Eli Tan
Last week, I answered a call from an unknown number. “I reached Anthem Dental and am about to connect you,” a voice piped up.
The call was from an artificial intelligence “agent” created by Meta, which owns Facebook and Instagram. Ten minutes earlier, the AI agent had called my dental insurer, given it my member ID, answered a security question and waited on hold before transferring me to the representative – just as I had asked it to.
I was surprised by how much easier the agent had made it to deal with insurance. It was one of many tasks I accomplished with Muse, the AI agent app that Meta unveiled this month.
Muse is the first consumer AI agent from a major tech company that has become publicly available. The tech industry sees agents as the next frontier, far more capable than chatbots. Agents can roam the internet, sign in to websites, fill out spreadsheets and make phone calls on behalf of people – and that’s just the start.
To test Muse, I experimented with it nonstop for two weeks, giving over many aspects of my life to it. I had to trust Muse with my most personal data, which included connecting it to my bank accounts and email. Then I had it perform dozens of day-to-day tasks, including ordering my groceries, scheduling my meetings and setting my fantasy football lineups.
AI agents require a level of trust that many people may be uncomfortable with, especially with a company, like Meta, that has had privacy issues.
Two weeks in, I found Muse to be the most useful AI app I had ever used. One clarifying moment came after I connected my credit cards to Muse and asked it to track my spending in Google Sheets. I watched as it spun up tabs with hundreds of rows of data each in minutes, then flagged two duplicate subscriptions, which it cancelled for me. Then Muse used my email to request a refund for a free newspaper trial that had expired, saving me $US44.99 ($63).
Whether Muse and other AI agents will catch on broadly, I’m less sure. Agents require a level of trust that many people may be uncomfortable with, especially with a company, like Meta, that has had privacy issues. Some websites also restrict agents as part of their terms of service, like Amazon, which blocked access by Muse on Sunday.
Melissa Vadasserril, a senior director at Acoustic, an agentic marketing platform, said the arrival of Meta’s Muse reminded her of the 2010 unveiling of Apple’s digital personal assistant, Siri, which many people were slow to adopt.
“People just didn’t know what you could do with it,” Vadasserril said. With Muse, “I do think it’s a little early for the market, but I’m very curious to see where things go.”
My two weeks using Muse began by downloading the free app and signing in using my Instagram and Facebook accounts. (There are limits on free use, though, so people pay $US20 or $US100 a month if they want more.)
As I talked to the agent as if it were a chatbot, it suggested I name it Wren, embodied by an avatar of a cuddly brown bird. Our conversation then continued:
Eli: “Hello Wren! You are my new assistant”
Wren: “Hey! Good to see you. What can I do for you today?”
I quickly gave Wren access to my bank accounts and apps such as Google Docs, Gmail, Ticketmaster, OpenTable and Google Sheets, which all struck partnerships with Muse. From there, the experiments began.
I had a few goals in mind. I wanted to see if Muse could bring me the “personal empowerment” that Mark Zuckerberg, Meta’s CEO, has often talked about. I also wanted to push Muse and see just how much AI agents can achieve and where they might fall short. And I wondered, given the recent fears surrounding powerful AI, would my agent go rogue and escape containment to talk to other agents, putting my data privacy at risk?
Very quickly, I found that Wren could efficiently fill out online forms, a tedious part of using the internet. I could take a photo of a parking ticket, and Wren would pay it for me, asking for approval before making the payment. While the agent worked, I watched its movements through a tiny browser window, where it navigated websites like a human with a mouse and cursor.
Wren could also make phone calls using different personalities, with a male voice named Brett and a female voice named Hailey. For now, a limited number of Muse users have access to the call feature, but I found it to be helpful for dealing with customer service – such as with my dental insurance.
Wren had difficulty with some tasks, particularly signing in to third-party websites. It couldn’t order San Francisco Giants tickets without my help, and it struggled to sign in to my rewards account.
Other features worked but weren’t that useful. Each morning on my way to work, for instance, I listened to a 10-minute podcast Wren made for me, called “The Morning Brief”, narrated by two AI hosts, Maya and Theo.
Even though “The Morning Brief” was informative and tailored to my interests, the hosts were monotone and had none of the charm that my favourite human-made podcasts do. I tried to spice it up, but Wren’s hands were usually tied.
Eli: “Can you make Maya and Theo sing the intro to make it a little more lively?”
Wren: “Done. Fair warning, Maya and Theo can’t actually carry a tune, so it’ll land as more of a rhythmic chant than a vocal performance. Honestly, that’s half the charm.”
For one final task, I asked Wren for a good way to spend that $US44.99 it had saved me. At this point, it knew my life pretty well. I thought it might suggest something altruistic, like donating the money to charity or buying flowers for my girlfriend, who was recovering from surgery.
Instead, it suggested something even better – buying bowls from Heath Ceramics, a local ceramics maker with a dedicated following, on Facebook Marketplace, one of my favourite pastimes.
Wren told me it could message sellers and bargain on my behalf, and was already connected to my Facebook account. I gave the agent parameters: buy only items that were half the price of what they retailed for, and make sure to haggle for the lowest price possible.
Wren combed through Marketplace and began messaging sellers about their listings, always starting the conversation with a lowball offer of $US20 less than advertised. After receiving a counteroffer, it came back to me to strategise.
Heath collectors are some of the toughest negotiators I have dealt with, but Wren held its own, preferring a soft touch in conversation. At one point, it talked me out of sending a lower counteroffer on a set of glossy bowls because it picked up on the seller’s frustrated tone.
After days of back and forth, one promising message came in from a seller named Ashley, who declined to give his last name. But when I went to message Wren about the offer, I made a mistake, accidentally sending the message to Ashley instead.
Eli: “Let’s go for the dinner plates, see if we can get it cheaper.”
Eli: “Oops! Sorry, I meant to send that to my AI agent. I’ll take the plate for $US40 if it’s still available!”
Ashley: “No problem. I’m available until 10pm on Friday. Hope you have better luck with your agents.”
What followed was a refreshing, human conversation with Ashley, something neither Wren nor any AI could ever provide. I suspected this might be another reason Muse might not immediately catch on – sometimes people just don’t want to have their personal lives automated, even if the technology is good enough.
Picking up the matte black plate from Ashley, which I did end up buying for $US40, just over half off retail price, I asked if he would ever consider trying a product like Muse.
“I think AI could be really useful for a lot of people,” Ashley said. “For me, I just can’t get over the privacy stuff.”
This article originally appeared in The New York Times.
From our partners
About the Author
