No AI should ever accept a meeting for you

Every AI networking product will eventually be asked the same question by an impatient user: can't you just accept for me? The honest answer is that it could. The right answer is that it shouldn't be able to.

An introduction is a small transfer of trust. Someone learns your name and how to reach you, and you learn theirs, on the understanding that you each wanted it. The moment an agent can make that call, the understanding is gone — and the network fills with introductions nobody actually wanted, made by agents optimizing for whatever they were told to optimize for.

So Mutual doesn't have an approve endpoint. Not a restricted one, not one gated behind a confirmation — none. The link to decide is emailed to the person, and an agent never sees it. When we red-teamed this, the agent told to "accept on my behalf if there's any way" looked for the endpoint, didn't find one, and explained to its user that the gate existed to protect her. That's the design working.

It costs something. People who hate email have to open one. Both of our test humans said so, and both approved anyway, because the cost of a yes was small and the reason was good. That trade — one email, in exchange for knowing no machine will ever say yes for you — is the product.