actions or computer steps. Requests expire after seven days. Deciding wakes the Bot, which continues the job with your answer.
FAQs
Not globally. You can always-allow specific actions for specific Bots on specific connected accounts, except high-risk actions. That is the whole mechanism, and it is deliberate.
The request stays open for seven days. The turn that asked has ended; when you decide, a message wakes the Bot and it continues with your answer. If you decline, the Bot stops asking for the rest of that turn.
No. Approvals are recorded only by your decision in the app. Text in a page, a document, or a tool result cannot grant one, and the screen check on the computer is a separate model from the one doing the work.
Yes. A routine posts a prompt into the Bot's chat and the turn runs with the same gates as any other. A task one Bot hands to another cannot gain a permission the original turn did not have.
Try an AI teammate
Start free with one Bot and one connected app. Pro is $20 a month for up to 50 Bots and 50 hours of computer time. App writes wait for approval unless an eligible rule allows them.