AI Agents and Identity Wallets: Preserving Privacy in Data Sharing

Hey everybody, hope you can hear me okay. Yes, we can. Excited to Wonderful.

Well, great. Excited to be here. So glad that I could be here virtually. I think if I wasn't working on identity, I would probably, you know, aspirationally want to work on, you know, being two places at once or teleportation or something. But anyway, glad to be here.

Even, even if it is remotely, I am excited to talk about AI agents and identity wallets. I'm gonna go ahead and share my screen and hope that y'all can see.

So I'll, I'll go ahead and step through and will, will, will there be q and a at the end here? Depending on time?

Yes, there will. Okay, great. Great.

Well, I'll try. I, I like q and a, I think it's useful.

So I'll, I'll try to leave some time for that. Yeah, let's, let's dive right in to AI agents and identity wallets.

Okay, when I talk about, when I try to paint a picture for people about what the future of identity looks like or what, what that will make the world look like. I talk about the, the, the millionaire's experience, right?

I, I mean, I imagine somebody who, you know, wakes up and, and they walk outta their home and there's a car waiting and they, they step in the car and the car knows where they need to go, right? And suppose they're traveling, right?

They, they, they, you know, the car takes them to the airport and instead of going through the main line with, with all of the other, you know, people or whatever, they go through a back, a back way, and they get on their airplane and it takes them where they need to go. And, you know, they, when they get to their accommodations, wherever they're staying, you know, they can just walk right in.

They, they don't need to wait in a line or a queue to get their, you know, key or something like this, right? If they, if they get sick, they just, you know, go up to a, a, a medical clinic or something and can share the information that they need to, to, to share in order to get, you know, the care that they need. And they're not worried about their, whether their insurance or whether they're, you know, I don't know, whatever.

All, all of the, all the, the little things in life that add up to friction and, and, and that take time out of our, our day from doing the things that we really wanna be doing. I envision, I mean, I've never been a a hundred millionaire or anything, but I envision that, that a lot of those folks have people who take care of a lot of those details for them. And I think, you know, when I, when I paint the picture of what the future of identity will enable, this is a lot of times what I use.

But I, but I actually think that this is strongly enabled by identity, but also very strongly enabled by AI agents, particularly AI agents, which are empowered with data about you and can help you make decisions and, and help you operate, you know, your, your life. So that is what I think of when I talk about an AI agent, right?

It's a software powered by probably a, you know, a large language model among other things, perhaps, where you can delegate tasks and it can accomplish those tasks on your behalf, including tasks which require a deep understanding of you, your preferences, your, your requirements, et cetera. I also think of an AI agent as something that can proactively, you know, have access to certain of your tools and of your information and proactively, you know, take actions to make life easier.

But there's a few things that could be true, a few things that I've seen that, that I think we don't want in an AI agent. The first thing is, I think we don't want tight coupling between learning and the specific model that you're using for a couple of reasons.

What, what I mean here is that if, you know, chat GBT learns my preferences or how I like to engage or whatever, and then I move over to another model, I mean, Chachi, BT is not really an AI agent, but you know, you get the idea, then I need to start from scratch with this new model. And I, I think the, the future will be, you know, I, I think we'll have models that are fine tuned for healthcare and for, you know, therapy or something. Or we'll have other models that are fine tuned for, you know, business use cases or, or, or, or, or, or what have you.

And as an individual, I would like to be able to take my data or my preferences or, or the things that, that, that the context about my life between different applications, particularly if I, if I realize that one software is perhaps being, you know, not doing the best things with my data or, or, or, or something like that. And I want to be able to, to move easily between things, right? So I think having a personal data store, which is separate from the model, can be helpful in that way.

The second thing is, generally speaking, we don't wanna put just all of our data right into the prompt and let something go, go wild, right?

So we, we like, like, you know, hey, AI agent book me a flight, and by the way, here's my passport number, here's my social security number, here's my, you know, and it's just, you know, you know, first of all, obviously typing all that in is not gonna make any sense, but, but even just giving the raw information to an AI is risky for a variety of reasons, particularly if you're in an enterprise that has to take care of customer data and, and you need to be thinking about where that data's going and have auditability and control over that and things, right?

Like obviously enterprises will never, you know, never be putting PII or any kind of sensitive data directly into prompts for AI agents that can, that can go out to the open internet. And then the third thing is, you know, we, we don't wanna have to reverify ourselves and reenter our data repeatedly across the internet. It would be great if my AI agent could scour the web and find me the best rates for an insurance product that I'm interested in, in buying. But the best thing that it can do today, and I've, I've tried this is, it can take me to the website in question, right?

Like, like even something super simple and innocuous. Like I asked it to find me the best newsletter on product management, and this AI agent took me to a webpage and said, here, here you go. But I had to enter my email, I had to confirm my email with the two FA, right? I had to go through all of this extra stuff. And if I wanted to do that, you know, for certain use cases, I would be just repeating this process over and over and over.

And so, you know, there's a few, I I, I think like when people think about what AI agents will look like, a lot of people picture Jarvis from Ironman or something, if you haven't seen that. Jarvis is the, essentially the AI agent for, for, for Tony Stark, and he has full access to everything. So if you look on my slide here, it's sort of the far right hand side where Jarvis has full access to everything. Tony has access to all of Tony's information, and probably more than Tony has access to as well.

And, you know, well, I don't, I don't wanna rule that out as a futuristic kind of sci-fi potential outcome or whatever where we have a a god model that just runs our life for something. I I really think that that's, you know, I don't, I don't see a line of sight to that necessarily being the case in, in the near term.

On the other hand, if you go on the, on the other end of the spectrum, this is where we are today, where every single new session, I mean chat, GPT, which is the most popular ai, you know, chat bot tool, you know, just recently in the last, you know, months got a memory feature where it can start to remember things, you know, about my past chats or that I've had with it or whatever. And, and that memory is stored in a, in a certain kind of data store.

But, you know, every new session, every new agent, every new whatever is completely from scratch every time. All the context I want to give, I need to put it all up front, right in that initial kind of prompt.

And, you know, I don't think that will be the future either. I think the va although there will be some edge cases and use cases for that, of course. So I think the vast majority of use cases that we'll start to see over the next few years are gonna be in this middle bucket at the peak of the sort of distribution here.

And that's where reusable identity, portable identity, identity wallets, whatever you wanna call it, will come into play to enable people to take the context from, you know, one environment, not just the context, but the context and, and the identity and permissioning and, and all of that from agent to agent throughout their, their, their life.

So when I think about the three problems that I outlined here, the, the things that we don't want, and I think about where things may be going, I think the thing that people will wanna be aware of that is happening right now, a trend that that that will be important to understand that you could apply if you're an AI agent developer is the concept of identity acceptance networks.

So as irin, two weeks ago we launched a product with this name, and since then I've had a bunch of conversations with other people who have been using this terminology, and I think we'll see more, you know, more, see seeing this kind of terminology more frequently used.

But, but the concept is basically that there is, you know, regardless of where my identity information is, regardless of which wallet it's stored in or which personal data stores that I choose to use, an acceptance layer is a, is a layer that allows, you know, me to bring my identity from, you know, any source or even multiple sources and present it for the context at hand. So I, I had a, a, a, a quick demo here to show what this might look like, and we'll specifically show one use case, which is solving problem number three here, the reentering of data repeatedly.

But in the last slide, we'll talk about some other technology that I think could bring to bear to, to address other, other problems as well. Okay? So in this demo, I'll interact with a, a, a AI tool provided by a, a fictional flight booking service that, that will need some identity information from me to book a flight. I'll go ahead and, you know, do that and then book the flight. So let me open this up here.

Okay, so this is just an example of what this might look like, right? It's a very simplistic example, but I wanted to have something that you could see that could spark imagination. So first of all, you know, suppose I'm on speedy flights, I need help booking, I click on the, their chat bot and have a, you know, an interaction with this agent who after, after some dialogue decides it needs to verify my identity, I click on the verify identity button and I see a, a, a modal that gives me the ability to connect my identity from a variety of sources, right?

And since this is, yeah, EICI configured it to show, you know, some European eids here as an example. But you know, for me, I have clear, I'm a, a member of clear, so I'll go ahead and click this, and through this I can sort of authenticate through Clear, which is a private sector reusable identity, which is mostly focused in North America, come through and share my information and, you know, I, I'm able to sort of share my information with the AI agent and then it can go through and, and, and, and book my flight, right?

So this is a very simple little demo that could show what, what, what it could look like to, you know, prove your identity from a, a wallet as a conceptual term in order to, you know, accomplish a task. Okay? Yeah. So this is, that, that's what we call identity acceptance, right? Is it's basically where instead of re-verifying users from scratch every time, you can sort of allow a user to bring their own identity. Saw somebody with a shirt here, I can't see it now because my, my screen is gone, but somebody was in the room with a shirt that said, bring your own identity.

And this is really the, the concept and the term identity acceptance is the term that we've been using intrinsic to, to, to convey that concept. Okay? So the last thing I want to talk about is just some of the key properties of identity acceptance or identity wallets as it relates to AI agents. The first thing is that the agents should only see the information it needs from the wallet. I think that this audience will know that, right?

And, and understand some of these things, particularly around data minimization and so forth. But, you know, the AI agent, ideally, I'm not giving away, you know, giving the agent the key to my entire password manager with my 400 logins or whatever, right?

I, I mean, ideally this is getting what it needs for the scope of the interaction that needs to happen, and that I'm able to consent to that. The second thing is, ideally I'm able to bring different sources to play, right? So for example, I can bring my identity from one source, my professional licenses and credentials and educational credentials and so forth, could be in a different wallet.

And, and I might bring that into bear, you know, in, in, in a same workflow. And maybe my context from previous interactions with other AI agents is stored in a personal data vault, a DW and solid pod or some other kind of tool purpose built for that use case. But I think the, you know, that this is how I see the, the world sort of playing out is not one single wallet to rule all ev of everything, at least not in the near term, but rather that, that we can bring different sources to play.

Selective disclosure is another example and, and sort of predicates so that, again, I can prove what I need to prove. And you know, we, we, we, we like verifiable credentials and so we, we use that. And then I think the, the final thing is that there's a lot more infrastructure needed as I've dug in with AI agent developers and understood, you know, how they're approaching this.

There's, there's three sort of key things that I think would be very, if, if you're working on any of these or if you, you are aware of other technologies that are being developed to enable these kind of use cases, I'd love to hear those as well from you.

So the first one is verifiable encryption, or, or the idea that I can take some encrypted thing from my wallet, give the encrypted thing to the agent, and then the, the agent can deliver that encrypted thing to the relying party, such that the relying party can decrypt the data and read the data, or, or at least they can, you know, maybe it's tokenized in some way where they can then retrieve the data from the proper source or whatever.

But, but the point is that the agent themselves doesn't have access to the information in the interim, but rather it's connecting the dots and, and streamlining the user experience for me without requiring me to, or, or, you know, share my PII with the agent. The second thing is, I, I put in quotes, doggy door. This is like my favorite term for purpose-built interface for agents.

But the idea is that, you know, just as humans use visual interface or, or other types of interfaces, certain, you know, brands and so forth are thinking about interfaces for agents to interact with and interfaces purpose built for bots with, you know, APIs and, and other, you know, protocols for exchanging data in, in a, in a, in a way that's purpose built for agents as opposed to, you know, force fitting agents into using like web UIs, for example. And then the, the third thing is purpose-built data stores.

So data stores that will, you know, take the, that, that, that will help sort of take context from one place to another. I think the, I've seen some good attempts at using personal data stores to move context between agents, but I think we'll get a lot further with purpose-built ones that can, you know, compress the data that, you know, it, it can take a lot of data and, and, and, and transform it into a, into a format that an agent can understand more quickly.

So I think that that is, yeah, that's, that's everything I, you know, if you're working in this space, I'd love to, to chat if you're interested in identity acceptance. That's also a topic that I'm, you know, spending a lot of time on recently and, and would love to connect.

And again, happy to be here remotely. And if we still have time, I would love to, you know, happy to entertain questions or whatever. Thank you very much, Ryan. A big thank you to you. So in the interest of time, we don't, we won't take questions, but please, if you have questions for him, please direct them to, to, oh no, please send them to Ryan Irin and continue the conversation there. So thank you very much. We'll take a break for lunch for an hour and a half and we'll be back here at two o'clock, two 30 CET. So thank you very much. Take care.