Agentic AI Use Cases with Real ROI - The Boardroom Memo

0:00 We are now really checking on their processes and hey, how can we automate it? Sometimes the solution can even be not with an agent >> because we can just build better IT systems that support your process and automate the flows. >> It will have a completely different UX. It will not be just clicking. It will be combination of clicking next best action. So there will be some kind of a support agent. people need to be involved and need to understand the output of the AI and take the right decision. I mean I disagree with you but you said that Google is not I mean maybe in the last one year or two years maybe they were not with the innovation at the forefront for some things.

0:44 >> The boardroom memo powered by Net Economy. Welcome back to the boardroom memo, the net economy podcast where we cut through the noise on enterprise technology and focus on what actually moves the needle. Today's topic, aentic AI and the use case that deliver real ROI. We'll move past the hype of 2023 and 2024 into the Atlantic era of 2026. Joining me are two people who live this every day. Buban Chavich, our AI and data department lead and Maximillian Tornado, our head of technical enablement. Let's get into it.

1:25 >> I would say when the chat deputy got released, I think that the next year was a lot of this kind of a PC. So, everybody wanted to do do a PC. So, it doesn't really matter about which topic. It's only about like, hey, I need an AI use case. I need to show this to somebody. But this led to having like PC's that really didn't prove to be successful. I think some kind of a percentage is like extremely high like 90 something% like not not being successful. But over the time it's the the customers also in the the market changed a bit. So it's a bit more like do we really need this? So it was PC era then last year it was a Atlantic era. I still think that the the whole value is not yet proven as it's like hyped already. Uh but yeah, this year I think the expectations are really high. I think the the CEOs are a bit losing the the patience of you know they would they like to find the value for that. I mean for me what where I see a lot of difference now is in the beginnings we had like this chat GPD wrapper which basically just used the API of chat GPD >> and doing some nice system prompt on top um at the end the value was quite limited to to all of those um applications but now we are also going a bit deeper with agent architectures basically where we really hook into into the core of the business or into our um e-commerce shops and try to to solve real issues and not just having something that you can throw in a chat at the end >> um and get to the same result. Um for me also as you said the value is now a lot more asked and and and basically um we are going a lot deeper into what can we actually achieve with this technology and not just hype around in my opinion >> exactly and I think part part of this story is also about like everybody was expecting like in few days you were done. So traditional machine learning you need months maybe you will get some value and in this I just create something in few hours few days I have it in a production it's working which is far from reality so and I think like many clients really were disappointed because you maybe you can achieve something really fast in like couple of days but to productionalize this to have this everything running in a secure environment to add scale to you know to serve the thousands of customers I think this is in the end I think the complexity just shifted to uh the governance and the infrastructure and the micros service layer building agents is kind of easy of course it still has challenges but the things around it are still not less complex to the you know comparing to the traditional software development cycles >> yeah I think also that's that's what we experienced in the recent years um that building AI agents at the is not far away from traditional software engineering. So we are >> started somewhere with machine learning engineers building machine learning models but it turned out that LLMs are powerful enough um in order to have more a lot of use cases I would say covered without any touch of the machine learning >> model itself or the model itself >> which is quite interesting. Um yeah, but also I mean at some point I think in the in the beginnings >> we had like use LLM for everything. So like like don't even write a deterministic logic just let the LLM do it. And initially if you just look at the results I mean LLM by nature are pretty confident what they're saying and you're just like oh this is perfect this works perfect. I don't need to code anymore this stuff. But then when you dig under the hood you see that it's completely like wrong the results that it's spit out. Then we also had this kind of a deterministic workflows which is still I would say the most important stuff that you just really have a clear separation. What is a workflow? Where do I need determinism? Where do I need some LM magic to actually get some data which is not possible.

5:48 >> Yeah, exactly. Yeah, I mean that's that's something where I really discuss a lot with my engineers currently because my engineers tend to to do everything as deterministic as possible and the LLM somewhere in the middle of of of the workflow very tiny very very testable >> approach at the end but I think this is also something we need to learn over the next projects where is the cut between deterministic workflows and how free can we provide let the LLM run at the end in order to still get a good enough or the perfect output at the end. I also think for the engineers it also needs to be this kind of a mindset shift because they are really used as you say for these deterministic workflows also like coding in a way. I mean this is maybe a separate topic for the for the vibe coding and and stuff but we recently had a situation where we had to build like the API services for prices and availability are not working. we are now waiting for the client for weeks to give us the access >> and I was just like why don't we just create a mock service and just white code this in like one hour and it doesn't really matter for the use case and they were like yeah sure we can do it but please don't evaluate the the code quality and don't judge me well I don't care we will just throw this away >> and when we get the API we just integrate this and just do it like that >> I mean I think that's that's for Sure.

7:20 One way we need to find for ourselves like how much testable is something at the end or how perfect need does the code need to be and when is it okay to just write throwaway products which you touch once or or never. >> Yeah true. I mean also I see I mean based on the on the Google next that we attended I mean also Google is investing a lot into this whole agentic game as everybody know they now did this whole rebranding of vertex AI to agent platform so this is also like going in a direction which I was talking initially like have the surrounding things around the agents as pluggable as possible in a way so that they can just focus on agent building and then you have agent registry you had all the monitoring in place, all the stuff because this is what where the most time was spent and also with their Google ADK which is really integrated in their ecosystem really in theory you can really fast deploy the agents and make them work at scale and I really like their approach.

8:28 Yeah, especially what what makes it really interesting is this uh scaling from proc code agents where you can build everything define everything it's registered in the agent registry but also the noode agents which are on the Gemini enterprise platform built by our business users at the end >> um are registered there and you can provide them access into every kind of uh GCB um service etc which is the first time I I personally see that we could probably bridge business user build and automations with actual engineering work um in in one platform.

9:11 >> How do you how do you see this split in the future in a way? How much would it be pro code? How much drag and drop and click and click? H that will be interesting. I I think there there are many many platforms currently going around like Naden um things like that where they are really interesting and business user tried to get into it. Um but at the end you still need a lot technical knowhow in order to properly build such workflows and and and understand them. And since Google is now going into the direction to say okay we have an agent builder in Gemini Enterprise but we want to keep it as simple as possible for business user to understand and use and on the other hand we have the pro code with Google ADK uh the agent development kit um where we can build nearly everything as engineers I think it would will be even like 70 30% a split where 70% % are most probably even the business users because they know what they need.

10:16 >> They know what what they want to achieve and they somehow can build it in their in the workflow with guidance probably but they can build it and 30% are really the the difficult hard nuts we need to crack as engineers where just it's not possible because the connectors are hard to achieve. We have still a lot of legacy software in our companies where we need to to attach them where no MCP server every or anything is available or we have workflows that are just going over multiple steps with with many involvement of other systems where you need to to prepare the context in a very very um refined way and that's that won't be possible I think for any business user to achieve. Yeah, true, true. I'm also thinking I mean uh there's a lot more and more requests that we we see across the company about you know vibe coded apps like small scale only for my department or only for few people >> which I think I mean we are talking about AI a lot but in the end we are all just want to you know make our people more productive or save costs on incre increasing efficiency and so on. with it this widecoded apps. I think there is also a huge potential in there because not everything needs to be a agent in a way. Maybe somebody just pipe code an app in two three days and if we we provide some kind of a platform where they can just easily deploy this and make accessible to users in a secure way and govern way. I think this would also bring this performance boost that we all expecting also. But don't you think we we we then push the people a bit away from their core work actually so they are now fumbling around with a lot of AI tooling and building apps in in uh for their work but not doing their actual work. So what do you think is there any >> I I think I mean how I see the people I think there is a lot of motivation for people to try this out. I think many people are even experimenting this outside of work just because it's interesting. There are people who are really like naturalb born builders. So maybe they're not coders but they like to build things.

12:33 >> And of course I'm not saying now that everybody should just do it in their free time. But in a way if I now can invest one day two days to for one year to save for my department I don't know hours or days of work uh per per person. I think in the end that's the core benefit because in the end I think during the software development we had a lot of this kind of a problems where developers don't really understand the client or the domain expert but now if we're bu bridging this gap and bringing the like capability for the domain expert to build the application not really overblowing it like they are now spending months on building this but something really small scale which makes sense that they can just easily deploy it provide a lot of value. But for me exactly that is is also a bit the the challenge we now have um because we we have now technology where people that are by far not used to building software are capable to build something that provides value. That's fully true and and we saw it also in in our company that this is now happening >> or we see it actually. Um but still the outcome of this tools or or the how the tools are built is not what I would call production ready and and something you should scale on on a whole company and things like that.

14:04 >> And I think they and this is something we we are also currently working on with with management and so on. we need to define a way on how to provide such tooling to the company because I think it's fine if if you build such a tool and use it for your own if it works for you it's perfect and yeah >> then maybe two or three other users see okay hey someone built a very nice tool how I also want to have it now you come into this deployment problem I would call it like you need to change your application that it works for multiple users. You need to deploy it on a platform that it's accessible. Um probably you have a database somewhere flying around because you want to store the data and things like that. And then the things start to get um difficult to achieve especially for users that are not used to building s such a software.

15:04 And at this point also software engineers need to hook into this process. um or operations engineers at least to support and guide the people and this is something where we are currently thinking about multi-ear projects where you can say okay you have a small application you want to have it for a few users feel free there's a platform deployed there with some readme or whatever you will figure it probably out but if you want to scale it to a company level you most probably still need some engineering work in order to to make it multi-tenant um capable and scale it at the end. Yeah, true. I mean for the company scale that for sure cannot live within one department and on one person because there is also so many things to to consider. I mean with this smaller apps I think all of them even like on a lead level like I don't know engineering leads have some alignments about the stuffing. This is still used really like in in Excel. They're filling this out, populating all this stuff. And in for these five people, if they can do this things automatically, just a nice nice UI where they click and collect things and they can just make it visible. Of course, we can think about the agents doing this for them. But I think this is even the next level and we should really >> try to see >> what am I trying to achieve >> before I really introduce an agent to do this for me fully automated end to end.

16:33 But I think this this is also where the where the companies are quite stuck currently because we currently have the promise in the market like hey just build an agent it will solve it for you but for building an agent we need to have data foundation and in addition we also need to understand the human process in order to build up those capabilities and this is I think where most of the companies currently um are stuck in the in the process of figuring out and um also we stuffing process is I think one of the best examples in our company because it's very human-driven >> um we have department leads that understand how what people we have in in this whole department and we have um a lot of discussions until one of our projects is stuffed to find the best possible will wait for the project at yet. Um and there is a lot there are many attributes included in this process like okay um what are the skills of the employee?

17:45 What is the situation of the project the employee is currently in? What is the situation of the project the employee will be shifted to? >> Um do we have location specific um restrictions? Do we have language specific restrictions? >> Motivation of the employee. >> Motivation of the employee. Exactly. Things like that. And um this makes it hard because we need in order to build an agent that can do it, we need to have this data somehow available. So we need to provide it and this is really hard because it's in the brain of the people and not somewhere in our systems right now.

18:23 >> Yeah. True. True. I mean also you know customers now are getting smarter of course and they would like to have like a concrete >> KPIs behind everything. Okay, how much of a performance boost will I get if I get this agent? But exactly what you said, they're not even currently tracking how much time do they spend. So if we take a simple example like I don't know recruiting and how much HR is spending on the first screening phase.

18:50 We would like to maybe introduce an agent that is doing the CV screening and then giving advices to the HR maybe doing some kind of a uh you know not even grading but some kind of a helper for the HR. >> We don't even know how much time we spend now and people are many companies not even tracking and they're now expecting okay please you need to promise us we expect a KPI of 20% of performance boost which is always like a you know chicken and egg problem. We should not stop the innovation before we get all the data right. But still it needs to somehow make sense because in the end we are just ending up in a gut feeling this will probably help me. Did it really help? We don't know. We think it is. Some people think it it doesn't.

19:36 So yeah, but for me it's also interesting to see because this is I would say the first time at least that I saw that companies are trying to push so much for return of investment because in earlier days or before AI basically started um there was a project we built some features be and barely looked at at KPIs or thought about okay what are the KPIs we will measure as soon as we implement this this feature um and now every agent use case is basically okay but how much more efficient can we get now with that and >> yeah yeah yeah >> that's what the mind of people >> exactly exactly that's also what we we saw with with with one of our customers where um we we started uh building up agents for for internal processes And the first thing was okay what is the efficiency gain we get and we really needed to ask the the the customer first yeah okay how efficient are you currently and the question the answer is most of the time we don't know >> yeah it's really the fun thing >> exactly I mean also I mean I also interesting how when you go into this kind of a workshops and everything so when you are discussing about the process how much people you need to understand the process end to end even the people from the same department it's hard really to to drill this okay this is how the whole flow looks like >> and people are often I think also surprised on the workshops when they get the full process driven the the like design and drawn >> they're like confused that this looks like that >> also but still what what I I really like is that we now somehow started to think about process efficiency and how we can automate more and more processes with actually computer systems >> because yeah okay every every um company uses computers and and um IT systems and many of of our bigger customers or all of them actually have ERP systems where you think a bit about automation but now we are also heading into into different other departments like um marketing where We are now really checking on their processes and hey, how can we automate it? Sometimes the solution can even be not with an agent >> because we can just build better IT systems that support your process and automate the flows.

22:21 >> Yeah, true. Do do you see more potential in the this kind of a internal use cases when it comes to the agents and AI >> compared to >> compared to for for the clients like for for our indust industry for example for the retail business for the end consumer or more on the internal efficiency gains. Um, I think jumping into AI for retailers and and our e-commerce shops is something we cannot and should not avoid. This is something we you need to do. I think Daniel mentioned it in the in the podcast recently. Um, he also said this is something that will just happen and the the the companies need to jump on this train in order to to survive and also to continue. um selling products at the end. Um for here the efficiency gain is most probably on the end customer side researching products things like that >> or experience maybe better experience >> better experience yes but there is also a huge potential when it comes to internal process automation with AI and and sometimes I tend to say it's also easier to start with because >> if you have an e-commerce shop and you serve that to a few thousand clients every day. Um, building an AI agent that is bulletproof. I think we have a lot of cases where this didn't work out well.

23:51 >> Yeah. Um but internally you have you have your trusted user base basically and you can also scale um and try it out and find use cases um that are small enough in order to to to start building gain experience and then can scale bigger in order to to really um have an end to end flow. So I since also from my natural role um as as um automating our own company I see a lot of potential and also use cases internally and I tend to say start internally gain experience because it's a safe environment and then we can always scale to to the end customers. Yeah, I also think that somehow customers or organizations in general tend to be a bit too how how to say it like they're looking for the golden AI use case that will enable them days of increased efficiency across the organization. But I think if you just they just get lost into this kind of a search for this use case for this golden use case which doesn't really I mean maybe it exists but until you find it until you implement it probably it's something much more sophisticated and it requires time if you are just focusing on small gains maybe these departments say few days this for some other use case few days on a smaller things these things add up in the end you will achieve much more than just running around in circles having a pool of 50 use cases and asking yourself okay what which one should I pick I don't know >> yeah and the the interesting part here is we are always looking at use cases from processes we are currently doing and I think internally we have we have a very interesting project in in agent areas where um we actually built an agent for process which was before not possible to achieve and not working um because it took so much time in order to do it that we actually didn't really fully do it. Um so what we did is we have in our um pre-sales uh area the problem that we are faced with a lot of company names basically and we have somehow an understanding what our ideal customer is. So basically we um know okay a customer is on certain range of revenue um it has a certain structure and it has a certain technology portfolio then it's a good fit for us and now you're faced with 1,00 and more customer um names and you need to choose what customer might be the next best customer for us. something that is not achievable or not easily achievable in with just human work. And so that's why we basically um choose customers based on personal connections to to certain certain customers um but did not really go into datadriven selection of the next best customer. And this was the first use case we actually did with an agent. now where we then said okay we can now do something like deep researching with an agent in a matter of minutes instead of someone researching over days for a single customer.

27:41 So what we really what we really achieved is um is we can provide the agent a list of names 100 200 and the agent can come back in a few hours with a completely scored ranked list of potential customers which is is um shaped for the needs we have on our customers. And this is something we were not able to do before. And even with AI help, it was it took us a month in order to get to such a list before without >> building the Exactly. Exactly. Exactly.

28:22 >> And now that we with we are really able to say, okay, we have a new list, 100 customers, throw it into into the system. The agents go on a deep research journey in different areas and um put in a new scoring for for compared to the existing list and we know okay those are the 100 next best customers for us and this was really an interesting journey for us because it was first of all the first journey we did on Google ADK and second the first really big agent system we built and uh it turned out to actually quite successful because we now can do something in ours which was not possible to do in months. No, what is in the in the end the feedback from the salespeople? I mean I don't know how deeply are we evaluating then these results. I mean of course you cannot really cross check everything but at least for the customers that you kind of already know you can just skim through the list of I don't know key stakeholders key technologies and see if there is some hallucinations in there or not. Yeah, exactly. So, basically we it's not perfect. We're still talking about deep research and and AI systems.

29:39 Um it's not perfect, but it's good enough to give us an indication um where we could look into it into next and it's also good enough to have comparable um results between the customers um potential customers which which we have. So for us it's a very valuable system uh already also the sales um it was in very very near collaboration with the with the sales department um and also they told us this is really valuable already um to to continue and we also um agreed together with management that we will invest more into it actually but I'm I'm also often asking myself like >> what what means good enough and when does it you could stop in a way because in the end you can improve this indefinitely >> and when until when you say okay let's pause it for now let's continue in if the technology gets a bit better and so on so how do you see that >> I think I think good enough in this case for example is if we can do a decision on what are we going to do with with this list next Um and we also said okay now we can introduce more third-party API integrations like providing LinkedIn data providing >> um more tech feed data like we really know that this customer is using SAP CRM products things like that um but at some point we also figured out okay we know it from googling >> because that's what research does. It goes to Google, tries to find a match between a company and and and the technologies that are relevant for us.

31:36 >> And most of the time it it works at the end. And um >> this is also where we need to to understand where to make this cut because at some point we say, okay, we now have this 100 customers. We want to approach them in some way. Um and then the actual work of our sales department starts because you then need to understand the customer. You need to to to find context. You need to to get in touch with them. Um and this is also where the people needs to start researching on their own as well because this understanding is not by just getting AI output at the end. It's really like investing time into into this customer or a potential customer. Yeah, I guess it's it's also like you know until the 80% you get pretty fast >> and then the remaining next 10% is kind of difficult and the next 10 is even more difficult and I guess also it's on the on the management and CEOs to decide okay when is the border and I think also now it will go a bit into this kind of what you say new features but to evaluate why do we need it and what do we expect to gain from it >> because maybe you know we can I customers and and organizations can really go into this kind of a trap.

33:04 Let's try to get something more. But if you now need to invest additional I don't know one month to get a fraction of a percent of increase does it really make sense? I mean maybe in this case scheming through LinkedIn if if we don't have it could potentially make sense but if we want now to go in I don't know social networks to see what the companies are discussing about all of this stuff maybe there there is not a huge value compared to some other other topics. So it has to be prioritized >> also. We should not go into the trap to say we want to replace people here.

33:40 >> That's that's not not the goal. We want to augment the people and their their 80% with AI is most probably the right fit because the last 20% is the percentage basically where people need to be involved and need to understand the output of the AI and take the right decisions because overall if we would provide AI decisions or decisions fully to AI I don't know if a company would survive.

34:12 at the end because also there are multiple studies around could an AI lead a whole company and manage a whole company and we are not at this point I would say >> yeah I think it cannot even at this stage really maybe some specific roles of course but in the end I think the biggest fear of people realistic fear in my opinion would be like my role will change >> and I will do some things differently >> and maybe this differently. I don't like it >> because for example maybe I would like I'm if I'm doing like I'm if I'm a designer right and I was enjoying working in a Photoshop doing everything manually all the layers all the image changing and so on for now I think we still people like for a professional grade uh image uh generation you still need to use Photoshop but now it's different now you have AI you circle this say I need to change this please move this around >> this is a bit of a different work and I think some people maybe not maybe it's not the best example but I think simple some people are also reluctant to go in this direction because they like their job. Also we we had like you know people browsing through you know to the website looking for some sources and materials and trying to get into this kind of a creative mode. But now AI already did this for you in a way because it has access to all of your data and then if you just ask the question you would get probably like 90% good response and it will be different but some people maybe have a a joy of going deep into different websites leading to the reading to the blogs investigating the topics and so on. So this I think it's it's a big bit of a mindset shift also for people that is needed.

35:58 >> But for me it it's it's like this this going deep into a topic. Um this is also sometimes at least for me the the necessary part in order to get a real understanding of the of the matter. So >> I don't know I'm I'm into mountain biking. If I would just throw things in into AI, deep research, whatever for the next best mountain bike, I will most probably not understand the the geometry of of the frame. Um, what forks are very well currently, things like that. And this is something if you have a hobby or something like that, you most probably want to jump into that. And that's where I think um also sometimes you just need to to go deep and do it on your own in order to fully understand it. But sometimes if I'm not into that like um I don't know coffee machines. I I'm not into into the coffee machine so much. So I just want to have a coffee machine that makes my coffee in the morning. then I would just use deep research because hey, what's the best price for the best coffee machine I can get? Um, let's go with it.

37:18 >> Yeah, I mean my workflow is is a bit different for example because I start I'm into coffee machines for example >> and I like to use deep research but just as the initial trigger >> and then I go deeper and deeper into these topics. Maybe >> it can happen that I don't even open a website for that. Maybe I now see uh different what tools do I need to make a perfect espresso. So I see I need a port of filter. Then I will continue the interaction with the AI to ask things about that. So you know my way to understand what is relevant also maybe I will trigger another deep research to look at it. I mean I'm losing maybe you know all the nice images and different stuff but in a way I also save a bit of time not going to content which is there is always like boilerplate content in all these blog posts and the informations and so on. So I just take it as a different medium for example.

38:16 >> So you're trusting AI deep enough that the answers are correct enough let's say it like that because it won't be 100. >> I mean if if it's something really important for me which I don't know it's about safety or something like that. I for example tend to also you know trigger a deep research focused only on like uh scientific publications and things around that. And based on that you get also a feeling that this is much more correct because if you just focus on the blog post >> anybody can write a blog post.

38:49 >> I mean also for the for the research that also can be true because there are many researchers which don't really make sense. There may be some I know tested on 100 people but also if it's really really important then I go and I open the publication scream to the skim to the abstract for different uh see also not every scientific paper is equal in its like uh references. So yeah but >> that's already maybe too deep but I think we went a bit off off the topic in general. uh but what do you see like uh for for the clients like what are the the main use cases they are interested in?

39:31 I think since we are coming from traditional e-commerce um most of the clients until now approached us for e-commerce features um and and agentic commerce at the end um which is very relevant and we we also need to be experts and gain a lot more knowledge in in that area. Um but now I I have the feeling and you are also involved in in many many of the or all of the projects actually um the clients are really now coming also with internal automation >> um internal agentic workflows um where can we hook and also there I have the feeling clients really ask for guidance.

40:25 >> Yeah. I don't know what's your your feeling on on that. >> For sure. I we also had recently a situation where a client had like really nicely identified use cases. Uh there is even some kind of indication of you know costs or expected effort saving and stuff like that. But it's like 50 use cases and you know I mean this is also done between the departments. So I mean in theory you can have some kind of a ranking list >> but to really properly evaluate the impact I think this is where they need our supports most often. So even we even have that situations where you know we are not maybe that deep into organization but just because of our experience we're just looking at the use cases and like just challenging this from from our perspective we really see something what will work what will not work maybe something where technology is not ready yet maybe something that they can just use a tool they to build a custom use case so and and in a lot of cases I think there is also this kind of a struggle do I need a chatbot or do I need a workflow?

41:34 >> Yeah. >> Do I really need this chat experience? >> Exactly. And >> especially this chat experience um is is something Yeah. Okay. We have large language models which are language models. Um so that the chat experience is quite the natural thing to to um start with. But also here I think we now need to learn on how to get a bit away from this whole chat experience and providing >> um a better h more human friendly interface because I think every one of us is a bit bored of reading through all these AI messages. I think there is even a trend called AI fatigue where >> where you get >> get really really um tired actually of of reading through it and there are many AI um vendors that that are now looking into different approaches on on integrating um AI output or providing AI output. I think that the number one use case that I see that the clients in in our industry are interested is chat bots.

42:51 >> Yeah. >> Maybe sometimes there is not even a vision behind what this chatbot is will provide to me. Uh because there is still not a proven value. I mean there are numbers you know 20% of uh conversion rates and all of this stuff but you can never you need to put all of this into a context to really see does it make sense for me as organization or not and somehow it's a bit natural because also the competition hey maybe my competitor also now has a chatbot I need to have a chatbot do I really need it is is my uh search for example regular search working >> is my product data ready for this because If you have 10 attributes for each product and no descriptions, the chatbot cannot help you. I mean, it can guide you a bit, but if you ask any kind of a specific questions, hey, I'm looking for a dress which is good for wedding and has some kind of a luxury vibe. If there is nothing mentioned in the product data, chatbot is not magical to just, you know, do the stuff. Of course we can now discuss about okay we now trigger a Google search we find the data but this is like over complicated.

44:05 Yeah, sure. Especially we should not compensate um bad data with more bad data because googling can also go horribly wrong at the end. >> We had I mean u when it comes to googling and and bad data. I mean we had in the beginnings like really few cool use cases which are around product data enrichment. We also have for one client like implemented in production like product description generator based on the attributes but also this is where they have a flexibility to do some kind of a configurations tone of voice style.

44:39 So it's not just a simple block of text. It's really quite complicated and this is already something which will will give them additional boost. But also when it comes to like really researching the product attributes because many customers are really struggling to get this >> I don't know 20 30 attribute per product they don't even know where to start. Some companies simply don't want to invest time because it's definitely time consuming. But for some other companies, it takes a massive amount of time just because it's a simple workflow. You get a new iPhone if that your supplier didn't provide you exactly the attributes.

45:17 >> They are googling this or reading through materials and saying, "Okay, this makes sense. This doesn't make sense." It's a errorprone and timeconsuming process. And we got some use cases in a in the end in a PC stage where you can do a deep deep research about the product attributes. Maybe sometimes there will be a clash for one attribute and you need to select as a human which of them makes sense. You will get the references and you will probably believe more the official iPhone website than some local dealer.

45:50 >> And this is something which I think provides even better value. The problem with this use case is for example what when we build it the connection with the actual enterprise system because currently how we build it it was working in an isolation >> you can really do it nicely and have it exported maybe in CSV but since the whole human in the loop thing you need to validate to see to check some stuff because it's it cannot be fully automated yet this is where then the real investment comes in because building this kind of an agent is easier here than having it integrated end to end when I accept this that it's already in my pin that it's already on on my website. So this is where the complexity also somehow it's hidden but it's huge especially this human in a loop um problem I would even call it is is still something we need to figure it out when when to pull the human in um in into the AI process. But for me that that's a very interesting use case. Um because if first of all you have work that is very manual and requires a lot of time in order to get this attributes uh right. Um and the second thing is if you have your attributes right most probably your search will also work better. You have already in your in your um shop and also your SEO ranking as well. Um and and that's I think the for me the right way to think about AI features. Not hey everyone else has this chatbot I also need a chatbot.

47:29 Um more like where can we really make a difference for us for our customers at the end. Yeah. >> Um, also we had one customer that was kind of bored of the of the um chat interface and then we also here um in in a PC build up a system that recommends um order improvements, order efficiency improvements >> like next best best action or something.

48:00 >> Yeah, exactly. Exactly. And and this was also quite interesting because we hooked it completely different into the interface. We we did not provide any chat like interface for now. We just said, "Okay, you're on the card. We know you're buying different other stuff um from time to time. Maybe you also want to add that to the cart now." Um or you are you're ordering in different clusters the same product. Um maybe you want to combine it in order to save delivery costs or or or um handling costs. And um I think that's also also a smart way of thinking about AI features.

48:40 Um where can we provide the feature to the customer and not get the customer into the AI at the end? >> True. True. But there are also customers on the opposite side of the spectrum where they're like >> I don't like the chats as it is in a way. There is a chat bubble. I click in here I get a small thing. I would like to have like a half of the page now in a chat chat interface. I mean I think the or the full page. I think also in in the future probably the way how do we interact with the web shops will be quite different.

49:15 >> You know maybe we will get into some kind of a this is just my expectation potentially the next next two years >> that it will have a completely different UX. It will not be just clicking. It will be combination of clicking next best action. So there will be some kind of a support agent or somebody who is like from the AI observing what we do providing next best action. I can ask questions it can recommend products for me. So I think this whole dynamic UI ch changing and shifting this is somehow future. Let's see if we get there and how how soon that will come. But that could be quite interesting. I think we should start building a prototype just providing a prompt box nothing else and then let's see how how how yeah but now I also saw like yeah this whole >> training of the models with the corsor so that it can do actions for you so in a way with a cursor voice and not even text maybe uh maybe neural link and then we just >> Exactly exactly >> browse and shop our our stuff >> but maybe maybe for sure this will be a lot stronger like just talking to Siri on on your iPhone, uh recommending products, things like that. This is for sure >> uh a much stronger sales potential in future.

50:35 >> How do you see in general like the the Google position in this whole AI game? And are we on the right track? Uh betting on Google, >> it's always always a back and forth, I would say. um when when it comes to selecting a vendor because sometimes vendors are much upfront um in in this game like OpenAI was at the very beginning. Now people or companies are are um a lot nearer or even before uh Open AI. But what what I also saw on Google next is that Google has a solid foundation I would say. Um and we are we are also early adopters of this whole platform. Um we we jumped we jumped on the platform right at the beginning. Um and we also saw the evolution over the years now. Um and Google was maybe or is very often not the first one to to to provide something. Um there are also entropic and so on on on the market but Google was always providing um capabilities on a solid level um where we can also count on them and also with Google ADK Gemini enterprise and so on and also the the CX stack now um we we have really a platform where we can fulfill I would say 90% of the customers needs It's >> e either with just pro configuring the products or building it on the platform.

52:18 >> Yeah. I mean I disagree with you but you said that Google is not I mean maybe there are >> in the recent in the last one year or two years maybe they were not with the innovation at the forefront for some things. If you take a look at Nano Banana for example this was a game changer for everybody. This is like sooner with LLMs. You know, over the course of the last 15 years, they were always, in my opinion, the innovators.

52:43 If you compares to some other companies, companies like Microsoft or AWS, I mean, all the AI innovation came from Google. Transformers, it came from them. So, all the topics, I think they made a mistake. Maybe they took this whole LLM game a bit too lightly in the beginning. Yeah, I think the the whole size of the organization was something which is which was blocking them to be there at the forefront. So, OpenAI recognized the opportunity. They had a smaller team, much leaner, as smart as on Google, but they didn't have like this whole you know uh anchor of the big organization that stops innovation.

53:24 Fully fully agree on that point. Um also we we always need to understand that Google is not not a small scale startup and and also people at Google told us >> um they have never seen such a dramatic shift as it now happened with AI basically in their organization because every every part of the organizations there is going into AI >> um which is really interesting and also they always proved us that they can deliver uh in form of LLMs with Gemini 3 3.1 um the models are a lot a lot better and comparable with any any competitor on the market like OpenAI and and anthropic >> models >> I mean they also have it like end to end in a way they close the loop because they have GPUs they are deeply integrated into ecosystems across the world they have talented people and they have So in a way I think I I really think they're they're on a good track and I'm happy that we are like their partners.

54:32 >> Um but maybe I mean compared to SAP and this whole collaboration because we're also SAP partner how do you see the yeah this collaboration >> when is good to use Google when it's good to use SAP? >> Exactly. Um I I think SAP will will always be needed and strong in in their product area like ERP um CX and so on. And for me it's not uh we do one or we do another it will it will be a good synergy at the end. Um, so we can build very very good working and customer tailored agents on on Google and connect it hopefully with SEP at the end and SAP knows how to interact with their systems.

55:20 >> So um with protocols like A2A um things might have a very very well working synergy in future. >> Yeah, exactly. Because I also think that in the future building agents will not be that difficult. Um, >> in the end, SAP, let's be honest, came into the AI game a bit late, but they're catching up. I mean, I was last year in the SAP Te, they announced all of this, you know, opens aic platform also on their end. So I think as you said through A2A and this collaboration between the agents in the end we'll probably have some kind of a aentic pool of agents across different mediators that we don't even in the end need to do all of these big data pipelines across different uh systems and so on. So maybe let's see maybe the agents will you interact with with each other and just exchange information because in the end you just need something maybe for your agent you still need a textual response.

56:18 Yeah. >> If you instruct the agent to hook him directly to a database or ask another agent in the end nobody cares as long as you get the the final outcome there. >> That that is right. But at the end, how we we need to find, I think, a good way to to talk to our clients and and um provide them a way of of getting into this agentic game. Um what I mean, we we have we have in general like this whole approach depending on where the customer starts. So are there already some kind of a use cases that are defined?

56:57 >> Yeah. or they are completely clueless where to start or maybe they already have a a department or area of business where actually they would like to start with the use cases and we have this kind of a structured workshops where we are really drilling into asking a lot of wise really trying to get this business value out of this first of course focusing on the business problems because I think this is often the case where maybe there is an AI tool maybe there is some kind of a idea but if you don't know the problem that you are trying to solve and if you don't track the data uh how much it is of a big of a problem and if it can be improved by AI you are already like starting without a vision and there is no guarantee that you will actually succeed and on these workshops we have a lot of interesting discussions and how people see these things differently also you know challenging is the data available challenging also if the technology is ready you know sometimes maybe are over creative. This this is where you know maybe the technology is is not ready or there or you know because they still see it as a magic thing and magic box that can do everything what they like. I think this is also a bit the the hype that has been produced around AI. Um and this is also our task to do at the client to to ground them basically guide them through and and um but also be realistic like not everything will be possible but a lot more will be possible now.

58:29 >> Exactly. But it's also funny because sometimes we end up in the thing that for this specific problem we don't need AI at all. >> Yeah. It's simple deterministic flow programmatically that can be done integration within systems and yeah that's