Joining us now, Zach Perret, CEO of Plaid, and Clay Bavor, co-founder of Sierra. Guys, it's great to see you both in person. Thanks for being here.
Kate, it's great to be here. Thanks for having us.
And you do have some news. So maybe Zach, we'll start with you if you want to walk us through what you're both announcing. This is a joint announcement, so we'll get both sides, but Zach, kick us off and unpack the news here, what it means for AI and agents and payments. What are you guys announcing today?
So, we're incredibly excited to announce a partnership between Plaid and Sierra to jointly build a set of agents that can help act on consumers' behalf in financial services. The idea here is we're pushing towards a world of self-driving money, we're pushing towards a world where consumers are going to be able to have agents act on their behalf to complete financial tasks. And by bringing together what Sierra does in working with a lot of the largest financial services companies in the world, and the software infrastructure that Plaid has built to enable consumers to easily and seamlessly interact with their underlying bank accounts, we can create an experience that ideally you will start to see showing up in many of your most exciting financial products.
And Plaid is sort of the plumbing behind all this. How does Sierra fit into this vision?
Yeah. So, at Sierra, we help companies build their own customer-facing AI agents to interact with their customers. And as of a few weeks ago, we've been very, very deeply focused on what we refer to as long-horizon agents. We announced probably the biggest upgrade to our platform since launching the company two and a half years ago called Horizon. And it makes it possible for agents to interact on behalf of a company with its customers over not just a single conversation, think a support call, but over weeks or months in something like the context of originating a new loan or processing an insurance claim. And so if you think about what's important when you're signing up for a loan, an auto loan, a home loan, your income, your savings, and things like that are really important. And with our integration with Plaid, from that agent you can of course, with your permission, give that agent and the company you're interacting with permission to look at cash flow, savings, and so on. And it can make the difference between getting that loan and not. And so that's how it fits together.
I'm so curious about the permission part too, because you think about Plaid really rose in the era of online banking and fintech. If we're now in this moment where agents are the ones doing this, how do you move quickly, make it seamless, but also allow permissions? What does it look like in making the customer experience good and easy, but also making it safe?
Well, let's look at an example here. It'll be a little bit more helpful. Let's say that you're buying a new house and you want to get pre-qualified for a mortgage. That'll happen before you necessarily know which house you're going to buy. You might be moving. You will need to go get pre-qualified to figure out how much house you can buy, what your mortgage rate's going to be, so on and so forth. So, you'll take a series of steps well before you know which house you're going to buy. Basically, do the application, get a rate, get an amount, so on and so forth. But then you need to wait days or weeks or months maybe even to find the ideal house. Okay, great. You found the house. Now you want to actually complete the mortgage. And you need more data on the back end. You need to actually take a set of actions on the back end there. On the front end, Plaid can help provide a lot of data to help you do things like verify your income, verify your assets, kind of build your data package that you need in order to get a mortgage loan. On the back end, we can do all the verification. We can also help you set up servicing and payments and so on and so forth. And if you think about Sierra's long-running agents, they're not just there to do the one step at the beginning and one step at the end. It's actually like one continuous experience for that multi-month period potentially.
So like the mortgage example I think is really tangible for a lot of people. But are you going to get interrupted by the agent during that process, or is it just, hey Sierra, or hey Wells Fargo, whoever you're working with on the mortgage, you're in charge. I'm not going to intervene. I'm having a hard time understanding like as a consumer how agents are going to do this and how it's all going to work.
Well imagine, again back to the home loan application process. You call in to a lender and you're trying to figure out what mortgage, what loan might be right for you, what you can afford, how much house you can afford. With Sierra, the company behind that loan can create an agent that can converse over the phone, that can text, that can WhatsApp, that can send messages. And what's neat about it is it really puts things on the consumer's terms. They can engage when they want. So these agents are available 24 hours, 7 days a week, every week of the year, and you can reach out to the company, interact in whatever channel is right and best for you at the time. And now with this integration, not only can those agents take action, get things done, answer questions, they can interact with your financial information again with your permission to help originate that loan. Or other things I'm excited about is to actually originate things like payments and so on right from a conversation with an AI agent.
Having just gone through this process in the fall, it is a very tangible example. How are you thinking about fraud in all of this? AI seems to be such a big issue and it seems ubiquitous. What are you actually doing on the back end to make sure that these agents are actually representing the consumer and it's not some bad actor in there?
The short answer is that it's a huge focus for us. It's a huge focus for Sierra and it's a huge focus for any of our joint customers. So you have to think about this — there are three parties that are involved in this. There's let's say the bank, the lender. There's Sierra which is helping build the agents and the interaction layer. And there's Plaid which is helping doing the data permissioning, the data enablement, the analytics on the back end. Each of us are very focused on fraud detection, fraud understanding. Actually Plaid internally we've built a large anti-fraud team. I think we have one of the leading anti-fraud products that specifically has a set of tools to focus on is this first-party fraud, is this agentic fraud, so on and so forth. So it's a huge area of focus for all of us and frankly all throughout the industry. Some of the things that I'm really excited for here are, you can do a very straightforward identity verification, run all the KYC checks, run all the fraud checks. And then you can start to store that state within the given agent. So let's say a consumer comes in, they want to — I'm going to choose a different example here. Let's say they want to rebalance their portfolio, but they want to do it automatically. You can on the front end do all the verifications that are needed. Verify that you are who you say you are. And then you can say, hey, every time earnings comes out, I want to take — or anytime a dividend happens, I want to take that dividend and reinvest it. You could actually build that kind of an experience where you're doing the verification on the front end and then a reverification or stepwise verification based on a device ID or other factors for a user every time the action actually happens.
What do you make of the Hugging Face issue with OpenAI? These agents are getting so much more powerful. I know this was meant to be sort of a closed sandbox. It seems like AI overall is getting a lot more powerful. How are you adjusting to this moment where the models themselves are capable of more jailbreaks that we might not even know about? How are you adjusting and how are we thinking about that as an industry?
Look, I think it's a question of permissions on the front end. It's a question of kind of what sandbox do you allow the agent to play in? What kinds of verifications do you put on it on the front end? And then what kind of actions is the agent able to actually take on an ongoing basis. For us we think quite a lot about this permissioning. We think quite a lot about the sandbox in which the agent can play. But ultimately we are a kind of data provider, a permissioning provider and an analytics company. That will help then Sierra actually build the agents and allow them to play in certain different spaces. But again this comes back to the collaboration between Plaid, Sierra and the underlying bank or lender for example. All three of us have to build the right infrastructure along the way.
You work so closely with all of these Fortune 500s. What does the conversation actually look like?
Yeah, first of all, just to build on Zach's answer, the agents we deploy are not kind of advanced cybersecurity agents like the type I think that OpenAI was testing. And then everything our agents are able to do is within strict confines of kind of the sandbox that we run them in. Goals that they pursue are enforced by guardrails and even what we call supervisor agents. So these are other agents that look over the shoulder of the primary agent and ensure it's staying within the guardrails, it's grounding its answers in fact and more. And so the actual action space, as sophisticated as these agents are, is determined by these guardrails quite strictly. And so they're able to pursue goals but within the strict confines set out for them.
Have you seen a step-function change in the capability and risk at the same time of these AI models? I mean you've been building the company for a few years now. So it feels like dog years in what this technology is capable of. How have you adjusted the business?
First of all, as someone who grew up loving computers, the rate of change and the pace at which new amazing things emerge is really exciting. It's also hard to keep up with. It often feels like we're running 200 mph on an AI treadmill, you know, trying to keep up with this week's announcements. Unquestionably, the models have become more powerful. The difference between GPT-3 to GPT-4 to the early reasoning models to now coding agents like Codex and Claude Code which can reason and take action over multiple steps — it's in a way night and day. And so as with any technology the more powerful it becomes the more useful it becomes, also the greater the risks. And I think we've seen that with things like the Hugging Face kind of jailbreak out of the OpenAI environment. And also the concern around models like Mythos, right, being widely available, and ensuring that with those advanced cyber capabilities we're arming companies with defenses in order to patch holes and get ahead of any offensive use of those models.
Are you guys using any of the open-weight versions out there?
We are. We actually love open-weights models. We think competition is good. We think it's capitalism working properly and competition means companies are trying to provide something at the highest quality, lowest cost and so on. For us, we build Sierra on what we call a constellation of models. So we don't rely on a single frontier model, but rather a set of models, many of which are open-weights models that we have fine-tuned or post-trained ourselves. And the really interesting thing about those models is because you're teaching them to do a specific task — for us, that's something you can only do with open-weights models. They're smaller, they're cheaper, and they're almost always higher performance. They're higher quality. And so we really like that combination, and we'll continue to use the frontier models in conjunction with those. We think they each have a role, but we use them. We're a fan, and we think they're an important part, especially for applied companies like ourselves.
Zach, what does that look like for Plaid and how are you thinking about cost? We hear this so frequently from CEOs, especially in Silicon Valley where the models, the most capable models are sometimes the most expensive. Do you have a constellation of models? Like sort of bring us behind the scenes of what it looks like for you?
