AI chooses brands. Yours?

Moderator: We're going to kick off with the first session with the keynote and um the topic of the keynote is of course AI related. So um the subject line is AI chooses brands yours question mark. So that is where we're going to dive in. Um I do there will be two amazing speakers on stage today. Um, and I would first like to uh introduce uh Jean-Marc Noël, CEO and co-founder of Trusted Shops. Jean-Marc, I would say the stage uh is yours.

Jean-Marc Noël: Thank you so much for all this, you know, great presentation. I'm so happy to be here today and you made a wonderful work of introducing that. I'm also want to take the opportunity to congratulate all the trusted shops team here in the Netherlands and also Jessica here on all the team the event team to set up this. It's a great and look they even also ordered the right weather. So that's fantastic. So let's continue with this. Thank you so much.

Jean-Marc Noël: Yeah, today it's you know I'm thinking about trusted shops and when I started trusted shops a few years ago there was being this revolution about internet and e‑commerce etc. And today uh it's again things are changing and what we want to do and it started three or four years ago. We started to figure out what can we do you know to make sure that our members stay visible in this new world of AI. What does that mean? And also what will it mean when um millions and and maybe billions of AI agents are roaming the internet and knocking on your door? What what do we need to do? So uh today we are showing you what we believe will be the future and I think it's good to be prepared. But the very first thing I want to start with you is something different something to tell you a personal story about why I started all this discussion and I want to introduce you to somebody is very dear to me.

Jean-Marc Noël: Um this is Michelle and Michelle is my mother is she just turned 82 this year and she's living in France and if a few years ago it was I think three three years ago or four years ago I don't know I was there and always you know we discussed about when and he showed and and they showed me said look I just received this letter from a bank and I don't understand a thing this was and you know as the you know the letter the banks are usually sending our insurance company very small script everything and it was all the things she she needs to do and she needs to avoid in order to avoid fakes fishing and everything and it was so complicated she looked at me and said I don't understand anything so I started to say okay look you have to look at the URL and you have to check what is the you know if you look at an email and she looked at me and said uh I don't understand anything and then I realized you're Right?

Jean-Marc Noël: Because you're not a cyber security expert and you don't need to be and when it started to and that was the time where generative AI was started to become a thing. And if you realize that it's so easy to create things which are you know um nearly like real things but are fakes because actually we are living in this era of fakes currently. So I said yes you're right. We have an issue here and we need to we need to figure out how to solve it. I don't know if you've learn if you hear something called Bogus Bazar maybe. Anyway, Bogus Bazar is um we had a few months ago uh somebody from Hamburg Hecker Company and they showed us this Bogus Bazar is an international network. It's originated in China, but it's it's acting worldwide. And these people, you must, you know, it's incredible. They created something like 75,000 fake online stores.

Jean-Marc Noël: And you we are not speaking about just copying one online stores in different in different uh with different domains. We are speaking about completely different online stores with different brands etc etc.

Jean-Marc Noël: Um what we know which for sure is something like 850,000 people um fall for that scam and mostly in Europe by the way and and you know what this and and the guys were working with with with the police and FBI and and still they're still active and this is one network from many so this is really something what's going on actually if you think about you and uh I was discussing with the other day with other with other friends and everybody has an experience on fake whatsoever even even I felt for a fake I must say okay I shouldn't shouldn't say that but it's it happens and especially and this is the time where we thought okay it's not just fraud if you think about online store and I I was thinking that about myself I went um a few weeks ago we made a big you know big trip in France and I was booking my hotels and I must say at each booking stage I said okay is is it the right page is it not the right page you know am I safe or not it was not always clear and and this is the situation you have all the time so it's really something that is happening which is not only fraud but it's basically something like hampering fraud and sometimes and I and I must say even myself I'm looking at a store and I'm looking okay that does not look really right and I want you know I want to break and I want to cancel the buy maybe I come down a little bit later so it's really a think especially if you think that each time you want to buy something brands are so important for us as as people actually you want to you want to connect to the brands you know the brands you you recognize the brands you buy from the brands so this is really something we had to come up with a solution so my my mother, you know, could say, um, you know, here I want to know that this thing is real.

Jean-Marc Noël: And the solution we call trusted login. Um, how does that work? I took I took an example from one of our member here in in the Netherlands. And if you look at the trust badge, some of you recognize that there is an additional functionality what we call um cutely the switch. Okay? and and in a switch it's an opportunity in a very simple way an opportunity to check it's everything is fine you are on the real side and if I think about you know this morning or yesterday I went through this site and what happens if I if you download the trusted app this is what happened lous.nl L verify by trusted for Jean-Marc and at that time I know this is the right side.

Jean-Marc Noël: I'm I'm sure I can go on and I'm can you know I can buy say it's basically a two-factor authentication a very simple way and I was thinking about my mother the other day that does not only work for e‑commerce this is al also working for today already 20,000 shops but it's also working for companies like Debat this is a an insurance company in Germany and you see that small switch on the top so this a functionality we are rolling out in many many cases not only e‑commerce company and for this one if my mother was looking at that site this is what happened uh her phone and this is the only thing she trust really okay the iPhone looking at it yes I'm on the right side and that's fine and I could go on you know there is no thinking there is no u checking stuff it's everything fine and I can go and that was the the starting point where we you know if you think of that new world especially with geni and we will see in a minute what happens with agents etc.

Jean-Marc Noël: Actually there is one thing we need to keep in mind. Trust is not only it's it's not a promise only anymore. Trust needs to be verified in this world where you cannot really distinguish the real from the fake. And this is really important. This is becoming even more important and we will see not only for people but also for machines. So how does that work? How do we you know do we want to do because one is one thing is certain. I was discussing just a few minutes ago. We said okay things are changing so quickly. We don't know exactly when things are coming but there is one there is a few things we know already. Soon your next customer is not a you know it's not a human anymore.

Jean-Marc Noël: it would be an AI agent and that has a lot of consequences about you know how trustworthy and how your your website and your your offering will be discovered because if you think about how brands are working so far you look at the brands you recognize it you trust it you buy it but this is not true anymore for AI agents AI agents do not you know they they do not care about storytelling or things like that.

Jean-Marc Noël: They will care about information about what we call signals especially trust signals and more and more and in our discussion with you know people like open AI anthropic and everything not only they are getting this information but they want to verify that information because if it is not verified and your reporting information is not wrong that happened by the way a few months ago in the UK judge GPT was basically proposing fake shops and everybody went on a fake shops and bought something. So you can you know for the person using Chpt at that time said okay I would never ever again use that thing because it links you to to a fake shop.

Jean-Marc Noël: You need to be to be sure that this is getting the right information and that means all these trust signals what we're looking at for instance identity or reputation accountability assurance all these things this is something we will need to provide to all these agents because if you're not providing this basically if AI won't be able to verify your brands to put it bluntly uh you will become invisible and the the point is so and this is really something we really need to to integrate and I'll come in a minute why is it so basically for AI unverified brands will become really unvisible brands why because AI does not rank AI selects AI also decides so it's not you it's not that you're coming you know you're you're getting from uh position one to position nine. Uh if you're not within the three or four things which will be provided then we have an issue. You won't be there anymore.

Jean-Marc Noël: So that's that's the reason why it's good you're here today because we need to all be prepared how to do that. In order to do that, we also worked maybe some of you have already used that. We also created started to create a few months ago or nine months ago uh a tool to help you at least prepare your website and your own store from the structure point of view to make sure you're visible. And this is something what we call the AI trust audit. Uh if you go on that site, everything is is uh is free. And you will see a lot of information uh the the the system if you give your your URL in the system you will everything will be analyzed along many many categories and it's a good way to make it that to to make this preparation in very systematic way there is a lot of category there is I give you an example um you have to make sure that AI bots and chat bots, but also agents in the future.

Jean-Marc Noël: Um, make it it's it needs to be easy for them to catch all your information. That means you need to structure your information. Shima.org or something like that to make sure, oh, this is a product, this is my offer, etc., etc. So, you come there, etc., Prox and there are many more stuff to follow. So it's a good idea to do that. The report is something like 50 or 60 pages. So it's a lot of um bedtime uh stories you can you know read and for everybody who is managing a website I think it's a good way to structure your work. Of course it's not it's a prerequisite. It does not mean if you do this everything will be visible. But if you don't do this your visibility will be impaired. So now let's move again in into this new world. So how does it work? You know today uh I'm online buyer. I'm going to the online store I'm buying.

Jean-Marc Noël: Tomorrow if you look at AI um I will invoke my agent and my agent will work on my behalf will go on my behalf on your website and we'll have a look and maybe will even uh will even buy and this is a new thing it's not just a chatbot say do you want to buy no they will I will maybe delegate also this capability to buy on your checkout so this is really something something completely new and this this AI agent is operated by a company that could be open AI or other and they will be me over time. So the first question now imagine now a world where millions or billions of AI agents are roaming the internet. The first question we want to to answer is um an AI agent is coming to your to your site knocking on your door but you don't know anything about it. Do you want to let it in or not? Today basically you have either the capability say I'm blocking it or I'm letting it go.

Jean-Marc Noël: It's very difficult but it's a little bit like you know somebody call you by phone and there is very it's anonymous. I mean what do you do? You don't pick up the phone and that could be happen. So we will need as brands to develop a new ability. The ability to check all these AI agents not only about identity, we need to understand their intent. Are they there just to get all your content and train their model or are they there to really buy something? And based on that we will need to answer. And in order to do that we evolve the word and the the role of trusted shops into a trust what we call a trust authority. What we are going to do and what we are going to present here today is an ability to create this this what we call credentials. So we will provide AI agents what I'm calling I'm calling that a passport. Uh I know the analogy is not completely correct.

Jean-Marc Noël: I know that but it's it helps me uh to give this passport and this and this AI agent will come to you and say this is my passport. This is our and with this passport there is a number of information it's not only identity it's our reputation or stuff like that and you will be able also to check together with to um with the trust authority in that case trusted shops you will be able to check is that true yes or no and if it is true I will being able to make my mind and discuss do I let you go yes or no that's the first question we want to answer but trust is birectional so the next answer the next question we ask and we need to answer is an agent come in front of your website and say uh who are you? Are you trustworthy? Can I buy from you? This is the next question and that question will be answered in a seller.

Jean-Marc Noël: So this is a green question and the great question it's the same the agent will know and by the way the trusted login you just saw before will play a role in this. we know oh this is a trusted website so I know I can get you know I can get information that I can verify I can verify this information for instance reputation etc etc identity and by doing that I know this website is trustworthy and I'm going forward and I can buy from it and this is exactly what we're doing today and what we're presenting today to you this two question okay first is this agent trustworthy this new ability you have to develop Second, make you visible to agents, visible and trustworthy to this new AI agent. That's exactly what we want to do. That's the reason why we are building what we call trusted where trusted is for us the essence of trust. It's an infrastructure for trust.

Jean-Marc Noël: It's giving the opportunity to people, brands, and AI to create and to build that trust and to verify each other. And verify each other means what we just saw with my mother. My mother being able to verify one brand in a very simple way. This is what we call trusted login. So now what we were going to say is the other way around to make sure the agent will be able to verify your website to make sure okay this is a trustworthy website and I can and I know it because I can verify it. But it works the other way around. You need also to be able to verify that AI agent coming on your site because I can guarantee you and we will see see that a little bit later.

Jean-Marc Noël: There is already today an awful lot of bot traffic around you know if you're looking at that and sometimes it's not very clear who is going what you know what is this bot doing actually on on my website so we are going to see that and by the way we have created um website which we are trusted.com tdd.com and with all the information for people want to go a little bit more detailed understand what's going on and what what we're going to do but today we'll make some demo how you know how does it work uh in in practice. So let's let's get these two thing. Let's first look at how can you know how can verify AI verify brands. So looking at our agents going to you and saying yes you are a trustworthy website. And in order to do that I make an a quick example. I wanted to buy running shoes. So this is so I ask my my favorite agent go and buy running shoes. This is what happened.

Jean-Marc Noël: And this is simply you know very quickly click clack clack click and now yeah everything is bought it's there um and not only that it's not the shoe were not only bought but it was also I know from a trustworthy shop. So how did it work? So let's you know I want to walk you through all the steps. The first steps is the agent discarded the shops. Okay look at the product. Do you have the product? Yes or no? Once we find one shop with a product, the agent is starting to discover is this shop trust worthy. So looking what we call the trusted protocol. On this trusted protocol, it detects a protocol. When it detects a protocol, you know, it just um ask for verification. You remember the trust authority asking trusted shops. Okay, do I have the signature? Is it right? Okay, get the information. The verification is done. It got all the information.

Jean-Marc Noël: the verified you know trust signals for it's you know reviews it's guarantee it's it's um also the membership plays also a role when do we uh when did we uh issue that membership etc etc and once the trust signals are evaluated trust is verified so the agent make his mind oh okay this is a good price this is trustworthy let's go on and let's let's do the purchase the agent buys the purchase is done and you go that's it. This is how things will work in future for it. And that's a reason why it will be super important for all the brands to provide that kind of information so the agent can make that decision on on the spot. And you are also sure as a user you know as my user agent I don't want the agent to run a mock and to buy from somebody I don't know you know I want to make sure that everything is fine. So this is the very first thing you you want to do. So what does that mean?

Jean-Marc Noël: Well, that mean there is an opportunity for you here because for all stores who are working fine. That means that visibility um it's not only bought that you you know it's not only you buy a paid ad by um uh Google or whatever. It's if you're working properly M50 the agent is able to verify your information you will gain and visibility in that case and that's the reason why it's so important right now to start building all this and to provide this agent with all this information in a structured way. So this is the one think on you know what we just saw. We just saw AI agent verifying brands. But now let's continue with the other way around. An AI agent is knocking on your door and you don't know who is this guy? What do I know? So AI agent and that's important to understand are new kind of software.

Jean-Marc Noël: they are really not only you know reporting or this is the list of possible shops they will act they will do something I will for instance go the other day I had this trip in in France so I ask the my agent to run through and so find the hotels find everything and it was you know reporting should I book and this is where I stop where I say okay I'm not sure I will do that myself but this is we are short in time and if the agent is trustworthy and working fine. Let's do it because it's super convenient and this will come these are some information about some the development of you know the projection of agentic commerce uh from Mckini from different uh different organization and you will see between now and 2030 something will happen and I think it's good to be prepared right now we don't know exactly when things are kicking out but they will that's for you are.

Jean-Marc Noël: And one also one other um information is already last year if you look at the overall traffic on the internet the bot traffic was higher than the human traffic and together with this human with this bot traffic moving forward you have also a rise of the what we call the bad bot traffic of course in absolutely so that's a reason why it's important to understand it's important to develop that ability can I trust this agent on my website to do this otherwise uh you know I can block everything but if I block everything I'm losing business okay that's not a good idea if I let every everybody go well I will let also the bad one that's for sure so we need to develop that kind of ability how do we do that and this is an example if uh you know simulation of uh of the system how it works so you got an incoming agent coming on your website checking and if as that's what we said just before you imagine in an agent and you don't know anything about it.

Jean-Marc Noël: That's a that the phone thing anonymous. I don't know. I deny it. I'm not taking the call. Now, if the agent is coming with a passport and said, "This is the passport. This is right. Uh I will allow the website the the agent going on my website and doing things. But you could do also things like uh we don't know. We got an agent. We know the reputation is not that good, but it's okay." So, I can restrict. I can say yes you can do but within a certain constraint for instance you know you cannot buy more than I don't know €200 euro or something like that would be up to the brand to you know to use this information and to be clever about how how can I use this in order to provide constraints or etc etc and and and you get it if you have thousands and thousands of um of agent coming on your on your website you will need that kind of system in order to find out what do I do?

Jean-Marc Noël: And you cannot do that manually each time looking at you know oh what does he want no you will have to have your system in order to answer to to answer this in an in an automated way and that means um it's important because um it it gives you the ability you know today it's more a kind of a binary def uh uh decision either you are doing this you're blocking everything or not but you would be able to establish lish policies around that and say if this bot is getting this and this kind of trust level okay this is how you know how I want to respond and now I want to tell you and to give you some hints how does it works behind the scene. So we have developed a what we call a um AI agent trust model. Um and this model works like that. You have a several dimension. There is one dimension what we call the trust level.

Jean-Marc Noël: So the trust level is basically everything which is based on the uh the behavior of the bot. It's not only information. It's not only identity. How did the bot behave along, you know, in the in the past? Is it something we know something about? If we know nothing about, well, the trust level is zero. It's unknown. We don't know. We don't know you. But if you uh worked on you know several time and we we get that information, then the trust you will be able as an agent to earn your trust. The second dimension is of course how confident are we about that trust level because it's different if you have only 10 interaction or 10,000 interaction we can rely on sometimes you know people have five star but we are only three reviews or you have you know 10,000 reviews and you have a 5.2 or a 4.2 too. Uh okay, maybe the 10,000 review is very is very we are very confident about this information.

Jean-Marc Noël: So the confidence level will be provided but not only that because a good agent can turn rogue maybe sometimes and that's the re the reason why we are including something what we call the state and the state is simply if we detect an incident we'll need to have the ability to propagate this incident and this information very quickly around the network to answer about that and that is what we call the state with different kind of states which are moving forward. So uh trust level so I repeat trust level is really um how well the agent behaves this one dimension and the confidence is how certain are we about that conclusion and that will give you an opportunity to answer and to get oh this is information now I can act upon this. How does it work?

Jean-Marc Noël: Well, trust will be earned for um by the agents of course slowly as they are working actually what we do we are looking at what we call the data trails of all these agents we are storing everything that and we are monitoring how things are evolving I will show you in a minute how does that look in uh in detail so each time the agent will have to will have the ability to move up the ladder the trust ladder But you need to have a sufficient score. You need to have an evidence and you have of course not no no incident. So we know this agent has proved uh trustworthiness in the past and we can move on and and be quite certain that this is a good a good way to go or if it is not you know if we find out some incident you will also know and you will be able to to react based on that. So these are what kind of trust signal and it's it's um it's an excerpt of what we what we're doing.

Jean-Marc Noël: So all this for instance you know the t the tenure how many days as an agent showed up uh what we call the conduct how often do websites react to it you know if you if you want to uh you know an agent will be uh refused or or not refused we store that who is behind that agent what is the identity behind that uh behind the agent there is also an operator but uh what is the domain of that operator is it something which has been registered you know uh one month ago What is the which is also sometimes a very good indicator what and also the volume some of agents are just you know you got bursts of of activity in one day and nothing anymore you want to see something on a regular basis or a sustain activity that's an example and there are many more for all this dimension we are building a score from zero to 100 and there is a weight around that which we're creating what we call the trusted score and based on that trusted score you can move up the ladder And you will see also the development of all this.

Jean-Marc Noël: It's also important for the AI agent operator to see if their agent are trustworthy in this movie. So it's really coming into into both. This is an example and what we do also and it's quite new. We are able to look at the the what we call the behavior data of all this this agent and we are able to derive what we think the agent what kind of intent the agent has because some kinds of agent would say yeah I'm going on a website I want to buy something no they are not doing this they are scrapping all the website they they got they want to get all their information and and and to push or you or they want to train their model or they want to do something some of them just want to go on your website is okay to give this you know indexing the your website. Okay, that's fine.

Jean-Marc Noël: But you probably will issue some policies and we want we want to see if this agent is also um you know making sure that those policy are being uh um you know followed properly.

Jean-Marc Noël: one of thing very simple thing you have a robot takes the uh um a file I'm looking at this do I you know consider this file or do I don't care do I try to do stuff and this is what going and this is a a real-time example you would see where we saw a a bot on our website yeah trying to get you know there is a a pattern where it wanted to look at all our vulnerability in in the software so the the system just look at it and say no this guy for 91% it's about who wants to hack your your website and you don't want that so you don't want that you can block it you at that time so trust level and I don't want to go too much into detail here from zero to five with several from the confidence evidence I just show depending on the interaction but not only the interaction it's also depending on the diversity of this interaction because if you a bug only on one website all the time.

Jean-Marc Noël: Probably a lot of interactions but not a good thing. You want to have some com some more some more things you can spread over and making sure that this this bot is really something which is a you know very very somebody behind wants to do something good right and of course last the condition the state that can you know may override that trust level. So you see it's um it's not just one dimension. It's a lot of information coming into this in order to give us this ability to be you know to be equipped with the ability to you know check all these agents moving forward. So I want to I hope this is working because it's um I didn't check that before. So it's really uh a test with you. So don't don't be I will click on this link. I'm not sure what happened. Okay. Okay. Okay. This is working. So, what you see here, try to I don't know maybe I can I don't know. Do you see something?

Jean-Marc Noël: Because this is really a big pixels here. Big pixels. Okay. Let me try to can see that better. This is a back office of our information which in the last let's say 15 days we looked at all the traffic coming on our business profile. all this business profile which are you know of course we got a lot of traffic we got something like let's have a look at this profiles at trustedshops.com it's 17 million requests so it's already something which we can we can work with and we saw something like 400 yeah 70 480 agents and agent family coming into the website so it's a lot already today and you see there is already a mix there are different category. There are you know bots who are named etc. Today there are one 0.7% of all those bots who are what we call signed.

Jean-Marc Noël: So we can really you know know this is there is an identity behind everybody else has a name but the name you can do a whatever name you want.

Jean-Marc Noël: you don't you know you don't know exactly what's going on behind and there are different I don't want to go too much into detail but and you see here the various requests over the day so we have something around 1 million requests every day on on the profile and you see here there are what we call signed and named only so here is for instance some you know some issue on and some some bots we can we can see so let's let's pick a really a bad bot So something restricted for instance this one curl curl is you know for for the developers around you means something right but in this case it's it's only a name okay it's only a name so you so let's dive a little bit into into so what did we find about that bot we have a trust level of zero we have a high confidence and state restricted it's not a good bot all right it's not a good bot So if you look at what's going on is okay these three and you see the history of we follow the history of what we call our trust scores.

Jean-Marc Noël: It started at 33 but look there has been oh there has been a lot of incidents here and now it's five. So it's really not really trustworthy and that's and and if you look at this we are looking at um incidents what we call probing. So request in known attack pass a check on you know the PHP part. We also are using some honeypot kind of uh of website to check you know if the website are being attacked and and and and and and look what is the behavior of those boats what is the real face behind that bot really okay and you know robotics violation etc etc.

Jean-Marc Noël: So that's a reason why you see vulnerability probing to 90% we know this box want to haircut and you see also by the way here the uh the the traffic what happened in that in in in in this in this box so it's a lot of information moving forward that we uh and it's only one part of the information I want to see because the other part what we want to give you as an opportunity as a service provider as a brand we want you inside that that network to we want to give you the ability to feedback on that bot. If something is coming to you and doing something you don't want, I want you to have the ability to say this is a problem and I want to give this information to the network so we can propagate this information to everywhere you know. So this is really important and I think I was had this discussion just before uh looking this is a power of the community.

Jean-Marc Noël: If we are together if we are exchanging this kind of information we will be stronger and stronger and stronger and we will be able to have this kind of information to have you know this information will be you know more accurate and accurate to re to to give you this this ability to to choose who is a good one who is a bad one. So uh I don't want to move too much into uh details but here is for instance uh you will see uh all the list of the various the various bots going forward etc etc. So the the the the key takeaway here is we are building currently a a network an infrastructure where we gather information from all these bots who are doing you know what they are doing.

Jean-Marc Noël: We are analyzing it what we call with a trust intelligence to understand what does that mean and we want to give you back this information and because we are all together the more people are participating to this infrastructure and we have already as you mentioned trusted login is one part of this of this in this rail out if we are 35 and 45 we have 1 million transaction uh a day within trusted shop so it's a lot so this is really something we will get more and more and we will kicking kicking that that that um that development. So this is um something we are building but we are not alone and I'm very happy here to introduce you to Alexander Günsche.

Jean-Marc Noël: Alexander is uh senior solution architect at AWS and this what we are building is is not only something which is propriaratory to trusted shops actually Alex convinced me that it's a good idea to create it as open source and as and this is as you know this or based on open source standard And this is exactly what Alexander what will tell us today. What does that mean? What is the standards and what how we're working together with AWS to push that forward. So thank you for much and the stage is yours.

Alexander Günsche: So in fact uh it is a collaboration that we were together.

Alexander Günsche: We had the idea um late last year with a colleague of of trusted shops Sabrina for those who know her and we sat together and said okay we see the developments that you showed and what does that mean for trust in the internet because trust is something very human so it is emotional it is not really uh something discreet but you can trust or not in one situation or another but we wanted to as you said to kind of formalize that it is something that you can transfer into this world and uh as you correctly said we thought it should be open source.

Alexander Günsche: So we because trust is about transparency as well and uh what what sense would it make to create something that is closed source and uh concealed so we want to do this in the open and so I'm going to give a perspective at this whole technology from a different angle from the angle of well for those who are a bit more technically minded to dive deep into some of the mechanisms underneath what Jean-Marc was talking about. So, what does an open-source uh protocol for trust signals look like? Who of you would be able to define what an AI agent is? I think most of you have heard it, but can you give me a textbook definition? Anybody? So, we're talking about agents all the time. Nobody knows what it is. Quite funny though. Um, but I think it's a it is something that is all around us. Everybody's using it in a slightly different way. But let me just give a quick uh let's say definition of that.

Alexander Günsche: So an AI agent is something what you would see in the middle here. It's a almost traditional application but instead of you know if and else statements that would um by a programmer who decides this at the time they code would pre or foresee certain conditions. It uses a large language model. So generative AI to reason and to make decision based on signals and one of the signals is for example the input. So you had the example of building a U trip and I have the same example of build booking a trip to Paris and would have the agent um make all the arrangements for that. Not sure if you can read all this font is a bit small but as a merchant the question is where do you live in this kind of agentcentric perspective and the answer is you would sit there on the from your perspective right hand side as a tool.

Alexander Günsche: So from the agents perspective, an e‑commerce experience is a tool that helps the agents to fulfill a certain goal and it is sits between you and the user. So you may have invested a lot into creating nice user experience, building a relationship with the user, establishing that trust, but now there's an actor with a certain degree of autonomy between you and that user and you're just a mere tool now. So what does it mean for you? If you say I have a journey through my e‑commerce experience traditional funnel it could look like uh well someone comes and browses someone would then select an article put into the cart go through their check out process and uh finally would come to uh the payment and get the confirmation and delivery and everything.

Alexander Günsche: If you think about rogue actors who would exploit that that could be in every stage of that funnel and probably as merchants you see that as well. So, it could be for the browsing part, they could scrape your website. It could be a competitor who wants to see your prices and just um undercut them. If it's about the article selection that you would allow access to through an agent, it could just come and hope you can read that. Uh they would just maybe try to deplete your inventory and then say, "Okay, I'm putting you out of competition because I have put everything to one cart that I'm not checking out." and uh might that kind of uh hinder you from business. Could also be that they would attack you through inserting fake credentials to your databases. So you would have fake identities that you store um even in a payment you could kind of check this out of the fake credit card.

Alexander Günsche: So you see this at every stage of that funnel there is a way for attacking through rogue actors. So how can you prevent that? Do you want to trust those actors? Obviously not. So what you can already see is that it's not just um AI coming in the future and that will definitely increase but it's already happening. So for scar resources, for resources that are in demand, we already see that AI agents and bots try to cut the line, try to act on behalf of bad actors. And for example, if you have sneaker drops or concert tickets, an actor or an agent would try to get in line, try to get the the scars tickets for the next, I don't know, Taylor Swift concert and trying to resell this at double price.

Alexander Günsche: So for merchants who have those scars articles or would like to manage their um AI traffic a bit better, this is a challenge today and as we've seen and you see we have some uh duplicate slides because we're having the same narrative across different audiences. Um this is going to be more significant. What's interesting here is the distinction between what we call AI traffic and the broader um category of bot traffic. So we have seen bots since decades now. It's just a program that follows a certain algorithm to probe for example for typical security problems and if your website is uh has this vulnerability they would look and and exploit that and go further into your system. Now an AI agent as I said makes decisions on the fly. It would not wait for the instructions of the programmer, but it would iterate and reason on the signals it gets. And this can be good or bad.

Alexander Günsche: It's just a kind of a a fact of being an agent. So, automatic traffic can be good in the sense that an agent would discover what you have and then say, "Okay, this is what I'm looking for aligned with the mission I have or not." If it is a malicious AI agent would see, okay, I'm not seeing this being a WordPress site, so I'm trying a different attack vector. I see that this site is throttling me. I'm trying to be a bit slower, a bit sneaky. So, it can be much more difficult to manage an AI agent because it has a more adaptive behavior than just a simple script that follows instructions that have been set up at the beginning. And this is why we are making that distinction between automated traffic as a broader category and AI agent traffic that is a new quality still more under the radar but emerging very fastly. So, and again you've also seen this slide.

Alexander Günsche: uh so we see this coming at a rate of orders of magnitude in the next few years. So Jama has introduced several terms here. So we have talked about the agent operator. I think you've also mentioned the service provider and this is a more broader term for merchants for um e‑commerce sites but there are many more use cases where we can look into a genetic and this is a good thing. So the possibility of using AI and those autonomous agents to act on behalf of users has much much more implications just e‑commerce. So we can talk about productivity applications. We're looking about knowledge work can go even into the physical world where you have robotic applications having that degree of autonomy to act on behalf of a user or an organization. So that's a very broad spectrum that we are looking at when talking about this constellation between users, agents and uh operators.

Alexander Günsche: So as Raark already said, it is not a solution to block those agents. They can be good, but it's important to distinct between the good ones and the bad ones. And so how do we do that? We see a lot of approaches today in the industry to tie the notion and the concept of trust or reduce that to just merely identity. So when you look for agentic trust, there's a lot on the internet these days. But you will see that almost everybody confounds trust with knowing who the agent is. But that's not enough. So for example, if you think back about this e‑commerce funnel example, you see that damage can be done even without knowing the user or the agent. can be that you want to allow someone to put something in card and as a typical web shop you would say hey please tell me who you are first be before you put something in card.

Alexander Günsche: So you don't want to reduce the ability to interact with your website through introducing a identity gateway and this is where trust can exist without identity but it's also possible that you need identity plus an elevated level of trust signals for high stakes transactions. it is not enough that I can get some identity somewhere you know registering uh is easy but having that degree of trust signals is much more than that so when we talk about what are trust signals the ones you already showed there's a very broad spectrum that's this kind of emotional notion of trust that we have as human can be translated into trust signals that are understood from a technical perspective and you can look at this through different dimensions you can look at um mere attributes of an actor like well the identity but also something like do they have a credit card? Do they have a company register entry?

Alexander Günsche: You can look at attestations like um let's say ISO certifications or PCIDSS which are carried through by an auditor. You can look at reputation scores very much what trusted chops is doing and is probably the most dynamic category of trust signals and would allow you so even an actor having good standing for a long time but then turning rogue you could uh see much earlier through measuring the drift and the spikes in reputation and ultimately you can even have an actor like an agent with insuranceances or asurances. So, similar to the uh trust guarantees, you could just say, "Look, here's an agent. I know who they are. I see their behavior, but this agent even has a guarantee from trusted chops that says, "I am vouching for this transactions up to €1,000." And in this case, you have a third party guaranteeing for the success of this scoped transaction.

Alexander Günsche: So, this is all in the categories of trust signals which are used is up to the implementers. So you as the merchants together with the offering from the trust authority the the main point here is that the trust authority does not make that decision. So they will give you data signals and then you as a merchant can decide how you act upon them. So either block them or allow them as you showed earlier or just say this is such a great partner we even want to elevate this have a deepen partnership and the trust authority will not interfere with it. they will be the neutral actor to just allow you to make those decisions based on data. So in this actor model you see the main actors here. We have don't uh introduce the agent um operator here. What's very important here is that the user itself is not part of this protocol scope. Why is that?

Alexander Günsche: For an e‑commerce merchant, the user is of course the ultimately most important entity, but we leave it out here. And this is for a um conflict of interest in the sense of something we don't want to do is identifying the user through this trust authority. If you were to do that, you would have legal implications about tracking users, but also ethical ones. And we are quite conscious that this is not something you want to do. We do not want to rate human beings. We want to rate AI agents and their behaviors and of course they are acting on behalf of a human being human being but we do not want to be to create a system where we say this person is good. We are rating this person and would suggest that you block it through an rating score. But this is something that is concealed behind it.

Alexander Günsche: You would of course at some point as a merchant you need to know where the shipping address of a person is. But this is not part of this entire ecosystem of trust as an authority as a open protocol. So if you're looking for that you will would need other means and this should not be part by a central authority. So as we said this is open source. For those of you are a bit more technically minded you are very welcome to whoa you can hardly read that but there is believe me or not a URL actually two of those. So there's a GitHub repository that you could check out and there's a website where you find more information about that. So the question is now if we talk about those trust signals how could you actually verify them?

Alexander Günsche: So you could uh for things like a domain you could have of course look at the domain registry for a credit card you could make a microtransaction uh and for the reputation you would have a feedback mechanism. And as I learned we have uh five minutes left. I'm going to switch a bit forward and look a bit more into the feedback mechanism. So first um I'm going to to go a bit faster here. So there's an onboarding flow that you would have between the agent operator and the trust authority and uh ultimately but I think if you are interested in technical details we can look into this in a second. Um there's several standards that is built upon. So we are using the uh webot or for identity or kind of uh piggybacking on that standard and for the credentials themselves we are using a standard called verifiable credentials with selective disclosure.

Alexander Günsche: So these standards are embedded in the HTTP transaction for those of you who are a bit more familiar with that. And you would see that we would have another header injected where the agent could just convey okay here is my uh trust signals that I have gotten from trust authority. They are digitally signed so they cannot modify them and they would present them as a um payload as you don't see but you can imagine. So for who of you knows what uh JSON is? Okay, a few. So it's basically a format for conveying uh technical data structures and uh that is basically the way this is um transferred. I'm going to skip forward here first because it's probably not of interest of many of you and also because you can't really read that. So I'm going to skip to the part that is more interesting to you and that is the reputation feedback loop. So we have the agent as an actor.

Alexander Günsche: We have the uh service provider which is a merchant in most of your cases and we have in the green bullet up up there we have the trust authority. So important thing is that the trust authority does not interfere in the interaction between the agent and the merchant that is none of their business and again also they don't care about the user behind the agent. You can try to discover that through different means but the trust authority will take a neutral position in that. However, after the interaction, the trust authority will ask for feedback from both actors and they will say to the agent, tell me how the transaction was from your perspective and the same for the merchant or service provider. Now, this opens the door to different manipulation vectors. So, of course, the agent could try to give negative feedback for some reason or no no feedback at all. Same as the service provider.

Alexander Günsche: But as we have multiple parties from both sides accumulating their feedback through the thrust authority, you will see patterns in their feedback. So if you have one agent that is biased in a negative way, you would see that it has a certain um skew towards negative ratings and you could kind of reflect that in your overrating of the agent in the reliability of the feedback. You could also have situations for example where an agent uh and a service provider collude and try to you know manipulate the way rating the positive way so that others trust the agent a bit more but then again you would see that in the overall pattern for example if just one agent has 500 transactions with one shop and gives constantly good ratings then of course you know there's something fishy.

Alexander Günsche: So through the uh the amount of large data that we can gather here and the trust authority would be able to see across the whole market you see patterns of manipulation much faster than every single shop would do that themselves and this is basically this feedback loop. What's also interesting is that ultimately there are many many more actors in this egentic economy. So we've been talking about agent operators which are basically running the agents that will visit you. We have been talking about the service providers which are most of you as merchants. And then we have been talking about trust authorities. But you will see and you will probably at some point get offers from different companies who say we are offering many many more services in this ecosystem and this could be asurances to guarantee first transactions.

Alexander Günsche: You will see agent marketplaces where actually you can launch an agent that is pre-baked for you personally through a hyperpersonalized experience. We will see reputation aggregators who may even emulate the uh data for trust authority, observability, payment solutions and wallets. So there's a large ecosystem emerging around all these main actors and you will see those in the next months and years being part of your business overall. And this has been basically the view into this more technical perspective of the agentic commerce and how trust can play a role in there. Thank you so much.

Moderator: All right. I think the last hour we have gained a lot of insights about um the direction of trust that it's not only coming from your brands but that we as a consumer are also trackable and need to be trustworthy. And Jean-Marc and Alexander um I also have a related story to yours.

Moderator: Uh, I just bought my new running shoes, although I haven't authorized my AI agent to buy them for me just yet, but probably will do in uh in the near future. Um, and I think also for you all sitting here, uh, it's very relatable. A lot of things will be changing for you as well. Uh, how people will buy from your brands. We're a little bit short on time so we leave the Q&A for a little bit later. I believe uh to the both of you you will stay a little longer right after after the presentations. So if you have any questions please stay you can uh you can up can can come up to uh to the both of them. Uh and again please give them a warm applause for this keynote. Thank you so much. And we are just about to make the switch. So you will go upstairs inside uh and also for the English speaking table that is table number seven. Um but all the roundts are inside. Thank you so much.