Why an AI secretary should ask permission

An assistant that answers questions is convenient. One that acts in your email, servers or shop is something else: it can save you hours, but it can also get things wrong with real consequences.

The problem is not intelligence, it is control

When an AI moves from talking to doing, the question stops being "does it understand?" and becomes "what can it break, and how do I find out?"

The brakes we ask of it

• Risk levels. Looking something up is one thing; sending an email or touching production is another. Every action has a level, and the higher ones wait for your approval.
• Explicit approval. By voice, with a click or with a Telegram button. Without your yes, it does not happen.
• What is forbidden is blocked. Some actions it must never take, even if asked.
• Chained log. Every action is recorded in a way that is hard to alter unnoticed.
• Your credentials stay with you. The hands work in your environment; your access does not pass through us.

And the audio

Your voice is transcribed and discarded. We do not keep your conversations to train anything.

If you want to try it, reserve your seat: we are opening a first group of pilots.

Back to the blog