The first reaction: "wait, it got that right"
The most common first-day message we get is some variation of surprise at accuracy. People arrive expecting a rough experience — they've tried cloud dictation before and remember correcting every other word. AirTypes, running Whisper on-device, is considerably better for most people in most situations.
This shows up in feedback like:
"I've tried dictation tools before and always given up in the first hour. This is the first one I'm still using on day seven."
"It got my name right on the first try. No other dictation tool has ever done that without training."
Whisper (the underlying model) handles accents, technical vocabulary, and natural speech patterns significantly better than older speech recognition engines. The accuracy tiers amplify this — the larger models are remarkably robust. Users who download the larger accuracy tier often describe the jump as "feels like it's actually listening."
The filler-word removal also catches people off guard. Speaking naturally and then seeing clean text — without the "um"s and "uh"s — is a small thing that turns out to matter a lot to how confident people feel about the output.
"I didn't realise how much I disliked cloud dictation until I tried offline"
This is one of the most consistent themes in long-term feedback. People say they knew, intellectually, that their audio was being sent to servers. They'd accepted that as the cost of using the feature. They didn't realise how much low-level discomfort that was generating until it was gone.
"I work with confidential client data. I always avoided dictating anything sensitive. Now I dictate everything, including things I never would have before. That's a real change in how I work."
"I live in Germany. The idea of audio going to American servers always made me uncomfortable. With AirTypes I just stopped thinking about it."
This matters in specific professional contexts — medical, legal, financial, government — where there may be formal restrictions on what data can leave a device. But it also matters more broadly to people who simply value privacy and don't want their conversations indexed somewhere.
The "works offline" aspect has a secondary effect that users mention too: it works everywhere. On a plane. In a hotel with throttled WiFi. In a corporate environment with strict firewall rules. The absence of a network dependency is a reliability feature, not just a privacy one.
The workflow change nobody expected
People install AirTypes thinking they'll use it for long documents — emails, reports, blog posts. That often happens. But the most mentioned workflow change in 30-day feedback is something smaller: short-form communication.
Slack messages. Reply emails. GitHub comments. Jira tickets. Linear updates. The stuff that takes 30 seconds to type but that you write a hundred times a day.
"I thought I'd use it for long emails. Turns out I mostly use it for Slack replies. Holding the hotkey and speaking a message is so much faster than typing when I'm in the middle of something else."
"The thing I use it for most is replying to comments in PRs. I used to give terse replies because typing was annoying. Now I give proper explanations. My colleagues have noticed."
This "friction reduction on short text" effect is one of the reasons retention is strong after the first week. It's not one big workflow that changed — it's dozens of small ones, adding up across the day.
My Agent feedback: the most polarising feature
My Agent — the feature that lets you speak to your own AI and have the response typed at your cursor — gets the most extreme reactions of anything in the product. People either become heavy users immediately, or they try it once and stick to plain dictation.
Heavy My Agent users describe it as genuinely different from using ChatGPT or Claude in a browser:
"The reason I use My Agent is that the response comes to me. I don't have to go find it. I'm already in the document I'm working on and the AI just types into it. That sounds like a small thing but it changes everything about how I interact with it."
"I set up three profiles — one for code documentation, one for formal email, one for casual writing. I use all three every day. It's the closest thing to having an AI that's actually integrated into my work rather than sitting in a tab."
Users who don't engage with My Agent usually say one of two things: they already have a workflow they like for AI (browser, API, IDE plugin), or they haven't had a chance to set it up properly. The setup friction — bringing your own API key, creating a profile prompt — is real, and it's feedback we're working on. Once people get past the setup, the drop-off in My Agent usage is close to zero.
What causes friction in the first week
We read the friction feedback carefully because it's where we learn the most. A few things come up repeatedly:
- macOS Gatekeeper on first install. The "can't be opened because Apple cannot check it" dialog catches people off guard. It's a one-time thing and the fix is two clicks, but it creates uncertainty. We've added a more prominent note to the download page, and our macOS setup guide covers it step by step.
- Choosing the right accuracy tier. Downloading the largest model on a machine that can't process it quickly leads to slow results and disappointment. The in-app tier guidance helps, but it's not prominent enough yet. On Apple Silicon M2+ with 16 GB RAM, the largest tier is great. On Intel or M1 with 8 GB, start smaller.
- Talking to a computer for the first time. This is the honest one. A surprising number of people feel self-conscious about speaking to their computer, especially in shared spaces. There's no fix for this — it goes away with practice. Users who push through the first two or three days almost always say the self-consciousness disappears entirely.
- Hotkey conflicts. The default hotkey works for most people but occasionally conflicts with another app. Settings → Hotkey, takes 10 seconds to change.
The things people say they love most
In roughly the order they come up:
- It works in every app. "System-wide" sounds abstract until you're dictating into your terminal, your email client, your design tool, and your notes app without ever switching context. This is mentioned constantly.
- The hotkey model. Hold to speak, release to inject. No mode, no modal, no separate window. People who've used dedicated dictation apps — Dragon, Apple Dictation, Whisper Web — describe this as a fundamentally better interaction model.
- Audio never leaves the device. Not a marketing claim. Provably true — the app has no network activity during transcription. Users mention this when recommending AirTypes to colleagues: "I can prove to you it's not sending anything."
- Speed. "3× faster than typing" is the number we cite, and users say it feels right. For some people it's more. For people dictating in a second language through My Agent cleanup, the effective output rate is even higher because the cleanup happens in the background.
- The stats. The AirTypes dashboard shows time saved, word count, most-used features. Multiple users mention checking their stats and finding the numbers motivating. Seeing "47 minutes saved this week" is tangible in a way that feeling faster is not.
"I use it every day now"
The 90-day retention pattern in AirTypes is unusual for a utility app. Most productivity tools see a sharp drop after the first week as novelty wears off. AirTypes trends the other way — usage tends to increase between week 2 and week 8 as dictation becomes ingrained.
Long-term users describe a change in how they think about writing at a computer:
"When I'm on a computer without AirTypes installed now, I notice how slow typing feels. I didn't think of typing as slow before. Now I do."
"Six months in, voice is my default. Typing feels like a fallback mode for when I'm somewhere quiet isn't an option."
This is the shift we hoped AirTypes would create. Not a feature you use occasionally. A different default mode for interacting with computers via text.
What users tell their friends about AirTypes
Word of mouth is how most people find AirTypes. When we ask long-term users what they tell people when they recommend it, a few things come up consistently:
- "It's actually private." Not "it claims to be private." Provably, demonstrably, offline. No audio leaves your machine. That's a rare thing to be able to say about a dictation tool, and users find it easy to explain and convincing to hear.
- "Try it for one week on Slack messages." Not the big use cases — the small ones. If you can get someone to replace their Slack typing with voice for a week, they usually convert on their own.
- "The My Agent thing is different from using ChatGPT." It's hard to explain without showing someone. The combination of voice input and the response arriving at your cursor in whatever app you're in is genuinely different from anything else in the market right now.
- "Just try the trial. No credit card." This one comes up a lot. The 7-day full-feature trial with no payment required reduces the activation energy enough that most people who hear about it actually try it.