Five Things to Check Before Dictation or an AI Assistant Goes on Company Macs

More and more of the questions I get sound like this: our staff want to use dictation or an AI assistant, and we do not know whether to allow it.

I understand why they want it. Talking is faster than typing. A Stanford study on phones found speech 2.93 times faster than the keyboard in English, and 2.87 times faster in Mandarin with Pinyin input. Japanese was not measured, but typing kana and converting to kanji is a very similar chore to Pinyin.

At the same time, voices and screens are company information.

This month in Tokyo I built a Mac assistant called Yubi. Hold the right Option key and talk, and your words are typed where the cursor is. Hold the right Command key and talk, and she reads the window in front of you, points at the button you need, or presses it for you.

As the person building it, the first thing I decided was not a feature. It was what would never leave the Mac. Those rules work just as well as a checklist for a company deciding what to allow.

First: where does the voice become text

Dictation apps split into two kinds. Some send your audio to a server and get text back. Others turn it into text on the Mac itself. They look the same from the outside. Press, talk, let go. The difference is all behind the curtain.

With the server kind, the meeting you just summarized and the client name you just said pass through someone else's infrastructure the moment you say them.

Yubi transcribes on the Mac. Zero seconds of audio are uploaded. When you evaluate a tool, ask the vendor directly where the audio is processed. "We do not train on your data" in the terms is a different question.

Second: how does it see the screen

AI assistants that help with the screen look at it in one of two ways. They take a screenshot and send the image, or they read the accessibility information.

Accessibility information is the list of buttons and fields, with their names and positions, that screen readers like VoiceOver use. It is not an image, so the photos and documents on screen do not go along with it.

Yubi only uses the second. It takes no screenshots and never asks for Screen Recording.

Checking this in your company is easy. Open System Settings, go to Privacy and Security, and see whether the app is asking for Screen Recording. If it is, it can capture your screen as it is.

Third: can you see afterward what left

Screen help needs to talk to an AI model. That is exactly why the person using it should be able to see what was sent.

Yubi keeps a record of every send on the Mac and lists it inside the app. The record itself goes nowhere. The option that tidies dictated text is off by default, and when it is on it sends text only, never audio.

An app that cannot tell you what it sent is an app your IT team has no way to explain.

Fourth: does it stop before anything irreversible

If you can ask an AI to press a button, the next question is which buttons it must not press on its own.

In Yubi, any button near words like send, post, buy, pay, delete or book always waits for a yes, and so do their Japanese equivalents. That rule lives in the app's code, not in the model. However confident the model is, if the button says send, it stops.

And one yes allows exactly one action. Permission to send one email does not cover three.

Before rolling a tool out, ask the vendor what stops an accidental send or purchase. If the answer is that the AI decides, the honest translation is that nothing stops it.

Fifth: are there places it will never read

Yubi does not read password managers. Whatever you ask, it will not.

Whether a tool has a fixed list of places it will not read, and whether its makers can explain that list, tells you how the people behind it think.

My view

I think setting conditions works better than a ban. Ban these tools and people use them on personal devices, where you see even less.

The five conditions above are enough. The voice stays on the Mac. The screen is not sent as an image. What was sent can be checked afterward. Anything irreversible always stops for a yes. There are places it never reads.

Yubi runs on macOS 26 with Apple Silicon, and dictation is free and unlimited. You can play all seven use cases on the site before installing anything, or download Yubi for Mac directly.

If you want to talk through how to choose AI tools for your company, or how to build internal ones, get in touch.