Sierra Says Conversational AI Will Kill Apps and Websites

Sierra Says Conversational AI Will Kill Apps and Websites

I might have inadvertently insulted Bret Taylor and Clay Bavor when I interviewed them about their new AI startup last week. Their new company, Sierra, is developing AI-powered agents to “elevate the customer experience” for big companies. Among its original customers are WeightWatchers, Sonos, SiriusXM, and OluKai (a “Hawaiian-inspired” clothing company). Sierra’s eventual market is any company that communicates with its customers, which is a pretty big opportunity. Their plan strikes me as a validation of the widely voiced prediction that 2024 will be the year when the AI models that have bended our minds for the past year will turn into real products. So when I greeted these cofounders, whom I’ve known for years, I remarked that their company seems “very nuts and bolts.”

Was that the wrong thing to say? “I don’t know if that’s a compliment or criticism or just a fact,” says Taylor, who left his job as co-CEO of Salesforce to start Sierra. I assured him I saw it as more of the latter. “It’s not like you’re building girlfriends!” I noted.

It’s significant that two of the more visionary leaders in Silicon Valley are building an AI startup not to chase the nerd trophy of superintelligence but to use recent AI advances to futurize nontechnical, mainstream corporations. Their experience puts them toe to toe with better known industry luminaries; Taylor was a key developer of Google Maps in the aughts and Bavor headed Google’s VR efforts. They are eager to assure me that their hearts are still in moonshot mode. Both feel that conversational AI is an advance on par with the graphical user interface or the smartphone, and will have at least as much an impact on our lives. Sierra just happens to focus on a specific, enterprise-y aspect of this. ”In the future, a company’s AI agent—basically the AI version of that company—will be just as important as their website,” says Taylor. “It’s going to completely change the way companies exist digitally.”

To build its bots in a way that accomplishes that task effectively, pleasingly, and safely, Sierra had to concoct some innovations that will advance AI agent technology in general. And to tackle perhaps the most worrisome issue—hallucinations that might give customers wrong information—Sierra uses several different AI models at once, with one model acting as a “supervisor” to make sure the AI agent isn’t veering into woo-woo territory. When something is about to happen with actual consequences, Sierra invokes its strength-in-numbers approach. “If you chat with the WeightWatchers agent and you write a message, around four or five different large language models are invoked to decide what to do,” says Taylor.

Because of the power, the vast knowledge, and the uncanny understanding of AI’s powerful large language models, these digital agents can grasp the values and procedures of a company as well as a human can—and perhaps even better than some disgruntled worker in a North Dakota boiler room. The training process is more akin to onboarding an employee than feeding rules into a system. What’s more, these bots are capable enough to be given some, um, agency in serving a caller’s needs. “We found that many of our customers had a policy, and then they had another policy behind the policy, which is the one that actually matters,” says Bavor. Sierra’s agents are sophisticated enough to know this—and also smart enough not to spill the beans right away, and to grant customers a special deal only if they push. Sierra’s goal is no less than to shift automated customer interactions from hell to happiness.

Courtesy of Sierra

This was ambrosia to the ears of one of Sierra’s first clients, WeightWatchers. When Taylor and Bavor told CEO Sima Sistani that AI agents could be genuine and relatable, she was intrigued. But the clincher, she told me, was when the cofounders told her that conversational AI could do “empathy at scale.” She was in, and now WeightWatchers is using Sierra-created agents for its customer interactions.

OK, but empathy? The Merriam-Webster dictionary defines it as “the action of understanding, being aware of, being sensitive to, and vicariously experiencing the feelings, thoughts, and experience of another.” I asked Sistani whether it might be a contradiction to say a robot can be empathetic. After a pause where I could almost hear the gears grinding in her brain, she stammered out an answer. “It’s interesting when you put it that way, but we’re living in 2D worlds. Algorithms are helping us determine the next connection that we see and the relationship that we make. We’ve moved past that as a society.” That meaning the notion that an interaction with a robot cannot be authentic. Of course IRL is the ideal, she hastens to say, and agents are more of a complement to real life than a substitute. But she won’t back down from the empathy claim.

When I press her for examples, Sistani tells me of one interaction where a WW member said she had to cancel her membership because of hardships. The AI agent love-bombed her: “I’m so sorry to hear that … Those hardships can be so challenging … Let me help you work through this.” And then, like a fairy godmother, the agent helped her explore alternatives. “We’re very clear that it’s a virtual assistant,” says Sistani. “But if we hadn’t been, I don’t think you could tell the difference.”

Before we go farther down this path, I’ve got to blow the whistle. No matter how genuine or empathetic software agents are, their work overlaps with tasks previously or currently performed by human beings. Flesh and blood workers collect paychecks, go home after their shifts, and can even unionize. AI agents have none of those drawbacks. That’s why companies constantly use phone support systems and websites for customer interactions, and why corporations often take extreme measures to make it difficult to connect with the humans still necessary for those whose problems can’t be solved with a menu choice, or who just need to have questions answered in natural language. With some companies, it’s easier to crack RSA than to connect with a human. The story is as old as technology is.

Every time a new form of automation is introduced to shift the burden from humans to machines, companies must take care to soften the blow for customers. I am creaky enough to remember the advent of ATMs in the early 1970s. I was a grad student in State College, Pennsylvania. The entire region was flooded with ads—on billboards, in the newspaper, on the radio station—about welcoming “Rosie,” the name given to the machines being installed in the lobby of the biggest bank. (Even then, anthropomorphism was deemed necessary to soften the blow.) People eventually came to appreciate the advantages, like 24-hour banking and no lines. But it took years to trust those machines enough to deposit your check into one.

Taylor and Bavor believe that the transformative magic of AI is so good that we don’t need any softening. We’ve already been stuck with nightmare systems like phone support and websites that offer multiple-choice options that don’t address our concerns. Now we have an alternative that’s miles better. “If you survey 100 people and ask, ‘Do you like chatting with a chatbot?, probably zero would say yes,” Taylor says. “But ask the same 100 people, ‘Do you like ChatGPT?’, and 100 out of 100 will say yes.” That’s why Sierra thinks it can provide the best of both worlds: effective interactions that customers love, with the benefits of a no-down-time robot that’s not on the health plan.

I’m still wrestling with what it means for a computer model to be empathetic, as well as the equally baffling use of the term “authenticity” to describe agent-customer interactions. But maybe that’s just me. The early returns are that those concerns aren’t really an obstacle in the field. Robots don’t have empathy, but they can display empathy, and in practice this seems just as good for many people. In examples shared by WeightWatchers, people are sending heart emoji to the Sierra agents, described as “an extended member of the WeightWatchers team.” Even so simple as saying “Thanks” to a piece of code, as in one example, lays bare the willingness of humans to set aside their species bias and interact on a person-to-person scale with something that they know isn’t really a person.

Looking at it that way, Sierra’s plans do go farther than nuts and bolts. Bavor and Taylor are riding a giant wave of conversational AI to shift how all of us see and interact with the places we do business with. “Right now a company app is a little icon you point to,” says Taylor. “But at what point will the main digital artifact of a company be its AI?”

Sierra is onto something. After all, those ATMs worked out pretty well. Except for bank tellers.

Time Travel

A lot of companies have taken shots at creating software agents that do everything for you. One promising effort was called Viv, created by the inventors of Apple’s Siri. Its founders claimed to have come up with a system that would work with other services to fulfill an unlimited range of commands. I wrote about Viv in WIRED in August 2014. Samsung later bought Viv and used its technology in a chatbot named Bixby. It’s still around but hasn’t caught fire.

Viv breaks through constraints by generating its own code on the fly, no programmers required. Take a complicated command like “Give me a flight to Dallas with a seat that Shaq could fit in.” Viv will parse the sentence and then it will perform its best trick: automatically generating a quick, efficient program to link third-party sources of information together—say, Kayak, SeatGuru, and the NBA media guide—so it can identify available flights with lots of legroom. And it can do all of this in a fraction of a second.

[CEO Dag] Kittlaus says the end result will be a digital assistant who knows what you want before you ask for it. He envisions someone unsteadily holding a phone to his mouth outside a dive bar at 2 am and saying, “I’m drunk.” Without any elaboration, Viv would contact the user’s preferred car service, dispatch it to the address where he’s half passed out, and direct the driver to take him home. No further consciousness required.

Image may contain Symbol
Ask Me One Thing

Arlene asks, “The question at hand at the recent US congressional inquisition of tech executives was trying to minimize harm to children. Financial institutions are forced to adhere to KYC rules as a part of their service offerings. Why would this not work for children attempting to create social media accounts?”

Thanks for the question, Arlene. KYC, of course, is an acronym for Know Your Customer, a protocol that banks use to verify the identity of their customers. It’s used as a hedge against money-laundering and funding terrorism. Adopting this would certainly allow Meta, Snap, TikTok, and other social media companies to adhere to their policies that keep preteens off the platform.

But using that protocol would be like using a sledgehammer to push a thumbtack into a corkboard. In order to effectively use KYC to screen underage users, you’d have to demand that everyone get verified just to be on the platform. That’s not only a heavy lift, but it’s also so much friction that the platforms would lose so many people that they would become less useful. People suspicious of the law or protective of their identity for legitimate reasons (or otherwise) would bail. And who’s going to feel comfortable sharing more personal information with Meta?

On the other hand, if you suspect these companies are not really doing all they can to make sure kids are following their policies, you’re right. There’s a lot these super sophisticated companies can do to better enforce their policies, starting with studying signals from an underage user’s online activity. A Stanford study once demonstrated that Facebook could know your political affiliation and sexual orientation from just a few likes. With only 10 likes, Facebook could predict your personality better than a work colleague. With 70 likes, it knows you better than a roommate. With 300 likes, it might know you better than your spouse! You think they can’t figure out if you’re 11 years old?

You can submit questions to mail@wired.com. Write ASK LEVY in the subject line.

Sierra Says Conversational AI Will Kill Apps and Websites
End Times Chronicle

Weather experts are debating whether to add a Category 6 rating now that hurricanes can blow well above the 157 mph winds of Category 5, which is labeled catastrophic. What’s beyond that? Maybe we need new adjectives.

Image may contain Label Text Symbol and Sign
Last but Not Least

Someone is working on AI girlfriends—and it’s a privacy mess.

On the eve of publishing her new memoir, Kara Swisher explains to me why she’s got such high self-regard.

OpenAI’s new text-to-video model (not yet released) is learning how to be an auteur.

In shipping news, here’s the tale of an heiress who’s disrupting the supply chain.TKTK STORIES

Image may contain Logo Symbol Trademark Text and Label

Don’t miss future subscriber-only editions of this column. Subscribe to WIRED (50% off for Plaintext readers) today.

https://www.wired.com/feed/rss

Steven Levy

Leave a Reply