An AI Executive Assistant Should Prepare Work, Not Quietly Approve It

I want an AI executive assistant to do a great deal of useful work. I want it to gather the context, compare the options, prepare the draft, and show me what needs a decision.

What I do not want is for preparation to quietly become permission.

These are different actions

Researching an option, drafting a message, and sending that message are three different actions. They carry different consequences. The same is true when changing a price, making a purchase, committing time on a calendar, editing a customer record, or publishing content.

A system may save a few clicks by treating all of that as one smooth workflow. But if something goes wrong, who decided what? Where did the authority come from? Which version was approved?

Preparation can be broad. Permission should be narrow.

I am comfortable allowing an assistant to gather context, compare alternatives, identify risks, draft an artifact, and recommend a next step.

Permission is different. It should name the exact artifact, the action, the destination or audience, and the scope. If any of those changes in a meaningful way, the decision should come back to the person who owns it.

Your yes belongs to a version

Suppose you approve a proposal. Later, the amount, the terms, or the recipient changes. Your original yes should not silently apply to the new proposal.

That does not prevent the assistant from improving the work. It simply prevents a changed decision from being smuggled through an old approval.

After the action, bring back a receipt

If the action matters, I want evidence: the URL that was published, the exact message that was sent, the calendar event that was created, the transaction that completed, or the record that changed.

Here is the thing: a confident “done” is not evidence. The owner needs enough information to verify the result and recover when the result is not what was intended.

Human control should not mean approving every harmless step

I do not want the solopreneur to become the bottleneck for everything. A useful assistant should be able to read, organize, calculate, draft, monitor, and verify within boundaries that are understandable and durable.

The founder’s attention belongs where judgment and consequence are highest—not on mechanically approving every low-risk action.

Five questions before you delegate

1. What could change in the world if this goes wrong?
2. Who would be affected?
3. Can the action be reversed?
4. What evidence would prove success?
5. Does the current approval cover this exact version and destination?

This is the standard I am bringing to Orabuena

I am designing Orabuena around bounded authority, exact approvals, and evidence-backed completion. The goal is an assistant that reduces the burden without pretending to own the company.

Orabuena is still in private research and product discovery. Tell me which decisions must stay yours and what preparation you would gladly hand off. Research participation does not subscribe you to product-update email.