What area would you NEVER trust to Ai, in your inbox?
by•
I mainly follow the principle of "reversibility".
EG. An archived newsletter or a starred investor email can be always reversed.
If ai is right -> makes my life easier
If ai is wrong -> doesn't hurt me
Meanwhile, I would NEVER let ai auto-send things (especially to other humans) for me.
Because no matter how accurate its tonality is to me,
no matter how well it understands my context,
I just don't... Maybe there's some kinda humanity in pressing "send" yourself!?
What about you? Where would you draw the line? Or just raw-dog it and trust the ai?
137 views
Replies
For me the line is: seconds of cleanup vs. relationship damage. Archiving or starring = reversible. Sending, auto-CC’ing, or accepting the wrong meeting = not really reversible. Anthropic frames this well around human approval for higher-stakes tool use: https://docs.anthropic.com/claude/docs/tool-use Would you put calendar accepts closer to archiving or sending?
Dirac
@fabriziowexare Yeah, this is developing deeper on the "reversibility" I mentioned. So true.
Me personally? I would put calendar accepts into sending.
I believe in AI governance. No matter how powerful the AI or how intelligent, if the AI can not explain what it did vs what it was allowed to do and prove that it was allowed to do it....why on earth would I turn it loose in the real world?
Dirac
@thomasbcoleman Can you explain more on Ai governance? Seems interesting
@peterz_shu Hey Peter — you asked about AI governance, so here's the quick version, and you're going to recognize it. Governance = can I trust an AI to act on its own: see what it did, why, and undo it. You already built a governance rule into Dirac — "background tasks follow one rule: reversibility." That's exactly the line serious agent builders draw. You just didn't call it that.
What I work on is the layer up: continuity engineering. The premise in plain terms — an agent fails when it loses track of what it's doing faster than it gets smart. Doesn't matter how good the model is if it forgets what it already decided or what it's not allowed to do. For something like Dirac running in the background across threads (and more so once Sentry/Stripe/Linear are in), that "remember the rules and prove you followed them" layer is what keeps trust as it does more.
Easiest way to see it: I make a free, open-source extension, Prompt Accelerator (Chrome/Firefox), that keeps an inspectable record of a chat's real objective, decisions, and what was ruled out — so state doesn't drift. There's a heavier version (EchoGate) built for agents that take real actions: reversibility-plus, where every action is checked against fixed rules and logged so you can replay exactly what happened. That's more Dirac's world long-term.
Not pitching — you're already thinking about it the right way. Happy to send the link or just trade notes on agent-trust. Congrats on the launch.
Same line of thinking here. AI sorting, summarizing, drafting? Fine. But auto-sending, especially to a client or a stakeholder? Never. It's not even about accuracy, it's about accountability. When something goes out with my name on it, I want to have actually read it first, even if the AI nailed the tone 100%.
There's also something real about the act of pressing send yourself. Like a small moment of "yes, I'm choosing to say this." Removing that feels like it removes something that actually matters.
Reversibility is the right frame. I'd add: anything where the AI's "explanation" for what it did matters as much as the action. I've been deep in EU AI Act stuff lately and the line that stuck with me is explainability isn't optional once money or a person's record is involved, even if the action itself is reversible.
I would use reversibility as the first rule too. Drafting, summarizing, prioritizing, and reminding are all reasonable. Sending, deleting, changing commitments, or replying to sensitive customer/investor conversations should require human approval.
The highest-risk area for me is tone plus authority. An AI can write something that sounds confident but subtly changes the relationship or commitment. For inbox AI, I would want a clear explanation of why it drafted the reply, what context it used, and whether the action is reversible before it touches anything customer-facing.
id draw the line anywhere a bad action creates cleanup work with another person. drafting and ranking are fine, but sending, scheduling, billing, and access changes need explicit approval