Hark, a young AI start-up, has released Hark Pro on the web, iOS and Android. It is free with usage limits, and paid plans at $20 and $100 a month raise the limits. The interesting part is not that it can use websites for you. Other assistants already do that. It is that the assistant, not you, decides what needs doing.
The assistant starts the conversation
When ChatGPT Pulse arrived in September 2025, it began preparing a daily set of research cards from your chats and calendar. Those cards were things to read. Hark's home screen is a list of things to do, and each card ends in a button that pays the bill, makes the appointment or approves the expense report. Meta's Muse, which came to the Mac a few weeks earlier, also works with your files, calendar and mail, but it mostly waits for a request.
Behind a button like that is a chain. The assistant reads an email and notices a date. It decides the date means a chore. It opens the right website in a browser of its own, fills in the form, and stops where only you can choose, such as a time slot or a payment. Then it shows you the result. The oil change worked this way, and so can a rental car that has been booked at the wrong airport: the assistant spots the mismatch in your confirmation email and offers to fix it.
Noticing is also what makes this expensive, because the assistant has to keep working when nobody has asked it anything. That is why Hark built its own model made only for using websites, and prices it at under a tenth of what a general frontier model costs per token. A free tier is only possible when each small step is cheap.
What the work becomes
A good human assistant has always been valuable for the same reason: they remember. They know the registration is due and the supplier invoice is late, and they come to you with the form ready. Until now only executives had that. A shop owner who deals with 20 such chores a month, at about ten minutes each, spends over three hours on them. At 15 seconds a chore it is five minutes. Even if you read each card for a full minute before tapping, it is 20 minutes instead of three hours.
So the work moves from doing to approving, which is what managers already do with staff. The mistakes also move. An assistant will rarely forget a renewal, but it can book the wrong branch, pay the wrong invoice or misread an email. Reading the amount and the recipient on every card, and keeping payments on a manual tap, is the habit that makes the saved hours safe.
Why the privacy promise is the product
Noticing a chore means reading everything. Hark pitches itself as a company that is not in the business of selling ads or building profiles of your friends, and its design lead describes the work of understanding you as limited to the tasks it helps with. That promise matters more here than for a chatbot you only type questions into, because the assistant sees your whole inbox, calendar and cards.
An inbox is also full of text written by strangers. Attackers have shown they can hijack AI browsers without any click from the user, and an assistant that reads your mail and then acts on a website is exposed to the same kind of trick. Hark shows its browser working in a small window inside the chat, which lets you see where it goes, but the stronger protection is your own approval.
Before connecting a work account, read what the company does with your mail. Start with a personal one, let it handle low-stakes chores for a month, and see how often the cards are right. If it is nearly always, the remaining job is reading each card for ten seconds and tapping.