37,442 Voice Requests: What People Actually Ask a Mac Assistant

Every company building an AI assistant, us included, sells roughly the same picture: it reads your inbox, manages your calendar, and hands you back your morning. We have been shipping that story for a year. So we went and looked at what people using our assistant actually ask it for.

Email, calendar and messaging are 6.2% of requests.

What is in the dataset#

Every voice request Jarvis handles is tagged at the moment it is routed, with a coarse category (use_case) and, where one is invoked, the tool that served it. The tag is assigned by the router, not by a human reading anything, and no transcript, no message content and no filename is recorded — only the category. That is why we can publish this at all.

The window is fixed and reproducible: 2026-05-05 to 2026-08-02, 90 days, 37,442 requests from 591 people. Of those, 2,006 come from the Android app, which does not yet send the category field, so the shares below are computed over the 35,436 categorized requests.

What people ask for#

Two things stand out. The first is the size of search — nearly a fifth of everything, and touched by 372 of 591 people. A voice assistant on a desktop is, to a large extent, a faster way to ask a question without opening a tab.

The second is that control beats communication by a factor of three and a half. App control plus Mac control is 22.5% of requests; email plus calendar plus messaging is 6.2%. People are not, mostly, dictating replies. They are opening things, moving things, changing settings, and finding files.

Opening the biggest bucket#

A category holding 38% of the data and named "other" is the first thing a careful reader should distrust, so here it is, broken out by the tool that served the request.

The top two entries are not tasks. They are the assistant explaining itself.

348 of 591 people — 59% — asked the assistant what it can do. 329 asked what it is connected to. Those two questions, together, are the most common thing anyone says to this product. More people did that than sent an email, checked a calendar, or played music through it.

That is worth sitting with if you are building one of these. The hardest problem in a general-purpose assistant is not capability. It is that a text box with a microphone gives the user no idea what is behind it, so their first instinct — and their instinct again weeks later — is to ask. Discovery is not an onboarding step you get past. It is the most used feature in the product.

What changes as people stay#

Splitting each person's requests into their first seven days versus everything after:

The shape of a settled user is clear enough: they ask it things, and they tell it things to remember. The novelty uses — driving the Mac around by voice, poking at what it can do — fade. And communication, the thing the entire category is marketed on, does not grow into the habit. It shrinks slightly.

The distribution, because averages lie here#

37,442 requests across 591 people averages 63 each, which describes almost nobody. The median person made 28.5 requests in 90 days. The 90th percentile made 158, the 99th made 377, and the single heaviest user made 1,287.

So a normal user of a voice assistant talks to it a couple of times a week, and a small group lives in it. Any "average user" claim about this product category — ours included — is being made about a distribution that does not have a middle.

What we think this means#

Build for discovery first. If 59% of your users ask what your assistant can do, the answer to that question is a core surface, not a help doc. We have started treating it that way.

Search is the wedge, not the inbox. The thing people reach for most is the fastest question-answering path on the machine. It is also the least defensible feature we have, which is uncomfortable and true.

Memory earns its keep. It grows with tenure, which is what you want to see from the feature that is supposed to make the assistant yours rather than generic.

Be careful reading this as the market. These are people who chose to install a voice-first assistant on a Mac and got far enough to use it. It is a real sample of real sessions, which is more than most of the claims in this category rest on, and it is still one product's users.

If you want to poke at the thing that produced this data, Jarvis is free and open source — Apple Silicon and Intel, no signup. The dictation and routing that generate these categories run on your machine, which is also why we can only publish categories: the content never leaves it.