Forget Bots: The AI Agents Are Coming!
Forget Bots: The AI Agents Are Coming!
You’ve heard of personal shoppers. But a personal AI agent? Here’s what happens to the retail industry when AI has all the purchasing power.
Words by Emily Wengert
Illustrations by Lisa Sheehan
“It’s time to reimagine the way we interact with the internet.” That’s what HyperWrite CEO Matt Shumer wrote this past April on X, formerly known as Twitter, with a video showing the AI assistant his company had created ordering a pizza on his behalf. It navigated the web to Domino’s website and filled out the form fields, all based on one simple text instruction.
It’s a “Look, Ma! No hands!” kind of moment, almost literally. After a human types in the initial instruction, the AI assistant does all the rest. As the name might suggest, HyperWrite was originally intended to be an AI writing assistant. But as its Chrome extension took shape, Shumer’s ambitions expanded with it.
“We’ve always had this vision that it extends, that it’s this assistant that can do anything,” says Jason Kuperberg, who co-founded HyperWrite with Shumer. “Executives have an assistant who’s doing these things for them: ordering things, organizing travel, scheduling meetings, notes, research. What would it look like for literally every single person to have something like this?”
Fans are starting to help answer that question. YouTuber AI Jason used HyperWrite to read and reply to all unopened personal email. He also got it to respond coherently to a LinkedIn post about AI on his behalf (both of which he then paused before he spammed everyone). While the quality of these generated actions was mixed (sometimes fine, sometimes poor), the demonstration of successful multistep task completion gives us a glimpse into where things are headed.
In a different video, YouTuber MattVidPro AI asked the HyperWrite assistant to find the two cheapest humidifiers on Amazon as well as the cheapest 4080 GPU on eBay. Though clumsy and slower than most any human, it effectively completed the task.
“We know that it’s early. We’re pretty transparent: This is the worst it’ll ever be,” Kuperberg acknowledges. “But the first time you see it working and actually do something is crazy. It’s like, whoa! This feels different from everything else that’s out there.”
At a time when the most magical often transforms quickly into the mundane, this truly felt — and continues to feel — like an enormous leap forward. It is also likely to deeply impact the retail sector in particular, in part because customers — especially Gen Z and millennials — are already signaling interest in personal virtual shoppers and AI shopping assistants.
Call My Agent
There are many features frequently touted when discussing generative AI. Most obvious: It creates things (words, code, images, animations, voices, worlds, songs). Also, it can synthesize lots of content into summaries we humans can more quickly process, including spotting both trends and anomalies in data. And it uses natural language to communicate, which adds to the dazzle of it all. But there’s one emerging capability that remains under the radar. And that’s generative AI’s ability to plan and execute tasks, sometimes even autonomously, just as HyperWrite’s Chrome plug-in proves. There’s a new word for this task-completing function: It’s called an “agent.”
An agent might still be controlled through a text field, but it goes beyond the familiar bot capability offered by a product like ChatGPT or, going back even further, classic chatbots that are programmed to guide you down preset paths. Here’s the big difference: Agents actually autonomously complete tasks for you, interfacing with sites or APIs along the way.
“This is definitely a sea change thing,” says Andy Mauro, who more than a decade ago led the team at Nuance that created Nina, the first mobile speech assistant for the enterprise (like Siri but for businesses). He also co-founded, built up and sold Automat.ai, which incorporated conversational commerce into brand websites.
Mauro recently founded the new startup Storycraft, a novel video game in which the player mentors generative AI characters, who have their own agendas and then go off on adventures.
“You’re going to see a ton of new companies, and I think those largely will be something along the lines of an agent-style company, which can take a vertical slice of a useful problem that people have and build an AI that you can talk to and guide and direct that’s aligned with you and then send it on an autonomous path to do something. In my case, it’s a video game.”
The danger, he cautions, is creating things that people don’t actually want. It should be born from existing needs we can already see, like managing a calendar or shopping for the perfect gift.
The Puppy Project
“Max” is my perfect dog. That’s according to AgentGPT, another autonomous agent, which I tasked with finding a dog I could adopt. It had to meet a fairly complex set of criteria: under 50 pounds, less than five years old, good in an apartment, not a barker. I also added something you can’t filter for: a couch potato. After working for less than two minutes, AgentGPT reported back: “One option that meets all the criteria is a dog named Max, who is available for adoption at the animal shelter in New York City.” It went on to describe Max’s calm nature, that he’s three years old and 45 pounds, perfectly suited for apartments and not a frequent barker. I was sold.
Unfortunately, when I tried to follow the dog’s profile page, the link led to the Petfinder.com homepage. Despite that mistake, I was convinced the agent had found my dream pooch. After it searched for two minutes, I proceeded to spend another 20 trying to track down this Max. Just one problem: Max, his details and his location were all made up — there was no real profile.
Eventually I asked my AgentGPT itself why the link didn’t work. It explained it wasn’t connected to any browser and was limited by information up to November 2021. Since the tool was built using GPT from OpenAI, it’s not altogether surprising it calls out that date, which is when the OpenAI model was trained. Weirdly, though, the FAQs for AgentGPT insist it does browse the internet — a good reminder to me that these agent tools are still in beta (or even alpha) and suffer from the same hallucinations that GPT 3.5 and 4 are known for.
In just three months after launch, AgentGPT, which has $1.25 million pre-seed funding with support from both Y Combinator and Panache Ventures, captured 150,000 monthly active users and 25,000 stars on GitHub. But the founders shared in their press release that the likelihood of hallucinations and failure increases the longer you run your request. These are two of the biggest problems they’re still working to solve.
For Jennifer Fleiss, co-founder of Rent the Runway and former CEO of Jetblack (a Walmart incubator service that used AI to offer personal shopping over text message), these emerging capabilities are feeding long-held ambitions.
“The world of AI has accelerated and made it much more cost-effective and faster to accommodate new recommendation flows. It’s exciting to see that this is going to be an unlock for so many businesses.”
Jennifer Fleiss, co-founder of Rent the Runway
“At Rent the Runway, for example, we’re plugging in different versions of AI to find the right dress for XYZ occasion and to let the customer layer in that it’s a beach night but it’s gonna be cool. All the variables,” Fleiss says. “And so I think various companies will use it to accelerate finding that best product match for their consumer.”
In other words, the tools of the future might know the consumer better than most brands today are able to, by leveraging truly 1:1 personalization.
“What you’ll start seeing more of is ways of having these plug-ins scrape and understand your personalization preferences,” Fleiss points out, adding examples like allergies in the family or specific colors you gravitate toward or your price sensitivity. “So I think the personalization opportunities here are just so much richer in a way that will be even more efficient for the consumer.”
She even imagines the possibility that individual data could be masked behind an agent to anonymize it but still remain accessible by a brand to offer the best products and experiences.
With the easy availability of these emerging tools, consumers may prove to be the early adopters ahead of businesses and brands.
“This is happening so fast, and enterprise generative AI has not been figured out yet. Customers are not bound by the same concerns that brands are,” says Katherine Jones-Siemsen, head of product and commerce at Huge. “I think about extensions and apps that used to scour the web for promo codes and things like that. We might just be seeing shoppers hack shopping in that way at a pretty sizable scale given the tools at their disposal and given the low level of effort and skill required to make use of them.”
She gives as an example the policy many companies have of price matching after something is sold if the consumer can find it cheaper elsewhere.
“That task management can be outsourced now. And it’s not particularly hard to do,” adds Jones-Siemsen. “The adoption and the ecosystem around these generative technologies is only going to get faster and easier.”
Currently, price-match policies are reliant on the assumption that humans won’t put in the labor to take advantage of them. However, it is possible right now for someone with a modicum of digital savvy to ask an autonomous agent to scrape their inbox for sales receipts, read the items and prices therein and then scan the web to beat that price. Moreover, the agent could generate the written request for price matching directly to the brand and then track that action in a spreadsheet, including writing follow-ups as needed.
Kuperberg at HyperWrite offers another example: “Something that we’ve heard is credit card rewards and airline miles and how that is a system that can be optimized and that people are already doing that.”
While these numbers don’t signal that there’s a demand yet or expectation for this capability, history shows us customers can be quick to adopt something when it benefits their wallet or saves them time. Klarna seems to be betting on this with an AI-powered shopping recommendation engine.
How Should Brands React?
One of the critical questions AI of all kinds raises is how a brand exposes or protects its data. Brands want that data to reach their customers, but what if the only way for that to happen is a nonhuman browsing the website?
“For years brands have had ‘not a human’ navigating their site,” Jones-Siemsen points out. “There’s price-matching algorithms, there’s content and data scraping, there’s crawlers trying to index your site. Those are very beneficial.”
The change in this case, however, is that humans may never browse your retail site. They may abdicate a simple decision completely to an agent. Or they might have the agent serve them answers and recommendations in an interface of their choosing.
Kuperberg predicts that like SEO optimization, brands will want to do what he calls “agent optimization” to ensure that an AI agent — either engaging with your site or your APIs — is able to get the information it needs so it can advise the consumer.
It’s possible to imagine, in that context, that experience benefits like ease of human navigation may degrade in importance. If an agent can find the best thing from anywhere, then it doesn’t matter how hard the checkout was if the agent’s the only one suffering through it.
This could change the nature of brand loyalty by putting the core emphasis squarely back onto the quality of the product — not the experience of shopping for it.
“Theoretically this allows brands to go back to having great products, having great post-purchase experiences and focusing a little less on operating transaction websites,” Jones-Siemsen adds.
In fact, autonomous AI assistance may actually help people achieve a goal they often claim they want but don’t always act on: values-based shopping, like factoring in sustainability or the inclusiveness of a brand. If someone’s personal agent is trained to apply those values to all suggested purchases, brands won’t be able to reach shoppers without authentically embracing those values.
“If you want to have brand control, you have to quite literally pick a side of what space you want to be in,” says Mauro. “When you think about what authenticity is, this weirdo word brands have been using, it’s that. You’re taking a stance. And the only way you can do that is if you actually have a stated brand purpose and experience you want to deliver.”
Plotting the Future
Jones-Siemsen predicts that retail as we know it today will be completely rewritten within eight to 10 years thanks to these task-achieving autonomous technologies.
And as is already true with e-commerce today, some brands will adjust for this next wave faster than others, creating value for their business.
“The success in e-commerce, I don’t think anyone would call it even,” Jones-Siemsen explains. “There have been brands that saw where things were going a bit earlier and were able to build really strong, resilient advantages. They were able to figure out e-commerce, figure out how to invest, how to find profitability and how to find adjacent streams of revenue in doing so. I think this age of AI agents will be a landscape that is uneven, possibly to an even stronger degree.”
What might that distant future look like? Do brands create their own agents, hyper-trained on their own data, to interface with customer agents instead of having them use the website?
“Agents need data, but they don’t need interfaces,” Jones-Siemsen notes. Perhaps in the future, the brands will go to the agents instead of waiting for the agents to come to them.
Someday, hopefully in the near future, an AI agent will finally find my “Max,” that ideal dog I dream of. In the meantime, I’ll have to resume the hunt the old-fashioned way: clicking links myself.
Emily Wengert is the head of experience innovation at Huge.