What an AI Secretary Should Do — and the Three Things It Should Never Do

A good human secretary is not someone who does what you say. It's someone who knows what you'd want done, does the parts that don't need you, and brings you only the decisions that do. The value is in the filtering, not the typing.

Most "AI assistants" get this backwards. They're very good at typing — drafting emails, summarising documents, answering questions — and very bad at filtering. They wait to be asked, they ask for confirmation on everything, and they treat a request to book a table with the same gravity as a request to wire money.

When we designed Lazy OS, the assistant layer inside Sanora One, we spent more time on what it should *not* do than on what it should. This is what we arrived at.

What it should do

Handle the repetitive, low-stakes, reversible things without asking. Rescheduling a reminder when you've clearly missed it. Filing a receipt into the right folder. Replying "received, will look tomorrow" to a message that plainly expects an acknowledgement. Things where the worst case is a minor correction later.

Batch interruptions. Nothing kills concentration like an assistant that pings every time something happens. A good one collects the non-urgent items and presents them at moments you've said are fine — start of day, after lunch, end of day — as one list, not fifteen notifications.

Prepare decisions, not just present them. "You have three messages asking for a meeting next week" is a notification. "Three people want to meet next week; Tuesday 3pm and Thursday 11am work for all of them, and here's a draft reply offering both" is preparation. The second saves you the ten minutes of cross-referencing that the first leaves you to do.

Remember context across tasks. If you told it last month that you don't take calls before 10am, it shouldn't propose a 9am call this month. If you always decline a certain type of request, it should start drafting the decline for you and asking whether to send.

Explain what it did. Every action it takes on your behalf should be visible in one place, in plain language, with an undo where possible. "I moved your 4pm reminder to tomorrow because you marked today's task complete at 5pm" — not a silent change you discover later.

Get out of the way. The measure of a good assistant is how rarely you think about it. If you're spending time managing the assistant, it's a bad assistant.

The three things it should never do

1. Move money without an explicit, per-transaction yes

No matter how confident it is, no matter how routine the payment, an assistant should never initiate a transfer, a purchase, or a subscription on its own. The downside is too asymmetric — a mistaken payment can take weeks to unwind, and a compromised assistant with payment authority is a disaster.

Preparing the payment — filling in the details, checking the amount against the invoice, flagging that it's 20% higher than last month — is exactly what it should do. Pressing the button is yours.

2. Send anything irreversible to another person without showing you first

Messages, emails, posts, replies. Once sent, they're out. An assistant can draft, suggest, even queue — but the moment of sending should be yours, at least until you've explicitly said "for this category of message, just send it".

The exception, and it's narrow: pure acknowledgements with no content, where you've set the rule in advance. "Got it, will reply properly tomorrow" to an incoming message is fine to automate. Anything with substance is not.

3. Pretend to be you

An assistant that signs messages in your name, without disclosure, is lying on your behalf. When Lazy OS sends an acknowledgement, it says it's an automated response. When it drafts something you then send, that's yours — you read it, you chose it. But the assistant never impersonates you to a third party.

This matters more than it seems. The moment people learn an assistant was speaking for you without saying so, every message they've ever received from you becomes suspect.

The hard part: knowing which category a task is in

Everything above depends on the assistant correctly classifying what it's being asked to do. "Reschedule my dentist appointment" is low-stakes and reversible. "Reschedule my visa interview" looks identical and is neither.

We handle this with a combination of rules and asking. Categories you've explicitly delegated run without confirmation. Categories you've explicitly reserved always ask. Anything in between, the assistant does the preparation and asks once, and remembers your answer as a preference for next time.

Over a few weeks, the asking tapers off, because the assistant has learned where your lines are. That learning period is the cost of an assistant that's actually trustworthy, and we think it's worth paying.

What this looks like day to day

Morning: one summary. What's due, what came in overnight, what it handled, what needs you. Five items, not fifty.

During the day: silence, unless something is genuinely urgent by a definition you set.

Evening: what got done, what didn't, what's proposed for tomorrow. One tap to accept, or edit.

That's it. It's not exciting. Good secretaries aren't exciting. They're the reason you can concentrate.