ChatGPT Tasks and Operator turn the AI bot into a fascinating productivity tool

Every day for the last few weeks, I’ve received a notification on my phone at 7:30 in the morning. The notification comes from ChatGPT, and it always contains the same thing: instructions for a 20-minute full-body workout and a 10-minute meditation. The instructions are simple, and I’ve actually come to appreciate the daily prodding. I do wish it would stop recommending the exact same thing every damn day, though. The mountain climbers and positive intentions are getting a little old.

OpenAI has added a number of new features to ChatGPT in the last few weeks, a couple of which attempt to turn the chatbot into a straightforward productivity app. There’s Tasks, which all paid users can access and allows you to set reminders and make to-do lists in ChatGPT; and there’s Operator, a so-called “agentic” model for Pro subscribers that attempts to actually accomplish tasks on your behalf. As an incorrigible tester of to-do list apps, I decided to throw my life into ChatGPT and see if it could help me get more done.

After using it for a while, I’m sold on the idea of AI-capable task apps. The best way to use AI is simply as a way to get you started — it’s a brainstorming partner, a sounding board, a first-draft generator — which turns out to be exactly the thing I often need to get going on a project. Beyond that, though, ChatGPT is a terrible to-do list app. It’s smart, but it’s also a mess.

Even adding a task to ChatGPT is too complicated. You have to open ChatGPT, select the correct model — the name is “GPT-4o with scheduled tasks,” not to be confused with the half-dozen other models you can use — and then type or speak what you want. You can say one task or rattle off your whole list for the day, and see what happens. Sometimes it works, and sometimes it fails spectacularly. I’ve had a task on my list all week that just says “check phone,” and I have literally no idea what it’s supposed to be.

As long as it has notification privileges, ChatGPT will reliably ping you in an hour to switch the laundry, if that’s what you asked for. You can set recurring tasks, too; in addition to my workout and meditation, ChatGPT has been reminding me every third evening to make overnight oats and to take out the trash every Wednesday. The reminders are just notifications on your phone (or emails in your inbox) with a link to a message in ChatGPT, and I do wish they were a bit richer. But they show up on time, and that’s something.

That’s about it for basic to-do list features. There are no project management tools or a way to easily organize your tasks. I can’t even find a way to reliably see my whole to-do list — asking ChatGPT for a list of all my tasks brings up one list and clicking on the Tasks menu inside the app brings up another. Some things I added as tasks appear to have gone to ChatGPT’s memory feature instead, which has absolutely no connection whatsoever to the Tasks feature except that when I say “what are my tasks for tomorrow,” it recites the answer from its memory and not from my actual task list. It makes no sense.

I pretty quickly whittled my ChatGPT productivity system down to tasks that require some kind of information to get started. ChatGPT doesn’t just remind me to make overnight oats every third night — it offers me a new recipe I haven’t tried yet. (But it always has to include blueberries, per my prompts.) When it reminds me to go to the grocery store, it also suggests a bunch of toddler-friendly things that Trader Joe’s sells. If I need to email someone, it’ll draft the email for me. The information in the messages isn’t always helpful or even accurate, but I’ve been surprised at how useful it is just to have something there as soon as the task appears in front of me.

The real victory, though, would be for ChatGPT to just do some of those tasks. That’s where Operator comes in: it’s an AI agent that attempts to actually go out on the web and get stuff done. Instead of just reminding me to make a restaurant reservation, book an Airbnb, or go to Trader Joe’s, Operator aims to actually do them.

Operator is, in my experience, both very impressive and kind of useless. I asked it to find an Airbnb within two hours of my house that accepts kids and dogs this April. Right there in the ChatGPT window, it opened up a virtual Chrome browser and went to airbnb.com. Then the bot asked which weekend in April, to which I replied that I didn’t care. Eighteen seconds of browsing later, it asked again: how many adults, children, and pets? Two, one, and one, I typed back. Operator successfully navigated Airbnb’s confusing interface of drop-downs, and found me… a bunch of places to stay right down the street from my house. In a way, this was a success! And I suppose it’s my fault for not being clearer in the prompt — a human would understand what I actually meant, but ChatGPT simply doesn’t.

I then asked Operator to find me some cool merch from Schitt’s Creek, a favorite TV show of mine. ChatGPT opened up Chrome, went to Bing (it always goes to Bing), did a search, and landed on Etsy. It scrolled around for a while, seemingly scanning for items with lots of reviews and a high star rating, and eventually offered me three options with links. This is legitimately useful! It’s not hard work, but it’s busywork I simply didn’t have to do. I picked one, Operator added it to my cart, and then it shared a link. I clicked the link, and my cart was empty. Of course it was: Operator isn’t on my computer, and it’s not in my account. If I were a more trusting person, I could have taken over the browsing session and given Operator my login info and credit card details, but instead I bailed. I took the three links and moved on.

In general, the best luck I had with Operator was with tasks that involved scouring star ratings or some other scannable and quantifiable metric. It could quickly find highly rated recipes for almost anything or well-liked movies just by poking through Bing results. Without that external validation, though, it often just gets lost. And as with everything in the AI world, there’s always the risk that information will be irrelevant or flat-out wrong. Since the stakes of my requests are rarely higher than “what should I cook for dinner tonight with broccoli and couscous,” though, it hasn’t been an issue.

Still, it’s both fun and kind of wild just to watch the system work. When I asked Operator to find me a pop-punk concert in Washington, D.C., sometime in the next two months, it kept finding bands, searching Bing to see if they were pop-punk, then starting over when they weren’t. It eventually landed on a website called Hypebot, staring at a page for a Hawthorne Heights concert in April. It was technically more than two months away, but it was hands-down the best answer yet. I wish it had told me! Instead, Operator just gave up.

Somewhere in all these half-baked, half-useful features is the beginning of something potentially really powerful. A to-do list app that actually understands what I need to get done is a to-do list app that can help me get it all done. All the generative power of these AI models for writing emails, making suggestions, finding information, drafting documents, and now even surfing the web, combined with my priorities and timing, starts to look more like an actual assistant than any of the so-called “virtual assistants” that are really just conversationalists.
But that is, at best, years away. At worst, it’s just never going to get there or even close. For now, ChatGPT’s productivity system is like everything else about ChatGPT and most other AI tools. It’s messy, hard to figure out, and fails often, but it’s still kind of remarkable that it works this well. It’s genuinely useful when all you need to get started is something. What Tasks and Operator give you isn’t much, but it’s often something.