I asked ChatGPT Agent to order a pizza to my house — here’s what happened
Just a few months ago, ChatGPT launched its Agent functionality, allowing the chatbot to act on your behalf. By doing this, you can ask for its help to plan and book holidays, create detailed presentations, and book a table for a restaurant.
I’ve tried using it for a host of different things now, but one thing I’m yet to try is using it to order food. With how far AI has come in the past couple of months, it should be a simple process, right?
Using the Agent
ChatGPT Agent can be activated by clicking the plus button on the search bar of ChatGPT. It will give you a list of suggested prompts. For me, it gave some intellectual choices, like “retrieve WHO health professional density data” or “make a reading list of the latest top Financial Times stories.”
I instead opted for the prompt “order a pizza to [my address inserted in]. Ask me for confirmation before making a payment”.
ChatGPT quickly jumped into action, looking up the nearest pizza places to me. It then decided that Dominos would be the best option for me, quickly working through the nearest locations to my address.
ChatGPT Agent is surprisingly good at this step, quickly picking the nearest location and working through the many questions that come with it.
I hadn’t specified what pizza I wanted originally. At first, the agent added three completely random ones before deleting two from the basket. It then seemingly got stuck in a purchase decision, randomly picking and removing different types of pizza.
Once it had finally decided on my seemingly randomised pizza, I then spent a good five minutes watching ChatGPT Agent try and access my basket. When it can’t complete a command, a small text box pops up to inform you what went wrong and how it is attempting to fix it.
This can sometimes feel like watching a small child learn how to interact with the world. It kept telling me that the basket link didn’t work and that it would try again. Eventually, it got there, accessing the basket and moving onto the next step.
Just before making it to the final step, the chatbot fell for a bit of last-minute marketing, throwing in some chicken wings to hit a ‘minimum order value’. I’m not sure what minimum value it meant as this isn’t normally an issue. It also didn’t seem to understand the + and – buttons on the extras at first, clicking the image of a cookie a few times before releasing how to do it.
Finally, we had made it to the final step. This is where the chatbot passed it back to me to enter my details and pay. In total, the whole process took about 15-20 minutes. While that is a long time, especially compared to how long it would take normally take to order food online, I was surprised at how well it did deal with some of the obstacles thrown its way.
Now for something more challenging
I then tried adding a step of finding the best-rated pizza place in my city and then ordering from there. At this step, I realized just how much Agent struggles with cookie messages or websites that aren’t designed well, getting stuck on buttons and menus.
Somehow, ChatGPT ended up landing on a kebab van around the corner from me as the best option. Then the agent changed its mind to Papa John’s. I then watched as it went round in circles for a long time, scrolling through random delivery sites and having a panic over star ratings.
Both times, it did eventually get to the end of the process and was ready to order a pizza, but it just took a long time.
A bigger challenge
Understandably, the agent kept picking out the big-name pizza places. It kept defaulting to companies like Domino’s or Papa John’s, where there are lots of them scattered around, and they are easy to order from.
So, I wanted to give it a bigger challenge. I asked it to order me a pizza from a specific restaurant. I know this place does orders to my house, but it will be a bit more buried on the websites like Uber Eats or Deliveroo, and having used the restaurant’s own website before, I know for a fact that it is a nightmare to navigate.
I watched as ChatGPT Agent relentlessly clicked through the pop-ups of the pizza restaurant’s website and painstakingly tried to find the order section before finally giving up and googling how to order from there.
It couldn’t find an answer on Google, so proceeded to Deliveroo, searching there for a long time, repeatedly entering the restaurant’s name and trying again, using filters. It finally gave up and tried Uber Eats, repeating the process a number of times, before finally tracking it down.
By the time ChatGPT had gotten to the stage of just finding the restaurant, it had been about 10-15 minutes of endless clicking. It must have been tired at this point because it then got stuck in a cycle of adding a pizza to my basket and then taking it out again.
You’ll be happy to hear that it did finally get me there, handing over the controls to finalize the order. It had been all over the web, tried every delivery site, and got lost in the process, but through blood, sweat, and tears, and way too much time, we got there together.
Overall thoughts
Agent is an impressive tool, but these technologies still have a long way to go. Most noticeably, they just struggle with obstacles that we’ve become so accustomed to as humans.
Pop-ups were consistently getting in ChatGPT’s way, and it was oddly distracted by requests by websites, at one point seemingly having a moral breakdown over a prompt from Dominos of it would like to donate to charity.
Right now, these kinds of tasks are clearly possible by an AI agent. Each time I gave it a variation on this task, it did eventually get there. But that is the keyword, eventually. It was definitely not quick.
Maybe one day we’ll be having AI do our food shops and pizza deliveries, but it still has a long way to go.
Follow Digitpatrox on Google News and add us as a preferred source to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button!
More from Digitpatrox
Back to Laptops
Source link