Interview: Bret Taylor of Sierra and OpenAI

0:00When it comes to managing money, forget the hype and look at the results. Bill has a trillion dollars of secure payments powering our BillPay tools. Instead of just moving money, Bill is powering the financial operations of nearly half a million customers. So stop the guesswork and start scaling with the proven choice. Ready to talk with an expert? Visit Bill.com slash proven to get started. And grab a$250 gift card as a thank you. Terms and conditions apply. See offer page for details.

0:36The Economist

0:41Hi there, I'm the Sonos AI support agent. You can talk to me just like you would with anyone else. Oh, hi. Did you say you were an AI? Yes, I'm an AI-powered assistant here to help you with anything Sonos-related. I've borrowed a colleague's Sonos wireless speaker as an excuse to call up its customer support line. Because I know it won't be a human that answers the phone. I'm looking for help. So I've brought my speaker into the office today and I'd like to connect it to the office Wi-Fi. And I'm having difficulty. So can you help? Are you seeing any specific errors in the Sonos app? No. Okay, give me a moment here.

1:24AI customer support agents like this one are spreading fast. It's not perfect. We talk over each other a lot. But the agent sounds natural and is able to respond to any question I throw at it, however irrelevant. What's your favourite music? Oh, good question. I'm thinking about that for a sec. I don't have personal favourites, but I can help you play anything you like once your Sonos 1 is set up. If you want to keep troubleshooting, just let me know. It's clearly been trained to stay on task. What kind of personality have you been prompted to have? I'm here to be friendly, direct and helpful.

2:03Focused on getting your Sonos working with clear, step-by-step support. That's a pretty good description. The agent tells me how to connect the speaker. And when I hit a wall with the office Wi-Fi, offers to walk me through a factory reset or to hand me over to a human. I wouldn't call it a delightful experience. But if I really wanted one of those, I wouldn't be on the phone to customer support. I'd be reading Hamnet. The agent is patient and professional. And perhaps the best thing about it is that it answered the phone immediately. If there's anything I'm doing for society, it means you'll never have to wait on hold again, which I think that's a very popular mission.

2:50That's way more important than AGI, I have to say, as a mission. The man on a mission is Brett Taylor. He's the co-founder and CEO of Sierra, a tech startup that helps companies like Sonos build customer service bots. Unsurprisingly, he's a big believer in AI agents. And the value proposition is really simple. You know, I think if it were 1995, and I don't think we'd be podcasting, but if you could sort of imagine for a second, I'd be sitting here and telling you why every company needs a website and how it will change your business. I think in 2026, every company needs an AI agent. And the vast majority of the digital interactions you'll have with your customers will be via your agent.

3:38And that agent's going to do everything your website can do. up. I'm Andrew Palmer, management columnist at The Economist. And as for Brett Taylor, he's a lot more than your average tech entrepreneur. He built the foundations for what would become Google Maps. He was once chief technology officer of Facebook. He's been chairman of Twitter and co-CEO of Salesforce. And today, when he's not busy running Sierra, he's also the chairman of a little-known outfit called OpenAI. This bonus episode of Boss Class is an interview with Brett.

4:22At the start of this year, I spoke to him about the rise of AI agents and what that means for the future of customer service, the software industry, and human jobs. But also what it means for the current moment. Why is it that the managers trying to implement Gen.AI often feel like they're banging their heads against a brick wall? When a technology is new, whether it's the mainframe or the PC or eventually the advent of the internet, there's not a lot of off-the-shelf solutions to leverage that technology. So, you know, when mainframes and PCs came out, a lot of companies had to build everything from scratch.

5:01And then the first generation of software companies, essentially amortize the research and development costs of building that software across thousands or hundreds of thousands of clients, which is just rational. So you end up where every company in the world built their own, and then you transition to licensing software from these software vendors, and then the internet comes out. And there was an article in Wired in 1997, 98, that time period. And it was about a set of banks that were spending between$20 and$50 million to make their website's transactional, which if you read the article, basically meant adding a login form so you could actually see your stuff and not just like information about the bank, which is something someone who goes to a coding bootcamp could do in a weekend or now with tools like Codex and Cloud Code, you could do just by prompting.

5:51But at the time, it was$40 million of consulting fees to just make this website work. And this whole article was about how they spent all this money and weren't happy with the outcomes. We're roughly in that era of AI and AI agents in particular, where everyone knows that agents are going to have a big impact. If you think about onboarding a new vendor to your supply chain, you do that hundreds of times a year of your consumer packaged goods company, an agent should be able to take that and make it lower cost, faster, more reliable. If you have a toll-free number for customer support, it stands to reason.

6:26It's better your clients can talk to an AI agent rather than wait on hold and route to a BPO offshore somewhere. But right now, for a lot of use cases, there isn't an off-the-shelf solution. So you end up with a lot of people taking the raw component parts of AI, the models and agent building toolkits, and are trying to string it together. Some people can do it with success. Some people won't. slowly but surely, we're emerging with off-the-shelf solutions for some of the most important use cases and agents. Sierra, we're the leader in AI agents for customer service. There's a really neat company based here in San Francisco as well called Harvey that makes AI agents for the legal profession.

7:09If you fast forward four or five years, I'm hopeful that for your listeners, for each of the use cases and each of the key departments, let's say auditing your financials after a quarter to close, there's going to be an agent for that that you can just buy. In the meantime, because there isn't, you'll have to build it yourself. And you'll have to go through this question of, do you want to incur that complexity and cost? And if you do, by the way, you should probably prep yourself here to throw it out when a vendor is available because most companies don't want to be software companies, right?

7:38Most companies just want the job done. So we're just in the early innings. And I'm hopeful five years from now, it'll be a very mature landscape of vendors who sell agents as solutions to problems rather than people selling models and saying, here's a bunch of wood, build a house, which is kind of the case for a lot of companies where they are today. I just want to zero in on this sort of interim period. So the argument is if you wait long enough, the ecosystem will be there. There's obviously a bit of self-interest here. People be able to buy from firms like your own. In the meantime, though, should people be experimenting?

8:15I mean, a lot of this is as a general purpose technology, there must be benefit to experimentation, to playing around, to developing an intuition of the technology. There absolutely is. And it's a very nuanced decision. So I'll sort of walk through what I believe is sort of the first principles view. So AI is inherently deflationary. I mean, Ideally, it helps you do more with less. So there's a cost savings aspect to it. And what do you do with cost savings? Well, you can pass it on to your shareholders, which is valuable, but I'm a capitalist. And if your competitors have access to the same technology, someone will find a way to reinvest that money to gain a competitive edge.

8:53So there's a risk to waiting too long because if you have a savvy competitor who's able to adopt this technology more than you can, what structural advantages might compound in the period where you're waiting on the sidelines and your competitor is not. I think that's driving a lot of the urgency. And I think a lot of boards and CEOs are actually driving the adoption of AI top-down for good reason. But the other part of it, though, is consumer behavior. And I think chat GPT is really becoming the front door, you know, the sort of the consumer front door for AI. and if you look at the role that search engines have traditionally played a lot of that is being reshuffled a bit and for brands if you think about how they reach consumers that changes as well there isn't the equivalent of a paid ads market right now with ai there's a lot of people trying to help you optimize your positioning in chat gpt but it's not exactly a science at this point and i think there's some really interesting questions which is what are the role of middlemen in this new world, like brokers and things like that.

9:59Similarly, I'm probably not alone in this. When I see my primary care physician, I upload my lab results to ChatGPT before I see him, and I ask different questions like, what is the role of expertise? All of your patients are doing that. How does that change healthcare, the law, financial services? That was a long-winded answer, but I think it's very hard to predict the future. I think it is very important to experiment as a consequence of things shifting so quickly. Hi, it's Matt here from P1 with Matt and Tommy, and this episode is sponsored by eBay. Ever tried to buy a car online and end up in a parking lot with a stranger, a paper title, and some blind trust?

10:38With Secure Purchase on eBay, you don't have to. Secure Purchase means sellers and titles are verified. Financing? Covered. Registration? Handled. Delivery? Get your ride, plates, and paperwork shipped to you so you can skip the DMV and sketchy meetups. Buy your next car on eBay. eBay. Things people love. Skill Purchase is powered by Caramel Dealer Services LLC, an eBay subsidiary.

11:10Anybody who has experimented with generative AI will know that there's a lot of room for improvement. Hallucinations still happen. Chatbots still occasionally go rogue. Security is a big concern. And although the models are getting better all the time, Brett warns that some problems are inherent in the technology itself. Yeah, the models are imperfect. And actually, even more challenging, I would argue, is that models are non-deterministic. So you can give it the same prompt two times in a row and get two different answers. That makes testing and the concept of robustness very challenging because you want to be able to say, well, the model never do this.

11:53I'm not sure it's possible to say that. One thing I think it's really important to remember, though, is people are imperfect as well. So if you think in financial services, you know, one of the things that's a big no-no is a financial advisor promising returns, right? That's illegal in most countries and should be illegal in all countries because it's impossible to promise and obviously inappropriate. But people have done it. That's why there's all the regulations in place, right? And in fact, for a lot of banks, people record phone calls and have transcription, have controls in place to inspect and say, did people do something wrong?

12:28I think that's actually a fairly healthy way to think about AI as well. If you stop waiting for it to be perfect and say it will be imperfect, do we have the technical and procedural controls in place to recognize when it is imperfect and remediate it? Have we put in guardrails in place to mitigate that risk? it can be a very constructive conversation. And I think the more narrow the use case, the easier it is to put in those guardrails. So if you're a retailer and you want to have an AI agent help return an item, well, that standard operating procedure is very concrete. We have a 30-day return policy.

13:06Has the item been warned? You know, all those things. And there's some risk of hallucination, all that, but you've narrowed the use case. So I joke it goes from being a science problem to an engineering problem. And then as a technology matures, you can generalize more and more and more. But, you know, for labs like OpenAI working on AGI, that's a science problem. For a company, if you think about building an agent for a process, can you narrow the domain so it becomes an engineering problem? And you can put the controls in place so that when there are errors, you find them and remediate them.

13:35So if you think about the edge cases in your own client base with Sierra, what is the frontier at this point? I mean, you know, if we had a conversation with a bank, which said, you know, we can get comfortable with a kind of pretty simple form-filling activities. If you start to move into the advisory space, for example, it just gets way too reputationally complex and problematic for us. So where are we on the continuum? What's the art of the possible right now? Yeah. So first I've been really pleasantly surprised how much regulated industries have adopted AI agents so far. As you intuited, there are some conversations that are more complex than others.

14:16In the context of healthcare, scheduling an appointment with, say, an orthopedic specialist is relatively low risk because it's scheduling an appointment. Having an AI agent do a triage diagnosis of which specialist you should talk to is much higher risk because there's medical decision-making there. And so I think a lot of people are saying, using an American baseball metaphor, getting some at-bats with, I think, some of the lower risk use cases so that as the technology matures and in particularly the regulatory landscape matures, they have deep experience with this technology so they can evolve towards those more sensitive use cases.

14:56I think it will actually evolve more quickly than people expect because AI, while it is imperfect, it is actually more consistent than humans. And so if you go back to the giving financial advice, there's a ton of risk. AI can hallucinate. If anyone tells you there's no risk, they're selling you snake oil. However, if you look at a very large financial services firm and all of the communications they have with their clients, what percentage of those are imperfect? it's probably very high. Our expectations of human perfection are just much lower. And so we accept that. And so I think AI can actually improve the robustness of control, whether or not that's the commonly held belief.

15:40Now, I believe it will become that over the next few years. One part of my brain is still trying to work out what an at bat is, but we'll leave that. It just in terms of customer reactions, because you kind of describing like people need to adjust to the idea of agents, of AI, humans are fallible, this is better. But to what extent do customers of your customers know that they're talking to an AI and how do they respond to it? All of them do. All of the agents built on CIRA identify as an AI, say, hey, I'm an AI. And actually, most of them say, I occasionally make mistakes too, because it's a trust building exercise, And the customer satisfaction scores of the AI agents built on CIRA are incredibly high and almost uniformly higher than the human interactions that preceded it.

16:28But one of my favorite conversations that one of our clients sent me was for a telecommunications company. And an elderly man called because his receiver wasn't working, so he couldn't watch television. And spent more than 30 minutes on the phone talking to this AI and ended with, thank you, you've been a good robot. It was fascinating to me just because the degree of patience this AI had with someone who probably in these interactions would have encountered a much less patient person on the other side just because of the inherent cost of staying on the phone with someone for that long. And just the empathetic gesture of thanking it.

17:07He was quite aware it was a bot. I think we'll be pleasantly surprised just because AI agents can speak your language literally. Here in California, there's English speakers, there's Mandarin speakers, there's Tagalog speakers, there's Vietnamese speakers. It's now free, effectively, to provide multilingual service. Infinitely patient. There's no one behind you saying, hey, get off the phone. You have to do 10 more phone calls a day to reach your quota or whatever it might be. And, you know, can accommodate even idiosyncratic things. It doesn't mean it's better for everything. I don't mean to imply that, but my co-founder, Clay Bavor, has a great way of putting it, which is for the history of computers, we've had to learn how to use computers.

17:48Think of the first time you used Microsoft Excel and just how intimidating it was. AI agents learn you and you just talk and it figures it out for you. So I think it's a really humane, really positive evolution in the history of computers going from punch cards to keyboards and mice to touch screens to now just speaking. What about some of the hidden costs here? So are there humans in the loop when people are interacting with your AI agent? And how often, what does a handoff look like in practice? There can be and there should be in some circumstances. I'll just give you an example like the mortgage industry.

18:24If you're refinancing a home or buying a home, some of it is the actual presentation of a mortgage, which probably should be done by a banker. But a big part of it is collecting information like your income and assets and credit. And that's an example where an agent can just sort of augment an experience that you have with a banker. and the banker can spend less time on collecting forms and PDFs and actually just doing what he or she does best. I would say it should just be a business decision. Do you want the AI to sort of be a co-pilot, if you will, to a professional subject matter expert? Do you want it to be autonomous?

19:01Similarly, the client should add agency in that. If you want to talk to a real person, you should be able to. The nice part about these AI agents, because they're good and they're not like the old bots that everyone hated, Most people are opting into using them. So you can end up actually having, I'll say, higher quality, more expensive interactions with your people because you've essentially unlocked a bunch of budget by taking a lot of the simpler transactional stuff off with AI. We have one client who had offshored their customer service who's now on shoring again, as an example, because the volumes have changed and they can afford it now and they thought it would be a better experience.

19:37So there can be some really interesting counterintuitive second order effects here. But to the extent that models can still go wrong, you've got to have some method of monitoring. Is your contention that that can be another model? How do you mitigate that risk? I like to think of it as defense in depth, which is a term in tech circles we usually use around security. Most companies now have a chief information security officer. And what defense in depth means is first you try to prevent anything from going wrong. So you lock all the doors. But then you say, okay, if a bad actor does get in, can we detect it quickly and limit the blast radius?

20:15So that's where you end up with, in addition to locking all the doors, you make sure there's good monitoring on everyone's laptop. And you end up with all these layers so that even if one of the layers ended up with a vulnerability, you've mitigated it. I think the same should be true of AI. So I think the first layer should be AI monitoring the AI. At CIRA, we use a concept called supervisor models. And they essentially supervise the decision-making of the underlying model in real time and say, you know, is this a hallucination? Did they actually follow the standard operating procedures? Did they break one of the guardrails?

20:52And that's in real time. And then you can have a longer, more intense model, which is essentially after the conversation is done, evaluating it. You know, was this a low sentiment conversation? Did the AI agent repeat itself too much? things that you can sort of look at with the context of the full conversation. And then you can, you know, maybe put those conversations on a queue so people can review them after the fact. And what's really nice about that is rather than just looking at maybe a random sample of conversations, you put the needles at the top of the haystack, you know, so the problematic ones are there.

21:26And the whole idea here is you use a combination of AI and humans in the loop, using AI to help those humans that are in the loop so that they're not just wasting their energy, but actually spending their time on the most sensitive, the most problematic conversations. I also want to just talk about evaluation. So if as an enterprise, you're trying to build things for yourself and trying to define what good quality is, what performance is, that seems to cause people quite a lot of trouble. You're inside organizations helping customers through that process, as I understand it. So what are the kind of big problems that people hit and how should they resolve them.

22:03This is going to sound reductive, but I think one of the most important things you can do as a business leader is specify a business outcome you're trying to drive more than a technical outcome. I think it's really important when you think about your business processes where AI agents can apply, what are the key business metrics you actually care about? I'll just give you one that's simplistic, which is what percentage of calls don't people need to talk to a real person. Well, it's easy to make it a hundred percent. You just don't let someone talk to a real person and you or any consumer knows that can be an insanely frustrating experience.

22:37So that metric's just gameable. It's not a great metric on its own. Things like customer satisfaction score coupled with a metric like that are really productive because you tend to get a mix of consumer sentiment plus the sort of business outcome of, you know, how many of your clients could help themselves without having to wait in a queue. And that feeds into your own pricing approach? That's right. We do what's called outcomes-based pricing at Syra, which means we only charge our clients when the AI agent actually solves the problem on behalf of the customer and we have to escalate to a person that's free.

23:13Our whole philosophy is that software, ancient history, four years ago, software was, you know, a productivity tool for a person. And if you asked a salesperson, was it your CRM system or you that drove that sale? Of course, they're going to take credit for it. Well, now an AI agent, if it's one of the CR agents answering a phone call or answering a WhatsApp chat and handling a customer service inquiry, you know whether it solved the problem. You also know the customer satisfaction score and why not pay for a job well done rather than pay for the privilege of using the software. So I want to just end by opening this out a little bit more into kind of where you think things are going.

23:51So you've described how in five years, say, the ecosystem for vendors will be that much more developed. You're in an interesting position where you've got feet in sort of two camps, right? With Sierra, you are one of those vendors, the application layer. You're also chair of OpenAI, one of the foundation models. If you take us forward five years, does the foundation model do everything? Is there a threat to your business at Sierra from a more competent Claude or ChatGPT? I don't think so. I wouldn't have started Sierra, but I think it's a really important question. And certainly in the cocktail parties here in San Francisco, it's what everyone's talking about.

24:30One of the main things that these foundation models can do well is generate code to produce software. And if you haven't tried it, you know, open chat GPT and, you know, ask it to write an app for you or a website and it will, and it's pretty good. And so you look at that and you're like, wow, the marginal cost of producing software is going down dramatically. What does that mean for the software industry? Why would I license a piece of software? Can I just go to, you know, an AI model as they generate that software for me? And for my entire career, there's been a shortage of software developers.

25:03It's the scarcest resource at most firms. So it's a really important question to ask for a variety of industries. Not this is a software industry, but the consulting industry, like the ramifications are dramatic. And as with anything so fundamental and so big, it's hard to actually predict the second and third order effects correctly. My personal opinion is that most companies don't want to build and maintain software. So even if the act of generating a piece of software goes down by a lot, you build it, you own it. And so I'll just take an ERP system as an example. There was a new accounting standard that came out for software companies called a decade ago.