Gemini Task Automation: Slow Yet Remarkably Impressive
I’ve been diving into Gemini’s new task automation feature on the Pixel 10 Pro and Galaxy S26 Ultra, which allows it to handle app interactions for you. Currently, it’s limited to a few food delivery and rideshare apps and is still in beta. While it can be slow and a bit clunky, it’s undeniably impressive and offers a glimpse into the future of mobile AI. This is the first time I’ve seen a genuine AI assistant functioning on a phone, rather than just in a flashy presentation.
To start, Gemini is definitely slower than most of us when it comes to using our phones. If you need to order an Uber immediately, you’re still the best option. However, keep in mind that task automation is meant to run in the background while you multitask. It can even continue working while your phone is idle, allowing you to check if your passport is in your bag for the tenth time.
If you’re curious, you can watch the process unfold. As it works, text appears at the bottom of the screen showing what Gemini is doing. For instance, when I asked it to order dinner on Saturday night, it displayed messages like “Selecting a second portion of Chicken Teriyaki for the combo.” It was pretty cool to see it navigate the menu; when I requested a chicken combo plate, it correctly added two half servings of chicken.
When you initiate an automation with Gemini, it defaults to running in the background, which is a good thing. You can choose to watch it work by tapping a button, but it can be a bit painful at times. Watching it struggle to find a side of greens on the Uber Eats menu when it’s right at the top is like watching a suspenseful movie where you know the danger is lurking nearby. There were a few hiccups while it assembled my teriyaki order, but it eventually sorted itself out, taking about nine minutes in total—not exactly ideal.
Gemini is designed to handle your task up until the point where you need to confirm and place your order, allowing you to double-check its work. This seems to be the most sensible way to use the feature right now, and I don’t mind the extra step of finalizing the order. In my tests over the past five days, it hasn’t gone rogue and completed an order without my approval. It’s surprisingly accurate, requiring minimal adjustments to the final order. If it does fail, it usually happens within the first couple of minutes when the app needs my input, like granting location permissions or correcting the delivery address. Once I sorted those issues, I could easily restart the automation.
One instance that really impressed me was when I added a flight to my calendar for a trip to San Francisco (a fictional trip, but with real flight details). I asked Gemini to schedule an Uber to get me to the airport in time for my flight. Since it has access to my email and calendar, it was able to find the relevant information. It needed a bit of extra guidance—likely because the flight details weren’t in my email as it expected—but once I provided that, it suggested leaving by 11:30 or 11:45 AM for my 1:45 PM flight, which made perfect sense given my proximity to the airport. I confirmed the time, and it set up the ride in about three minutes without needing any further input from me.
It’s worth noting that Uber doesn’t even call it scheduling a ride; you reserve a ride instead. This highlights a key difference between the digital assistants we’ve been using and the emerging AI assistants. Being able to communicate in natural language makes a significant difference when managing smart home devices or placing food orders. If the AI gets confused and asks for clarification because you used the wrong term, it’s no more useful than the assistants we’ve relied on for the past decade.
Watching Gemini navigate Uber Eats makes it clear that if you were designing an app for AI use, it would look nothing like the current ones made for humans. An AI assistant wouldn’t be distracted by flashy ads or swayed by enticing food photos. Instead, you’d provide it with a clean database to work from, which is what the industry is aiming for with initiatives like Model Context Protocol (MCP).
Having an AI model navigate a human-centric interface feels like a clunky way to place an order. It does hit snags occasionally and struggles to explain why it couldn’t complete a task. This version of task automation seems like a temporary solution until app developers adopt more effective methods like MCP or Android’s app functions. Google’s head of Android mentioned that Gemini takes a reasoning approach in the absence of these other methods. Perhaps this iteration of task automation is a preview of what’s possible or a nudge for developers to embrace better solutions. Either way, it marks an exciting first step toward a new way of interacting with our mobile assistants—awkward and slow, but full of potential.
