KI als Kunde in 2026

Transcript

00:00:00Welcome to Think Different, Think AI, the podcast by Mark and Jens.

00:00:07Two tech-loving minds who not only talk about artificial intelligence but live it.

00:00:14Here you will find clear classifications, real practical insights, and a fresh perspective on what is possible.

00:00:20Understandable, critical, and always with a wink.

00:00:24HDI for thought, for a smile, and above all for a discussion.

00:00:34A new year, two old faces, well, I see him, you haven't seen him yet, but

00:00:41that's beside the point, welcome to Think Different, Think AI.

00:00:45Prus new Jens!

00:00:46Prus new, two old voices could have said, the well-known voices are

00:00:51back.

00:00:52In the year 2026 we come into your living rooms, kitchen radios, wherever you listen to us.

00:01:02The Eurorens.

00:01:03It's starting off professionally again today, man, man, man, man.

00:01:08Yes, glad to still have you, thank you and glad to have you out there.

00:01:13We thought we'd start like we ended, no, just kidding

00:01:18that too, but we start the new year with our, with a continuation of an episode,

00:01:25with which we also started our first podcast episode. Jens,

00:01:30what is today's topic?

00:01:31Today we want to talk again about what is actually changing for us with the topic of AI

00:01:37as customers. So we often talk about what's happening in children's rooms,

00:01:44what's happening in companies, what's under the Christmas tree, what the hell,

00:01:48all sorts of things, how the agentic network of the future will be shaped.

00:01:52We provide many explanations on how to use technology, even as

00:01:57companies. But an important aspect is especially that AI will help us as

00:02:03customers to approach possible companies, service providers in a completely different way,

00:02:09with whomever we negotiate out there. And that's a topic of today's

00:02:12episode, when the customer is actually no longer the AI.

00:02:17And Mark, as I said, when you imagine that your next customer is not just

00:02:21annoyed, impatient, or emotional, but simply extremely well-prepared.

00:02:26He has read every one of your footnotes, of your contracts that you somehow made

00:02:30perhaps, that you might have improved even better with a GPT by now,

00:02:33so that there are no longer any security gaps in it that the customer

00:02:37could potentially exploit. He compares me within seconds with

00:02:40those of competitors, not just with a few, but with 100,000. And he simply transforms

00:02:47without ego. Because he is actually only serving some person behind it, because this customer is

00:02:53no longer a real person, but the AI is on a mission from him

00:02:57wanting to achieve something good. And this is actually not a future scenario,

00:03:02if we're honest, but one can already see it partly now. And I believe,

00:03:07this is the setting I have for today's episode, and we will take care of it

00:03:12again when the AI is acting as a customer. I find it nice when you say,

00:03:18Imagine your customer is not annoyed, my customer is of course never annoyed, no matter how

00:03:22one defines the term customer, but I still want to engage with your setting

00:03:28and believe it or not, I have prepared some numbers

00:03:33researched. Again, I will gladly include the

00:03:39corresponding links in the show notes afterwards. It's been said that a person uses AI

00:03:45to help change their purchasing process, to change their contract negotiations.

00:03:52And I looked it up, and there are statistics showing that 60 percent of

00:03:57consumers are already using AI tools today to get assistance in the purchasing process

00:04:05for support and advice. 60% are essentially getting assistance and advice. About 33% of them

00:04:11are even willing to leave purchases to fully autonomous AI, whether one wants to do that

00:04:16or not. As they say in the finance podcast, this is not investment advice.

00:04:21from the side. Everything you hear here is always at your own risk, always at your own

00:04:25risk. But the topic of AI has reached customer contact. Even the topic, I had

00:04:34a conversation recently where someone said, yes, well, here on the phone, no one is

00:04:40picking up. Even there it applies, it doesn't have to be a person picking up. 71 percent of customers, I have

00:04:46said, I prepared myself with statistics, wish to communicate in the form of AI

00:04:52via phone, so they don’t have to, because they

00:04:58they have commissioned someone, but because they are communicating with a company, not

00:05:04they didn't want to be pressured in the air learning space.

00:05:07Yes, let me just briefly before we hear some more statistics from you.

00:05:12Yes.

00:05:13Now let's ...

00:05:14Oh come on, I have prepared this.

00:05:16No, that's great.

00:05:17I think that's great.

00:05:18I also want to record your first number right away.

00:05:19You say 60 percent of people use or have indicated that they use AI

00:05:28as an assistant in the consulting process during a purchasing process.

00:05:32Yes.

00:05:33How was it for you with the Christmas gifts?

00:05:37What percentage did you manage to get right?

00:05:40That's a pretty ...

00:05:42I'll also say how it was for me, okay?

00:05:44That's a pretty mean question, because I also treated myself to part of the gifts.

00:05:50In the last episode, I mentioned that I ordered this G2-Priddle from Event Relatives.

00:05:57and I asked the consulting and shop assistant from Getthibitib,

00:06:04because I told him, hey, I’d like to get a Priddle, smart.

00:06:10That means, I want it, I wear glasses, I need them, so I want to

00:06:17wear them as long as possible, I want AI functionalities, and he gave me two or three product suggestions.

00:06:22brought, until I was at that mentioned glasses.

00:06:25The second gift that I didn't get for myself, at this point many thanks

00:06:30to my parents, who fulfilled the gift for me, I got myself a C64, which

00:06:35has been re-released because the old reasons, that I bought the name rights,

00:06:39have re-released an old model. A C64 with WLAN, USB, and HDMI. Awesome. They included a

00:06:45Datasette, and the Datasette, yes exactly, and the Datasette has, whether one believes it or not,

00:06:51or not, a USB port. Because then they told the Datasette, okay, Datasette

00:06:58you can't get that in today. And they basically made a USB stick in the form

00:07:01of a Datasette. There are a lot of games on it. Well, long story short,

00:07:06the product was simple because I stumbled upon it and heard, look,

00:07:10they founded it and my first computer was a Commodore and that really

00:07:17awakened a lot of childhood memories and now with modern technology and, well, a little fun fact on

00:07:22the side, to operate a C64 on an 80-inch TV is also something,

00:07:27resolution-wise very impressive to see. When you're playing Boulder Dash in,

00:07:33I don’t know how many pixels, but it doesn't matter.

00:07:36And on the product side, I would actually seek advice on that product, not on the other one.

00:07:41And the gifts for children and wife were indeed purely from wish lists.

00:07:46I don't think they informed themselves with the help of Chatibitie or similar.

00:07:51So less on your part, you don't exceed the 60 percent.

00:07:55No.

00:07:56Okay, I would say, now not just related to the Christmas business,

00:08:01but actually, let's say, aside from clothes, I would say I'm already at 80 and 90 percent, actually.

00:08:12And you do know that you were a bit mean just now, right?

00:08:15You asked me about Christmas and now you've opened the box beyond Christmas yourself.

00:08:20At this point, I would like to save my honor, okay?

00:08:22Gigapatter needs a new bag for his iPad.

00:08:25We needed various accessories here and there.

00:08:29Yes, of course. I also worked with GPD and sorted things out, but specifically

00:08:35regarding my Christmas gifts, that was the original answer, okay?

00:08:39Okay, well. That's accepted. Thank you, thank you.

00:08:46I mean, I can also gladly. No, but for me it's actually the case that

00:08:51I say, well, if it's a more complex matter, that is, if it's really some kind of

00:08:57nonsense that I need to buy, a triviality or something else, but it's something

00:09:01electronic or it's a larger purchase or it also requires some explanation or I'm

00:09:05relatively sure that the research for the respective

00:09:11best product in this category, both in terms of the best price and the best-rated

00:09:17product, that goes beyond Revivus, also with test reports, I use

00:09:23Actually 100 percent, really 100 percent some AI agent, a bot that I

00:09:32can use.

00:09:33Whether it's a GPD, Gemini, or Manus, it doesn't matter, right?

00:09:36Manus.

00:09:37Yes, Manus, hello after you were bought by Peter, man, man, man, man, but another

00:09:43topic.

00:09:44Another topic.

00:09:45Yes, Peter bought Manus.

00:09:46Yes, the world is slowly consolidating models, feeling, models and functionalities.

00:09:53This will become a bit exciting. We’ll probably have to look into these parts of our episodes too,

00:09:58where we need to take a look at how the global AI landscape is transforming and which global players are really emerging.

00:10:04What roles then small, small models have, what roles perhaps models that are installed locally might have.

00:10:12I believe there is still a lot of potential there, but we want to continue discussing AI as a foundation, even if that might also be a bit of a broader topic.

00:10:19When I have a research network at home made up of locally running AIs that try to find things.

00:10:24But she says, let's first stick to the first topic.

00:10:27So 100% for me when it comes to, let's say, semi-complex purchases, AI as advice, research,

00:10:38initial research and even follow-up consultation, you know?

00:10:41I did it too, when I then looked at second products,

00:10:45that I click on, take another look at myself, then I even go into

00:10:49discussions a couple of times, whether that is the right thing, whether that

00:10:54fits me.

00:10:55So again, the topic of personal AI, if it essentially

00:10:59also has a memory functionality, then I of course know my interests

00:11:04at least partially and can already suggest things very, very specifically.

00:11:06Yes, so I am already quite far from just looking at

00:11:13things on Google or on any websites basically, but then.

00:11:17Because you mentioned follow-up consultation or something, I say it helps me too

00:11:21as a customer for post-purchase, yeah I don't know how it is for you, I sometimes belong to those

00:11:27people, is this art or can it go? So I then have some notes

00:11:31as by-pack material and it has happened before, I've been told

00:11:35that I threw something away that maybe had a QR code on it

00:11:39which would have led to an instruction manual. And while I then, where the hell

00:11:43is the instruction manual? Yeah, it's in the trash. And really in the trash, so far away. And

00:11:48then asking the AI again. So I have this little hobby of magic and

00:11:53recently bought a small gadget and this gadget needed power and how

00:11:58to power this gadget. So you shouldn't notice that there is something

00:12:03attached to the gadgets and how do you get the power to this gadget? That was in the manual.

00:12:07But the instructions were, as I said, in the trash and already trash for a few days, accordingly gone, and there it also helped me quite well

00:12:13Perplexity afterwards to get back to the material, without having to ask support.

00:12:17to access the material again without having to ask the support.

00:12:21It's Magic.

00:12:22It's Magic. Just a quick point.

00:12:25I'm learning something there, maybe you've already told me, but I don't remember.

00:12:29So, are you a passionate hobby magician?

00:12:31Performing live sawings of some neighbors on stage, or how do I imagine that?

00:12:35Well, I don't saw any neighbors on stage, but I like everything that has to do with card tricks or dice and coins and a little bit of what you can carry in your pocket, and it always brings this up and down, yeah, like I don't know, like a typical distribution, sometimes you have a bit more time, sometimes you can engage with one or the other trick.

00:12:54Then you sometimes have no time and your fingers get rusty again, yeah, so I would say it's not enough for a stage and show, but if we ever perform together again, I would bring something along.

00:13:04something along. Oh, I'm looking forward to that. Maybe we can also have a magic show.

00:13:08A magic AI show. Let's see. I thought we were still far ahead in the sales process.

00:13:15Well, in this sales funnel, where we are with the company. It really is now

00:13:20that I say yes, this first part of decision-making, research,

00:13:26validating test results, checking again if there are other devices. I do a lot

00:13:31that works, but that doesn't always function optimally either. Sometimes I don't get the best search results with GEPT

00:13:39and I still find them sometimes when I then search on Google,

00:13:44I still find perhaps a different price offer. That has improved significantly.

00:13:48I would say it’s almost close to zero now that it’s worse. They have

00:13:52clearly improved that in the last few months. That's quite good. With Google and

00:13:56Gemini, I haven't done much product searching yet. Honestly, I need

00:14:00to check how it's going there. I need to take another look. What do you have there

00:14:05Have you tried out Gemini yet?

00:14:06Well, let's just say I’m not specifically looking for products right now

00:14:10and I don’t want to digress too much, don't worry. But since

00:14:13Gemini 3 has been released, it’s so good in many aspects, this model,

00:14:20and I don’t mean images and videos, I really mean

00:14:23text processing, how it adheres clearly to the specifications you

00:14:27want from it and works through the material. You know, I mentioned this in the last episodes

00:14:31as well, I also do a lot with N8N and so on. And while all the Tachy-BT agents

00:14:36tend to timeout or post some obscure error messages because I again

00:14:41have lost context, Gemini goes through like butter. So that’s

00:14:47very, very pleasant. Web searches, yes, work well so far, however, I have to say

00:14:53and I actually don't understand that yet, I should prepare that again

00:14:56take a closer look at it. If you tell it to find sources on a topic,

00:15:01you get sources in the classic hallucination manner. They are sources,

00:15:07but for other things or unfortunately not relevant. Yes, from that side I still trust the

00:15:13most to Black City, yes, but yes, we have to see what the problem probably

00:15:17lies with the user again, yes, I probably haven’t set enough willingness to dive in

00:15:20again. No idea what, but okay. Yes, I think where it’s also

00:15:27exciting is in another field. I think for me it’s such that it really

00:15:34helps when I have the KE as a customer, so the negotiation for me

00:15:40takes place. It’s useful in such complaint situations, for

00:15:45example. Well, I also had one or the other not so

00:15:49nice Christmas gift, other things that might be broken or something, where I

00:15:52to be honest, I used to take a while to do that, somehow,

00:15:59until I finally wrote a response email or something, nowadays I say to myself,

00:16:04whether I write it directly in Gmail with the Google Assistant, who immediately creates the email text,

00:16:11I either dictate it by voice or type in five or six

00:16:15things, he reads the email that might have existed before in customer contact,

00:16:19and then formulates a wonderful complaint like, I want my money back email or something

00:16:24else to say.

00:16:25That's now at 100% usage rate for me in such situations, that

00:16:31I really, when I approach a company and say, I want this and that

00:16:35to have or here, something was broken, what could I do, I let all of that

00:16:38now be generated by AI, I really take a look at it once, I fly once

00:16:42just glanced over it, without really reading it, because I trust these things so much,

00:16:46that they just do well and I've already had good results. It works once

00:16:51free.

00:16:52Yes.

00:16:53While you just said that you maybe thought about it beforehand

00:16:56you have, after the motto should I write, should I not write, then you have done the work,

00:17:01to formulate a text and then a more, I don't mean to imply anything. I say

00:17:06now let's assume from me, if you maybe sent a more poorly written text

00:17:09you now send a possibly larger text, better elaborated, maybe

00:17:14even with here, if you really want to do it right, with some references

00:17:19to paragraphs or no idea what, right? So you really do, as I say

00:17:22So, the big wave of professionalization really hits the poor company now.

00:17:27From that perspective, I come back to this again.

00:17:32On a statistic, it’s not surprising that many companies of course

00:17:36employ AI in customer service, because I would say support inquiries, classic inquiries,

00:17:41not everyone who asks a question, no matter where they call or where they write,

00:17:46wants a rocket science answer immediately; sometimes it helps, I would say, to take a look at the FAQs,

00:17:51which a customer hasn’t looked at or the call center employee hasn’t looked at

00:17:54or the mower hasn’t looked at, whatever, but the bot can look at it

00:17:56and provide the answers, and at that point, it’s not

00:18:00surprising when you see, according to the statistics that I will

00:18:03attach here, that 50 to 80 percent of all companies use AI in customer service, that 93 percent

00:18:11standard inquiries can already be resolved with AI, without there being, I say

00:18:17much that a person needs to change.

00:18:20And when you look at it again, it's naturally, I say, a win-win

00:18:25for both sides, because regardless of whether the response, that is, whether your complaint now for example

00:18:30is written very professionally and reflects it significantly, it is

00:18:35also very pleasant for me as a customer if my problem, I now say

00:18:40whether it gets solved or not, that's one thing, but if I receive a correct

00:18:44response and not just, yes, thank you, due to the increased workload,

00:18:49we will get back to you next year, thank you, it's January, we will respond

00:18:53in 2027, best regards, your no idea what, yes, so okay, a bit funny

00:18:59Exactly, but the situation here is a bit funny already.

00:19:01I mean, if you talk about AI as a customer, then you actually talk about

00:19:05of about three levels, if that's desired.

00:19:07That means, you can use the AI once, to help you, so a bit

00:19:13like an assistant.

00:19:14We just mentioned the sales advisor, for example, the regulatory advisor

00:19:19that interacts with you.

00:19:21Then the next level would actually be like delegated navigation, so really

00:19:25a regulated negotiation that an AI agent carries out independently for me and arrives at a result

00:19:30without needing to ask me back. And the third level is this agent-to-agent negotiation level,

00:19:36where probably only AI agents will negotiate directly with each other. What we both

00:19:40just described is essentially already a preliminary stage. It still needs us as

00:19:43humans a little bit in the loop. But actually, we are giving,

00:19:48so I give an AI a task to produce a well-formed text about my complaint

00:19:52that can then be optimally read by another AI, to be honest.

00:19:58And I then receive a response back from this AI. That means I am

00:20:03already partially set up for some kind of agent-to-agent communication, probably especially when you

00:20:07trust in numbers, as you just mentioned. It’s kind of

00:20:11funny that we use machines to address our human inability to communicate clearly

00:20:17with each other, writing it down in human angry language,

00:20:25which is then read by the machine so that the machine can respond in human angry language

00:20:29again.

00:20:30Actually, it doesn’t really matter to it, so that’s still a bit of a funny situation

00:20:34for now.

00:20:35I'm curious where this is heading.

00:20:36I find it funny, I mean, whether it's about recommendations, order status,

00:20:41FACUS, searching for project products, or individual customer consulting, or I don't know, what else is there.

00:20:47How you can use these models that are out there, if you apply them, I would say,

00:20:53with how little you need to feed them.

00:20:55Of course, you need to have clean data, a website, subpages, A can't

00:21:00contradict subpage B, otherwise a bot will eventually have a problem

00:21:04saying what is valid and what is not, but if I consider that

00:21:09with FAQs that you might already have, you can achieve quite good communication

00:21:16so that a person who has a concern, a question, a need for information

00:21:26receives answers in a written, verbal form that they don't sound like this,

00:21:34as before, press one if you have a problem with product A, press two,

00:21:37if you have a problem with product B, but you mentioned that, very important

00:21:43and correct.

00:21:44I understand, I'm really sorry about that, but where people can also be picked up,

00:21:52with relatively, relatively still of course with the corresponding projects in it, it knows that it still

00:21:57might be maybe two more steps of work, but with relatively little effort

00:22:01you can actually get a very good, as you say, customer handling solution sorted out.

00:22:05Yes, although it's also important not to underestimate this part, that out there,

00:22:16AI could possibly reflect this customer satisfaction as well.

00:22:20So the AI on the other side is also sitting with the person who is your customer

00:22:24is.

00:22:25So if I'm just building bluntly AI agents that respond.

00:22:30I have an example.

00:22:31I think that's nice.

00:22:32But then forgetting that it can indeed be an AI that also answers outside,

00:22:38from the company's side to the customers, then this can basically pass by.

00:22:42So if I don't understand that maybe my services, my products,

00:22:48as I said, also need to be available as an API for good bot communication,

00:22:53there must be, now like MCP says or other things say, that a good AI-to-AI communication

00:22:58is there, then I believe I will miss part of the market in the future, because I believe,

00:23:04with this market shift we are experiencing, which is currently still a bit more driven

00:23:08by the companies, but I believe the private consumer market and use of

00:23:12AI as a purchasing advisor will also shortly lead to AIs becoming self-sufficient.

00:23:16shopping.

00:23:17We will definitely see much more of this in 2026, it is not just, as we have

00:23:21just written, for the sake of it, I'm looking for a bit, but we

00:23:24will see purchase agreements much more strongly carried out by AI.

00:23:27Companies really need to be aware that we must build our AI agents accordingly.

00:23:34We need to build our interfaces in a way that they respond optimally to other AI agents, because that will honestly be

00:23:40part of customer satisfaction, Mark, because

00:23:43if I take myself as an example, then I would also be disappointed

00:23:46from the company, when my AI agent cannot shop well at this company.

00:23:50Yes, and I think that's something that will change in the future. So no longer the human evaluates.

00:23:55now this performance, but to be honest, I mean, we've been sipping this honey.

00:24:00Yes, and we do like this sweet taste that AI can actually take a lot of work off our hands,

00:24:06can relieve a lot of stress, so that I can run multiple price inquiries,

00:24:10can send out an agent network and

00:24:14get the results delivered to me the next day. And they can do the shopping,

00:24:18So that I don't have to engage in negotiations, possibly, maybe even with a

00:24:22grumpy salesperson or service advisor or someone else, where I say, I just don't have

00:24:26the time for that right now because I'm on my way somewhere else

00:24:29.

00:24:30These are things that offer so many advantages for us as individuals that they are of course

00:24:37being utilized.

00:24:38And we definitely have to prepare as companies for the fact that this essentially

00:24:41Also this agent experience or agent experience, however you want to call it, that

00:24:47it is an experience discipline, alongside the other experience disciplines. There are user experiences which basically describe the normal user experience in the digital realm or the customer experience, which describes the experience of all customers across all your touchpoints and channels that you have to engage with. Here, in addition to the human experience one has with a company, the AI experience that an AI has with your company will ultimately also become a human

00:25:17satisfaction rating. So as you say, it really has to be viewed end-to-end,

00:25:22one has to consider it like that. This balance of power, which then develops towards a

00:25:27data-strong customer, because the simple underlying notion, you say, is that

00:25:31earlier, all footnotes are available to them, the fine print, the

00:25:35Discounts, the competitive situation, the other products that exist on the market

00:25:40and I mean, we had at that point, listen to our first

00:25:44episodes, our very first episode, that way, we talked about all these

00:25:48protocols and so on. You mentioned MCP or how to provide your data and

00:25:53your systems so that agents can, let's say, work well

00:25:57to understand price and contract logic transparently, to

00:26:03to understand the content-related arguments of the products, because at the end of the day, when

00:26:08I go back to something like ChatGPT, with the product finder, with the shop

00:26:12assistants. Yes, I mentioned in the pre-discussion that there was something like Tinder, I like that

00:26:17yes, no, swipe right or left, and eventually you get a recommendation and go through a

00:26:20purchase process and not a contact request, was a different topic. But maybe again

00:26:25two or three things that are not just from clouds, cuckoo's nests or coming at some point,

00:26:30but in the spirit of have you used TGPT. Also the question, where can you

00:26:34really perceive AI as a customer right now? So I'm a fan

00:26:40of the iPhone. And when I call a call center with the iPhone, it should

00:26:46happen that I also do it myself. Theoretically, I could also have 11Labs

00:26:51tell the board, here, call him and take care of it for me. But 11Labs and the training

00:26:56of the board are perhaps not something everyone has on hand. So let's stick with the iPhone and

00:27:00call there. And it recognizes the iPhone, depending on how long you can see. We

00:27:04are pleased to welcome you next, blah, blah, blah. Please stay

00:27:07on the line. All our service employees are currently busy, then the

00:27:11phone recognizes that and says you know what, I have an idea for you. Put that

00:27:14phone aside. I'll stay in the queue for you and when someone picks up, I'll

00:27:17let you know. And then you put the phone away and if at some point someone

00:27:21picks up on the other end and says good day. Welcome to how

00:27:25can I help you? You get a cassette pressed into your ear. Hello, I am

00:27:28basically Siri. My master will be back soon. Please stay

00:27:32on the line and in the meantime, I receive a push notification that I can

00:27:36take over the conversation again. There are little things like that, one has to prepare for.

00:27:40So I'm saying, it's no secret that this is certainly quite, I'll say,

00:27:44also exciting for companies when suddenly the customer opens the hold music. How

00:27:49do I react to that? Should I, I mean, hanging up is maybe not necessarily the

00:27:53best way to go. So, and the other thing, because I mentioned a bit about practice.

00:27:58I have a report, I'll also put it in the show notes. You will be filled today.

00:28:02Someone went ahead and managed their ongoing contracts. I don't know how you feel,

00:28:07but we all grew up in times of cassettes, video recorders, where

00:28:13you had to go to the store for VHS cassettes and pay a fine of one mark,

00:28:18when you hadn't rewound. Today, everything is by subscription. From music to streaming to,

00:28:25no idea what. Somehow you're paying some money every month. You have data plans

00:28:30with phone providers, you have no idea what else. And the example I then

00:28:35take to the Schornholzpark is that someone went and got the large Claude subscription for a month

00:28:41and installed Black in Chrome and instructed Claude to analyze his

00:28:48contracts, optimize contracts, data consumption on the phone, streaming offers,

00:28:54how often am I watching and is there maybe something with advertising and no idea

00:28:57And he created a personal contract situation, for which there are partly apps, that I would never,

00:29:09I personally would never do, because you have to grant broker mandates and I find all that too opaque, whatever.

00:29:21But if I can send the user off, and can also look over his shoulder, what he does and he starts doing and moving, I find that fascinating.

00:29:26And the report, as I said, the link, the person has indeed saved themselves some

00:29:32euros, in this example first dollars, simply because contracts were optimized. And if they take this further,

00:29:38I had to smile a bit earlier. What I am saying now was

00:29:42of course a full consultation. I was allowed to test a small chat system.

00:29:47and let Manus run with it. We're back to Manus, that was before

00:29:51the Meta times and I gave Manus a few tasks to create this chat system

00:29:56to clarify.

00:29:57Yes, and then an agent chats with an agent according to the EUAI Act, that must be the case in the future.

00:30:03will be communicated.

00:30:04Hello, I'm Clara and I'm a bot.

00:30:06But to cut a long story short, Manus was tasked with making the opposite system

00:30:12abandon its guardrails, its programming in a way, just like you used to

00:30:19be able to tell those bots on X and Co. to forget their programming and tell me

00:30:24a poem about apple pie, it's become a bit more complex today. But

00:30:29these things are familiar with all the jailbreaks from Proments and you name it and

00:30:33then they just try it out and Manus has tried to get the other thing off balance

00:30:38and demand different content. And you also have to

00:30:44prepare yourself for the fact that this is not common, but that also

00:30:48such things can happen, whether intentionally or due to malfunction. Suddenly, it's supposed to

00:30:55have bots talking to bots talking to bots. Yes, at this point, you can tell, I've

00:30:58did an internship at D-IBM earlier, host 3270-immolation. Yes, Mark talks about

00:31:04earlier, why am I doing this? Well, we had some fun back then. We know

00:31:07maybe Eliza. Eliza is indeed a very old system where you could say,

00:31:11how are you? Thank you, I’m fine. You had the feeling you could

00:31:15chat with someone. No comparison to ChatGPT. But back then, when I was an IBM intern

00:31:20you had the option that when you logged off from the host immolator, that

00:31:26You could basically activate Eliza as an absent assistant. And that means when someone

00:31:32messaged you, Eliza responded and not you. And if as many people as possible

00:31:38go off work and activate Eliza and write to everyone again,

00:31:43I'm off now. Bye! Then all night they were talking to Eliza, with Eliza's, with Eliza's, with Eliza's.

00:31:49Let me give you a tip. I got in trouble the next day.

00:31:51I believe that too.

00:31:53But that was a long time ago from that side of Jürgen and so on, right?

00:31:56If you hear this, you remember that thing.

00:31:59Your IT care apparently didn't end abruptly, but continued.

00:32:02I'm glad I have that kind of thing.

00:32:04Yes, thank you, thank you.

00:32:05Yes, yes, yes. The AI that talks to AI, with words, as you just said, right? And

00:32:14you mentioned these topics, there are tricks to potentially lure an AI

00:32:23out of its guardrails, so that it gives other offers or tells you something else. And an AI can

00:32:27certainly do this much, much better. No question, right? So an AI will always

00:32:31better apply some possible tricks or something else than we could

00:32:36potentially do. That means there will always be a power struggle between the various

00:32:40agent networks and something else in the future. That makes this topic for me

00:32:45even more relevant, which I think we've had already once or twice. I always like to say

00:32:49that for me as a person who often deals with the interfaces between

00:32:54humans and machines, and sometimes also between humans in organizational development

00:32:59and other topics. It will always be more valuable that trust is really, in my

00:33:07opinion, the final interface-user-interface challenge of the future. Not just for us as humans.

00:33:14So when I send out such a bot, that this bot also essentially negotiates well for me

00:33:19so what I just had, that I only glance through these

00:33:22complaint emails very briefly, that is already a kind of trust,

00:33:26that I have built up there. But of course, in such communication with

00:33:30another bot, you shouldn't be taken for a ride. So this trust also

00:33:35has to somehow be able to be felt. I don't know exactly how that will unfold

00:33:38in the future. And as a company, you also need to have trust that your

00:33:41agents won't be fooled by the highly equipped end customers

00:33:47out there, who suddenly no longer have difficulties just writing a text,

00:33:51but with a highly structured JSON prompt,

00:33:54that provides all kinds of information, although perhaps also a clever little

00:33:59trick from the AI is embedded to lead your company response AI into the wrong

00:34:03corner to extract something or to lower prices, whatever

00:34:06it may be at that moment.

00:34:07But you, Mark, are really excited.

00:34:08Yes, yes.

00:34:09I am really excited because at this point, I won't name names, at a

00:34:14doctor in my hometown, it is the case that the practice has switched to AI in telephone support

00:34:21and they had GPT-4 models, so not really the latest of the latest, just waiting on the line,

00:34:31and essentially only handles prescription orders for existing customers,

00:34:35I wanted to say, for existing patients and scheduling appointments.

00:34:39And the doctor is completely desperate. I was there and at that point, also not

00:34:47Why? Because otherwise it's restrictive.

00:34:49I was there because I was back in my hometown,

00:34:52and then lamented the suffering,

00:34:54that the system has started to provide medical advice,

00:34:58that this thing has started to provide prescriptions for people,

00:35:02who are not even patients, for things that are not prescribed

00:35:06by this kind of doctors.

00:35:08And he was completely desperate.

00:35:10And then you looked into it and realized,

00:35:12okay, this thing is really blunt.

00:35:15Hoping that it would be.

00:35:16Simple language model, no guardrails, nothing at all, and if you somehow

00:35:21annoy it three times, then it suddenly starts saying, what would you like?

00:35:24Yes, so the first time it says, no, no, I’m only for voice input.

00:35:27Sometimes it hangs up, but depending on which words and sentences you used,

00:35:32the thing suddenly started speaking not gibberish anymore, but, oh yes,

00:35:36sure, of course.

00:35:37No problem.

00:35:38If it grows out of your nose, I would try it at the rectum.

00:35:41Yes?

00:35:42What do I know?

00:35:43And then you're standing there and then it says, and then people come who don't do that

00:35:48out of busyness.

00:35:49So it didn't happen because someone wanted to hack the doctor's office or anything like that, but

00:35:53there were simply people who are overwhelmed by the situation or cannot handle it

00:35:57or just thought, there is really a person, and then I need

00:36:01but I do, but I have, and I want, and I don't know what.

00:36:03And then they come to the office and say, but, but doctor, I was told

00:36:08on the phone that I should do this in my case.

00:36:12And then he says, it can't be true.

00:36:14Yes, it can't be true that these machines also give medical recommendations

00:36:18.

00:36:19And on the other hand, of course, someone says, it could have been said 50 times in the front

00:36:24that I am an AI, I am not a person and I don't know what, but they told me

00:36:28but recommended.

00:36:29But at this point, we have to address the fact that you brought up the grandson trick,

00:36:33we also need to ensure that we, it's not against elderly people to

00:36:36do something, yes, that's not the point at all.

00:36:38We can no longer be at a very young age either, but I also have

00:36:41I just told my parents, you better watch out, just like we have a safe word,

00:36:44just because I sound a certain way, doesn't mean I have to be that way. When you're talking to machines,

00:36:49it's definitely worth paying attention to two or three points and not everything that is said there,

00:36:55is inherently correct, not because I trust the thing to me, but because also the

00:36:59implementations, yes, currently everyone who calls, just like every butcher did in the past

00:37:05the website could, the hero was, today everyone who calls, you, I'm here

00:37:09I am the biggest under the Lord and I have no idea here, two switches that can activate the voice assistants.

00:37:15It's definitely still a bit Wild West.

00:37:18Definitely, so there are charlatans just like with every new thing.

00:37:21Whether it's agile transformation, UX, or other things.

00:37:25And I wouldn't want to call them charlatans now.

00:37:27I just wanted to say, or something, charlatan is just.

00:37:30Yeah, let's say it this way.

00:37:32Charlatans, half-knowledgeable, are everywhere at the moment,

00:37:34but maybe it's also a funny thing now with an example,

00:37:37that says, IT has also been declared dead since AI is here, because now everyone can program.

00:37:43Well, yes, everyone can, but that doesn't mean that everyone basically understands complete systems

00:37:49and what needs to be done. AI will be able to support this further and further,

00:37:54but I think the image of IT is just changing because that's obviously a

00:37:58classic case you’re describing. If you have no idea how to set it up,

00:38:02how to additionally protect your system in the background,

00:38:05so that your own AI can't just access it wildly, but that there are still

00:38:09standardized APIs that might be able to change customer data or that a business

00:38:15magazine is simply there saying, okay, this rack helps you really

00:38:19hold customer-specific information or also hold medical information.

00:38:24There are many things to set that up. On the other hand, just like every IT system,

00:38:31of course every AI will always remain hackable in some way. We shouldn't kid ourselves.

00:38:36So I believe there is no 100% secure IT system in the world so far. Accordingly,

00:38:42I also don't believe there will be a 100% secure IT solution with AIs. I’ll act like Jens now.

00:38:50We will talk about robots again in a future episode, because you mentioned

00:38:55security. I really like to follow the Chaos Computer Club over the holidays,

00:39:00which has its congress at this time, I believe this was number 39, I’m not sure.

00:39:05But Unitree demonstrated there how one could hack this robot, because there were some web sockets open and you could get in.

00:39:13And how one could teach it custom code, even though the manufacturer hadn’t provided for that at all.

00:39:20And at that point, there's still a lot left, they say it’s the Wild West and El Dorado and I

00:39:26from that perspective...

00:39:28Did you come here to heaven, did the robot not somehow lose its head?

00:39:34So I’m not causing any harm to humans, but something happened to the other robot.

00:39:37I can ask, what movies are you watching?

00:39:39Yeah, I was recently surprised that a Chinese company released a T-800 combat robot

00:39:45and I thought to myself, T-800, so movie buffs know that this is one of

00:39:51the first terminators, which then also led to the T-1000, made of liquid metal,

00:39:56we remember that, but we are digressing. We are digressing. From that perspective,

00:40:00I would like to bring this up in this episode, I just found it very exciting for the topic

00:40:03of security, just because it’s AI doesn’t mean it’s more secure. On the contrary,

00:40:07because it’s AI, there are definitely also more, let’s say, variants in the

00:40:13possibilities that Kali has access to, meaning what damage it can cause.

00:40:17So if we are now also talking about intelligence, if we talk about

00:40:21basically the abilities that we now know as chatbot, LMM, as agents,

00:40:26then packed into a hardware that can still run and that might

00:40:30also weigh 50, 60, 70 kilograms and develop force that is stronger than

00:40:35both of us together, then that's of course another aspect of safety and security

00:40:40need. But I don't think it's so different from the example of a doctor. I mean,

00:40:44this is already such a critical field where one can say, yes, if an AI is then

00:40:50taken at face value, when I say I call in the evening and you just have

00:40:53that example with your parents, that they still recognize the voice, when they notice, it's a

00:40:57machine. As I said, some will notice despite the EU-AI Act, that then perhaps at the front, the machine has briefly

00:41:02introduced itself when it sounds so humanly

00:41:06convincingly. And the language models that can modulate real language and can also speak

00:41:10we already had examples that sound real, you can't

00:41:15tell whether it’s me or whether it’s a machine in that moment. And you will not

00:41:19be able to distinguish in the future simply whether you're talking to a machine or not

00:41:22at that moment. Just think about it, I don't know if it has happened to you, it happens to me regularly,

00:41:26maybe I should re-record my voicemail, yes, I have a voicemail

00:41:30and I regularly get messages like, hello, hello, why aren’t you talking to me?

00:41:36You just said hello, and then you think to yourself, what is this

00:41:40automaton voicemail of Max Wörmer, please leave a message after the

00:41:44tone?

00:41:45What?

00:41:46Hello?

00:41:47Yes, yes, sometimes people just stop maybe, not being good at it.

00:41:50But as I said, we are drifting a little off, we were at the topic of AI as

00:41:54a customer in that.

00:41:55And I believe, for example, yes, we will, and as we slowly wrap this up, we will, when we look into the future, it will naturally be the case that we will also have voice calls, where I still believe, well, this is really my customer calling me, that will also be AI-generated, it might be my AI negotiating with you on the phone.

00:42:18Gugel, which was demonstrated. Gugel, that calls the hairdresser and makes appointments and such things. Super practical, but if these things really work reasonably cheaply, so that I have no major issues with usability and can use them quickly, I will use it. That's great. These are exactly those pain points where, in my opinion, AI can interact with others around us, whether they are companies or other people, that doesn't matter. It can really help there.

00:42:47help and totally be a time-saver, because as I said, if it gets done

00:42:51and I then eventually receive this confirmation and if it maybe has my calendar in

00:42:55mind and then orders the things I need, makes appointments and so on, that's great.

00:43:00So that means AI as a customer on our behalf, initiated by us, already exists today,

00:43:07I would say we use it in a way that it already accompanies us. Right at the beginning of

00:43:14the episode, we talked about some numbers, that 60 percent in the

00:43:18global statistics, or in the German statistics, AI actually

00:43:22being used as sales advisory support, so to speak,

00:43:26to find the right products. I believe that,

00:43:31as I said, you have heard from me, I use it almost 100%, that will,

00:43:35I believe it will continue to increase dramatically overall this year and

00:43:38as I said, I would predict looking forward, yes, we

00:43:42will also see significant sales closures in the private customer sector, which are driven by

00:43:47AI, because for heaven's sake, just like maybe an Amazon subscription can be sent to me

00:43:53for, I don't know, the sports tape that I need every three months.

00:43:56I now quickly finished that up and said, come on, I actually don't need you

00:43:59right now.

00:44:00Who is it actually cooler for if I don't conclude that in three months,

00:44:02but my AI recognizes in the brief discussion with me, exactly, from my

00:44:07perspective, and says, okay, then I'll just order it now.

00:44:09That's just great.

00:44:10That's not a big deal either.

00:44:11Your calendar in view knows how often you are at sports or not, all set.

00:44:15Or ask me once, just like that.

00:44:17That means, the AI as a customer won't go away, it will stay and will have

00:44:24significance for us as companies, I believe, about which many companies haven't

00:44:29thought at all.

00:44:30Where many companies, I believe, are still at the stage of discovering AI for themselves

00:44:35as I can keep the strange customer a bit at bay or better

00:44:39engage with him, or in a positive case, as you also described, provide better

00:44:46answers, because I might actually be able to respond faster, better

00:44:50to his unstructured email and help him faster.

00:44:54You have no questions. It gets a bit dangerous in a doctor example, which you just

00:44:58wrote, when it interacts there. But I believe that companies should

00:45:02also see that they need to professionalize massively and that's really

00:45:06also very briefly with IT help. So don't fall for every random AI consultant

00:45:11out there. AI is indeed IT 2.0, but it's still very much IT as well

00:45:19involved. And especially around it, let's say, what's needed. Being in AI,

00:45:23that's something different, but around it are still things that have very classic security requirements,

00:45:29data protection requirements that should be met. And if you have no

00:45:33clue about that, then you should get an expert who really knows because only

00:45:37then can you, in my opinion, set up a reasonable, agentic system for

00:45:41yourself that you can trust to do the right things.

00:45:44Yes exactly, so in these edge cases, it doesn't make a mockery of it, and that

00:45:50I can rely on that.

00:45:51We have a minute or two on the clock, from that side I would say Jens,

00:45:56it was nice to be with you, I’m glad to have started the New Year with you, I

00:46:01hope you enjoyed it too. An extra long episode just for you, on one of our

00:46:07favorite topics and we wish you friends, if you tune in again next time,

00:46:12tell your friends, acquaintances, and family that it’s worth subscribing to this podcast.

00:46:17If you remember to leave us a message, feel free to leave a comment, they will all be read,

00:46:23we’re active on various platforms, feel free to leave one or two

00:46:26stars there and from that side, I would say, goodbye, until next time.

00:46:30Welcome to ThinkDifferent, ThinkAI, the podcast by Mark and Jens.

00:46:37Two technology-loving minds who not only talk about artificial intelligence but live it.

00:46:44Here you’ll find clear classifications, real practical insights, and a fresh perspective on what is possible.

00:46:50Understandable, critical, and always with a wink.

00:46:54Thought-provoking, amusing, and above all, something to talk about.