Giving software the ability to act for you is a trade: convenience for risk. We think the right default is caution, and that trust should be earned in small steps.
Here is how the AI Assistant is limited.
Everything starts switched off
Autonomy is off by default. An administrator switches it on per channel: internal actions, content drafts, prospecting, invoice drafts, outreach email and ticket replies. Until then, the assistant only proposes.
People approve first
Every proposal is a card that a person approves, edits or dismisses. A trust ramp relaxes approval only after people have approved enough drafts in a channel, and some actions stay manual for good.
Hard limits
- A global kill switch pauses all autonomous actions at once.
- A daily cap limits external sends (20 by default).
- A freshness rule stops the assistant acting on stale cards.
- Invoice drafts are never sent by the assistant, and marketing posts are only saved as drafts.
Everything is on the record
Each action is logged, and a governance report summarises what the assistant did. Actions can be undone within 60 minutes.
Why it matters for you
If you are deciding whether to let an assistant touch customer email or invoices, the question is not whether it can be trusted on day one. It is whether you can see what it did, stop it instantly and undo it. Those are the parts we built first.
Read what the assistant does on the AI Assistant page, or ask us how we would set the limits for your business.