Full Transcript for E 776 — AI Agents Inside a German Mittelstand Company

times people copy and paste information, 00:02:10.381 --> 00:02:13.126 and how much inefficiency the process contains. 00:02:13.206 --> 00:02:17.520 For example, we had a bakery with around 40 media discontinuities, 00:02:17.600 --> 00:02:20.345 which cost 35 extra hours of work. 00:02:20.425 --> 00:02:26.425 So if you would have thrown AI on top of that you would have automated chaos. 00:02:26.505 --> 00:02:28.858 Then there is this leadership signal. 00:02:28.938 --> 00:02:33.252 For example, when the managing director describes a workflow completely different 00:02:33.332 --> 00:02:38.822 from what the team described 5 minutes earlier, that gap is a real risk. 00:02:38.902 --> 00:02:42.431 So we have to be very cautious of it. 00:02:42.511 --> 00:02:48.002 And that's why we run extensive workshops before writing a single line of code. 00:02:48.082 --> 00:02:49.650 The diagnosis determines everything. 00:02:49.730 --> 00:02:51.130 Diagnosis determines everything. 00:02:51.210 --> 00:02:52.610 That is interesting. 00:02:52.690 --> 00:02:57.004 You have an interesting background because you actually studied sociology before 00:02:57.084 --> 00:02:59.437 moving into technology and product work. 00:02:59.517 --> 00:03:05.007 What did that teach you about organizations that are purely technical founder might miss? 00:03:05.087 --> 00:03:09.009 I would say I don't take requirements at face value. 00:03:09.089 --> 00:03:13.402 When a client says we need that tool that does X, 00:03:13.482 --> 00:03:17.404 or even a lot of conversations with clients start with, 00:03:17.484 --> 00:03:21.405 I need AI to do this and that for me. 00:03:21.485 --> 00:03:25.015 Maybe many technical founders would hear a feature request. 00:03:25.095 --> 00:03:30.977 So I come from systems thinking, and I think that's what sociology has taught me. 00:03:31.057 --> 00:03:36.155 And I hear a symptom and I ask, why do you need that? 00:03:36.235 --> 00:03:38.196 What's the process behind it? 00:03:38.276 --> 00:03:42.198 And understanding the entire system is sociology at its core. 00:03:42.278 --> 00:03:47.376 And for example, with one client, a youth welfare organization, the brief was, 00:03:47.456 --> 00:03:49.809 we need a better documentation tool. 00:03:49.889 --> 00:03:53.418 But when we start looking at the whole picture, 00:03:53.498 --> 00:03:58.988 like how do you actually work and where inefficiency occurs in the entire process, 00:03:59.068 --> 00:04:03.774 we very quickly understood that they don't need a better documentation tool. 00:04:03.854 --> 00:04:08.168 They need an entirely different process of how they interact with 00:04:08.248 --> 00:04:10.601 the software on a daily basis. 00:04:10.681 --> 00:04:16.563 So in the end, we wrote an entire ERP system for them with voice-first interactions. 00:04:16.643 --> 00:04:21.741 and semi-automated processes where documentation is only one smart part of the chain. 00:04:21.821 --> 00:04:25.351 You basically taught

them how to use the tools. 00:04:25.431 --> 00:04:26.831 Okay, I see. 00:04:26.911 --> 00:04:28.479 ERP with voice-first interaction. 00:04:28.559 --> 00:04:30.520 That's also an interesting piece. 00:04:30.600 --> 00:04:33.345 Would love to see that one day. 00:04:33.425 --> 00:04:37.739 In the past, you have argued that AI made your own 00:04:37.819 --> 00:04:42.525 company possible because it lets you operate with a much smaller team. 00:04:42.605 --> 00:04:48.487 What exactly would have required 4 people before and what is now handled by agents? 00:04:48.567 --> 00:04:49.967 Exactly. 00:04:50.047 --> 00:04:55.145 So our situation is that we started with 2 people, with 2 founders, 00:04:55.225 --> 00:05:00.715 and were actually able to go a very long way with only 2 people. 00:05:00.795 --> 00:05:05.109 But I think we as you already said, we really did 00:05:05.189 --> 00:05:08.326 the work of 4 or even more people. 00:05:08.406 --> 00:05:13.896 The first area in which we work closely with agents is the most obvious: 00:05:13.976 --> 00:05:15.376 marketing and content. 00:05:15.456 --> 00:05:18.594 I would say 80% are outsourced to AI. 00:05:18.674 --> 00:05:22.203 Agents hold our strategies, our content plans, our pipeline. 00:05:22.283 --> 00:05:25.028 And I do the Q&A and finalization. 00:05:25.108 --> 00:05:29.422 So then there is this research because we work across industries 00:05:29.502 --> 00:05:33.816 and sometimes we start a project in a new industry that 00:05:33.896 --> 00:05:36.249 we are not yet familiar with. 00:05:36.329 --> 00:05:40.642 So we do deep dives into sectors and client processes before 00:05:40.722 --> 00:05:45.820 we start a first conversation, which AI is great at helping us with. 00:05:45.900 --> 00:05:50.214 And then there is the sales part, obviously, that were able 00:05:50.294 --> 00:05:54.608 to automate quite a big workflow that basically makes sure that 00:05:54.688 --> 00:05:59.002 during the sales process we have quality assurance, every piece of 00:05:59.082 --> 00:06:03.395 data we can gather enters the process so that when this 00:06:03.475 --> 00:06:08.573 lead converts into a project, we do not start the project from zero, 00:06:08.653 --> 00:06:14.536 but we already have, for example, a project outline created by the previous sales agent. 00:06:14.616 --> 00:06:18.537 And then the fourth and most significant area is coding. 00:06:18.617 --> 00:06:22.539 This is where we really have the biggest efficiency gains. 00:06:22.619 --> 00:06:24.580 We have multiple coding agents. 00:06:24.660 --> 00:06:28.973 We have, I would say by now we are replacing an 00:06:29.053 --> 00:06:34.936 entire IT department with a very coherent system of agents and it's not vibe coding. 00:06:35.016 --> 00:06:36.584 Please don't misunderstand me. 00:06:36.664 --> 00:06:41.370 It is an elaborate process designed to ensure consistent, coherent code quality. 00:06:41.450 --> 00:06:46.940 Previously, that would have required maybe one additional person that would somehow do sales, 00:06:47.020 --> 00:06:50.550 research, marketing combined, and at least three additional developers. 00:06:50.630 --> 00:06:54.159 When you talked about preparing for a new engagement, 00:06:54.239 --> 00:07:00.122 you were talking about all the data you add before the client actually says yes. 00:07:00.202 --> 00:07:03.731 I was wondering how, because it's from the outside, 00:07:03.811 --> 00:07:09.301 how do you ensure sufficient quality that you don't have the typical system problem, 00:07:09.381 --> 00:07:10.950 garbage in, garbage out? 00:07:11.030 --> 00:07:12.991 That's a very good point. 00:07:13.071 --> 00:07:17.384 So first of all, there is always a human in the 00:07:17.464 --> 00:07:22.170 loop and I always have to be able to understand the results. 00:07:22.250 --> 00:07:24.211 Otherwise, yeah, that's a risk. 00:07:24.291 --> 00:07:28.605 And then again, it's about setting up a system where you 00:07:28.685 --> 00:07:33.391 lower the risks of the AI producing low-value results or wrong results, 00:07:33.471 --> 00:07:37.784 which you can do by certain steps that you do while 00:07:37.864 --> 00:07:42.178 creating your agent system and not by simply throwing a prompt 00:07:42.258 --> 00:07:45.003 in ChatGPT and hoping for the best. 00:07:45.083 --> 00:07:46.652 Hoping for the best. 00:07:46.732 --> 00:07:51.045 Yeah, I think for some people out there it's still strategy. 00:07:51.125 --> 00:07:55.439 Where is the boundary between a lean AI-enabled company and a 00:07:55.519 --> 00:07:59.049 founder simply hiding the operational fragility behind some automation? 00:07:59.129 --> 00:08:03.050 That's a very relevant question, and, especially at this time. 00:08:03.130 --> 00:08:07.444 So I would say the boundary is where you understand what 00:08:07.524 --> 00:08:13.014 your agents are doing well enough to do it yourself if they stop tomorrow. 00:08:13.094 --> 00:08:15.839 Agents don't replace competence, they replace capacity. 00:08:15.919 --> 00:08:20.233 The judgment, the quality standard, the strategic thinking, this always has 00:08:20.313 --> 00:08:24.235 to stay with the humans, in that case with me. 00:08:24.315 --> 00:08:28.628 And I always think about it it's very tricky because AI 00:08:28.708 --> 00:08:33.022 is so good at producing plausible output that you can fool 00:08:33.102 --> 00:08:38.200 yourself for much longer before you realize that you've actually lost the plot. 00:08:38.280 --> 00:08:42.986 That's also something that we really take care of with our clients, 00:08:43.066 --> 00:08:49.066 with the tools that we create for them, that they always stay in the driver's seat. 00:08:49.146 --> 00:08:52.283 We do a few things to manage that. 00:08:52.363 --> 00:08:57.069 First, as I already said, there's always a human in the loop. 00:08:57.149 --> 00:09:00.678 Second, we document agent architectures just like software architectures. 00:09:00.758 --> 00:09:05.464

It needs to be scalable and not just living inside my head. 00:09:05.544 --> 00:09:09.466 And third, I'm really honest about what AI can't do. 00:09:09.546 --> 00:09:13.860 For example, client relationships, reading the room, judgment calls in whether 00:09:13.940 --> 00:09:16.685 and how a project needs to shift. 00:09:16.765 --> 00:09:21.078 And I would say an indicator of operational fragility is when 00:09:21.158 --> 00:09:25.472 founders or people talk about AI replacing this and that department, 00:09:25.552 --> 00:09:31.552 AI replacing this and that person because they are thinking about AI in terms of replacement. 00:09:31.632 --> 00:09:37.632 I think the discussion that we should be having is AI in the sense of amplification. 00:09:37.712 --> 00:09:42.026 The competence was there first and then AI makes it scale. 00:09:42.106 --> 00:09:46.812 Not only that I have learned that AI agents drift over time. 00:09:46.892 --> 00:09:51.205 So maybe at the start you get a marketing department and 00:09:51.285 --> 00:09:57.285 when you are towards the end of the year, the tool only does social media anymore. 00:09:57.365 --> 00:10:03.248 So I don't think there's right now a case for complete AI replacement for humans. 00:10:03.328 --> 00:10:08.818 So as you said, you always have to have a human in the loop. 00:10:08.898 --> 00:10:13.996 Your public positioning says Amadeni, your company, starts with a workshop, process analysis, 00:10:14.076 --> 00:10:17.605 and then decides whether custom software is actually needed. 00:10:17.685 --> 00:10:23.176 What is your decision rule for saying nope, you should not build software here? 00:10:23.256 --> 00:10:28.354 So it's important to say we always look for a standard solution first, 00:10:28.434 --> 00:10:32.747 because if you have a standard business case, and then there 00:10:32.827 --> 00:10:37.141 is a SaaS product that genuinely fits, then I would say 00:10:37.221 --> 00:10:39.966 that this is the way to go. 00:10:40.046 --> 00:10:43.184 That's part of being an honest IT consultancy. 00:10:43.264 --> 00:10:47.577 But what I've observed in the Mittelstand, especially in Eastern Germany, 00:10:47.657 --> 00:10:51.187 is that businesses that have been established for 20, 00:10:51.267 --> 00:10:55.973 maybe 30 years are the ones that are sitting in a niche. 00:10:56.053 --> 00:10:59.190 And they're at the intersection of 2 industries. 00:10:59.270 --> 00:11:05.152 So what happens quite frequently is that there is no proper software solution for them. 00:11:05.232 --> 00:11:09.546 What they have done until now is they have tried to 00:11:09.626 --> 00:11:13.940 use software maybe from an industry that they're slightly part of. 00:11:14.020 --> 00:11:18.333 But this is creating so much problems than in daily processes. 00:11:18.413 --> 00:11:24.413 That for example, a straightforward case where we go, okay, custom development really is the answer. 00:11:24.493 --> 00:11:28.807 And one signal that is growing stronger and stronger, for example, 00:11:28.887 --> 00:11:33.593 with an engineering firm in Berlin that we are currently working with, 00:11:33.673 --> 00:11:37.987 they were very happy with their SaaS solution, with their sort 00:11:38.067 --> 00:11:42.380 of project management solution, but they had been asking for a 00:11:42.460 --> 00:11:45.205 long time for a very specific feature. 00:11:45.285 --> 00:11:46.854 They didn't get it. 00:11:46.934 --> 00:11:48.895 They didn't get an API. 00:11:48.975 --> 00:11:54.465 So this, I would say, minor issue is really what caused them to say, 00:11:54.545 --> 00:11:58.075 okay, we have to say goodbye to our solution. 00:11:58.155 --> 00:12:02.076 And they proactively decided to go for a custom development. 00:12:02.156 --> 00:12:08.156 Because if a SaaS product doesn't have an API, it means that nobody can add automation. 00:12:08.236 --> 00:12:09.805 There is no flexibility. 00:12:09.885 --> 00:12:13.022 And this is something that surprised even me. 00:12:13.102 --> 00:12:15.847 This is really coming from our clients. 00:12:15.927 --> 00:12:19.456 So the decision tree, I would say, is simple. 00:12:19.536 --> 00:12:22.674 Is there a good standard solution that exists? 00:12:22.754 --> 00:12:24.715 If yes, go for it. 00:12:24.795 --> 00:12:29.500 If no, then custom software is really becoming a very viable alternative. 00:12:29.580 --> 00:12:32.325 One point is especially important for clients. 00:12:32.405 --> 00:12:37.896 If they don't have an API or some way to interact with their data, 00:12:37.976 --> 00:12:40.721 then they're also going for the alternatives. 00:12:40.801 --> 00:12:45.115 Yes, we should always keep in mind it is not only 00:12:45.195 --> 00:12:49.508 a question of standard software or not, but also is the 00:12:49.588 --> 00:12:52.333 standard software breaking the bank or not? 00:12:52.413 --> 00:12:55.943 I think you know what I'm talking about here. 00:12:56.023 --> 00:12:59.944 I've been going through a little bit of your data, 00:13:00.024 --> 00:13:04.338 and in your Impuls KBB case, the reported gain was from 00:13:04.418 --> 00:13:09.908 8 hours to 2 hours for development reports and a 4-week return on investment. 00:13:09.988 --> 00:13:14.302 What had to be true operationally for that result to be 00:13:14.382 --> 00:13:17.127 real rather than just a software demo? 00:13:17.207 --> 00:13:21.129 I really like it that you're looking at the numbers. 00:13:21.209 --> 00:13:23.954 So let's look at the whole picture. 00:13:24.034 --> 00:13:28.347 The four-week ROI and calculated savings of €21,000 per month did 00:13:28.427 --> 00:13:33.525 not come from one process, this efficiency increase from 8 to 2 hours. 00:13:33.605 --> 00:13:37.919 As I mentioned a few questions earlier, went through the entire 00:13:37.999 --> 00:13:42.313 process chain starting with case clarification from the youth welfare office, 00:13:42.393 --> 00:13:45.138 really to the very end of

invoicing. 00:13:45.218 --> 00:13:49.924 And we really looked at every step where we can increase efficiency. 00:13:50.004 --> 00:13:54.318 Across the entire process, there was so much room for improvement 00:13:54.398 --> 00:14:00.280 that in that case were really able to achieve this, I would say, impressive ROI. 00:14:00.360 --> 00:14:04.282 So first the client made a very, I would say, 00:14:04.362 --> 00:14:08.675 bold move at the time because we had already started the 00:14:08.755 --> 00:14:14.245 project and soon after the project began they decided to replace their old software. 00:14:14.325 --> 00:14:17.855 They saw that it did not have an API. 00:14:17.935 --> 00:14:22.249 Certain features in their previous software were causing a lot of 00:14:22.329 --> 00:14:26.642 problems and they decided not to supplement it but replace it. 00:14:26.722 --> 00:14:29.467 And that unlocked the enormous automation potential. 00:14:29.547 --> 00:14:33.861 For example, the voice-first interactions, because these are social welfare workers 00:14:33.941 --> 00:14:37.471 that go to the homes of children or youth. 00:14:37.551 --> 00:14:41.472 And then afterwards they have to sit in their car, 00:14:41.552 --> 00:14:46.258 open the laptop and type up documentation about what they discussed Now, 00:14:46.338 --> 00:14:50.652 on their way back to the car, they can simply make 00:14:50.732 --> 00:14:55.045 a voice message which then gets transcribed and then translated into 00:14:55.125 --> 00:14:59.047 official bureaucratic language and then placed into an automated workflow. 00:14:59.127 --> 00:15:03.833 Second, we built the system for how the social workers actually work. 00:15:03.913 --> 00:15:07.835 As I said in the example of the voice notes, 00:15:07.915 --> 00:15:12.228 They also had a lot of unofficial communication about what was 00:15:12.308 --> 00:15:17.798 the status quo of this document and have you thought about this and that? 00:15:17.878 --> 00:15:22.192 So we took a lot from that informal communication that created 00:15:22.272 --> 00:15:27.762 really a lot of inefficiency and put it into a process within the software, 00:15:27.842 --> 00:15:30.587 which eliminates a lot of wasted time. 00:15:30.667 --> 00:15:34.981 And then third, we integrated it with the rest of the 00:15:35.061 --> 00:15:38.198 system because they had a separate HR system. 00:15:38.278 --> 00:15:42.200 So they had to do 2 types of time tracking. 00:15:42.280 --> 00:15:47.378 And we said, no, the system has to work for the social workers. 00:15:47.458 --> 00:15:49.419 They need only one interface. 00:15:49.499 --> 00:15:53.813 And in the background, we will create a flow of data 00:15:53.893 --> 00:15:56.638 so that they have only one touchpoint. 00:15:56.718 --> 00:16:02.208 And I would say this entire process optimization built or visualized in the tool, 00:16:02.288 --> 00:16:05.033 that's what really created this four-week ROI. 00:16:05.113 --> 00:16:09.427 I had to smile when you said that you make a 00:16:09.507 --> 00:16:13.428 voice memo and then it gets translated into bureaucratic language. 00:16:13.508 --> 00:16:18.998 It's— that's really a good point, a translator between bureaucrats and normal human beings. 00:16:19.078 --> 00:16:21.824 But let's go a little bit further. 00:16:21.904 --> 00:16:25.433 When a client has 20 systems, 40 media discontinuities, 00:16:25.513 --> 00:16:29.042 and 35 hours of avoidable duplicate work per week. 00:16:29.122 --> 00:16:32.652 How do you decide what you will fix first? 00:16:32.732 --> 00:16:37.045 Yeah, so the numbers that you quote come from a case, 00:16:37.125 --> 00:16:40.655 from a bakery that produces its own sourdough goods. 00:16:40.735 --> 00:16:45.441 That's actually important for the process because sourdough makes things very complicated. 00:16:45.521 --> 00:16:50.227 But the CEO called us because they had a bad gut feeling. 00:16:50.307 --> 00:16:54.620 So we did an analysis in a 2-day workshop and this 00:16:54.700 --> 00:17:00.191 bad gut feeling translated into these numbers that are quite shocking or quite surprising, 00:17:00.271 --> 00:17:01.671 I would say. 00:17:01.751 --> 00:17:04.104 And I'm very open about it. 00:17:04.184 --> 00:17:09.282 We were a little bit overwhelmed at first, but what is really important, 00:17:09.362 --> 00:17:13.675 and I think that's great that the CEO actually took that 00:17:13.755 --> 00:17:16.893 step because you don't have to automate everything. 00:17:16.973 --> 00:17:21.286 You don't have to optimize for everything, but you need to 00:17:21.366 --> 00:17:27.249 prioritize and have a plan for how you want to deal with the other things. 00:17:27.329 --> 00:17:30.074 First, we looked for ticking time bombs. 00:17:30.154 --> 00:17:32.115 That's how I call them. 00:17:32.195 --> 00:17:37.293 So something that could potentially go off any day and threaten the business. 00:17:37.373 --> 00:17:41.686 So if you find a ticking time bomb, then that's obviously 00:17:41.766 --> 00:17:44.511 at the very top of your roadmap. 00:17:44.591 --> 00:17:48.513 And then you deal with it regardless of everything else. 00:17:48.593 --> 00:17:52.907 In that case, we did not find any ticking time bombs. 00:17:52.987 --> 00:17:56.516 But then you look at the fundamental structural things. 00:17:56.596 --> 00:18:01.694 In that case, it was actually that they didn't have a leading system. 00:18:01.774 --> 00:18:05.304 which led to a lot of distributed data flows. 00:18:05.384 --> 00:18:09.697 And so it was very clear that we first had to 00:18:09.777 --> 00:18:14.091 build a very sound foundation, like a building, and you don't 00:18:14.171 --> 00:18:16.524 want to build it on sand. 00:18:16.604 --> 00:18:18.957 And then we looked at impact. 00:18:19.037 --> 00:18:23.351 So where did we have the lowest investment or where could 00:18:23.431 --> 00:18:27.352 we create the greatest business impact with the lowest effort? 00:18:27.432

--> 00:18:30.177 Then you have a prioritization, a roadmap. 00:18:30.257 --> 00:18:35.747 And as I said initially, it's totally okay if you don't tackle every process, 00:18:35.827 --> 00:18:39.749 the important difference is that you are aware of them, 00:18:39.829 --> 00:18:42.966 have quantified them and you have a plan. 00:18:43.046 --> 00:18:47.360 And I think that makes a difference between a vague gut 00:18:47.440 --> 00:18:49.793 feeling and having worked through it. 00:18:49.873 --> 00:18:55.873 Yeah, you just have to have a framework built on numbers that you can add to. 00:18:55.953 --> 00:18:59.482 I was wondering, talking about numbers and metrics here, 00:18:59.562 --> 00:19:03.876 what is the metric you trust most when evaluating whether an 00:19:03.956 --> 00:19:06.309 AI automation project is worth building? 00:19:06.389 --> 00:19:11.487 Do you think about hours saved, reduction in errors, adoption rate, revenue impact, 00:19:11.567 --> 00:19:13.136 or maybe management relief? 00:19:13.216 --> 00:19:17.529 I would say I try to collapse all these things that 00:19:17.609 --> 00:19:21.923 you just said into a single number because at the end 00:19:22.003 --> 00:19:26.317 you can put a price tag on nearly all those things. 00:19:26.397 --> 00:19:31.887 What is it costing you right now in terms of time, errors, management firefighting, 00:19:31.967 --> 00:19:34.712 but also what would the solution cost? 00:19:34.792 --> 00:19:39.498 So if the solution has an ROI of less than six months, 00:19:39.578 --> 00:19:43.500 then it's a very straightforward decision to go for it. 00:19:43.580 --> 00:19:48.678 If it is 12 or 18 months, we have a more nuanced conversation, 00:19:48.758 --> 00:19:54.248 and it's also more about strategy, about potential risks also of not acting, right? 00:19:54.328 --> 00:19:59.818 But I would say there is a very big misconception that SMBs frequently have. 00:19:59.898 --> 00:20:04.212 It's a conversation that I have with a lot of especially 00:20:04.292 --> 00:20:09.390 potential clients at the beginning is they treat IT as its own budget. 00:20:09.470 --> 00:20:14.960 So for example, a €25,000 project looks like it would blow the IT budget, 00:20:15.040 --> 00:20:18.177 but and that's the difference from my perspective. 00:20:18.257 --> 00:20:22.571 You have to calculate the IT budget in terms of personnel 00:20:22.651 --> 00:20:26.965 costs of the inefficient process, and then when you compare the 00:20:27.045 --> 00:20:31.358 personnel costs versus the investment into optimizing these processes and the 00:20:31.438 --> 00:20:37.321 potential personnel costs that you save, not because you're letting go of people, but because. 00:20:37.401 --> 00:20:42.891 Suddenly you can come from a very reactive business to a more strategic one, 00:20:42.971 --> 00:20:46.500 or your people can actually do more meaningful things. 00:20:46.580 --> 00:20:51.286 That reframing is one of the most important conversations that we

have, 00:20:51.366 --> 00:20:54.504 and then suddenly the decision-making becomes much easier. 00:20:54.584 --> 00:20:55.984 The decision-making. 00:20:56.064 --> 00:20:57.464 making becomes easier. 00:20:57.544 --> 00:20:59.112 Talk about processes here. 00:20:59.192 --> 00:21:05.075 You argue that a a poor process plus AI becomes a poor process on steroids. 00:21:05.155 --> 00:21:10.645 What is the clearest sign that process must be redesigned before it's even automated? 00:21:10.725 --> 00:21:13.078 There is a very clear sign. 00:21:13.158 --> 00:21:18.256 And it is when nobody can explain the process without naming a person. 00:21:18.336 --> 00:21:22.649 So if people start saying, so I get the invoice and 00:21:22.729 --> 00:21:27.827 then Sabine does this and that and then Thomas does that for example, 00:21:27.907 --> 00:21:31.045 that's not a process, those are personal habits. 00:21:31.125 --> 00:21:35.438 And we don't want to automate people's personal habits because if 00:21:35.518 --> 00:21:41.518 you automate that and then the moment Sabine goes on holiday, the system doesn't work anymore. 00:21:41.598 --> 00:21:45.912 So especially in the Mittelstand, that is a process that we 00:21:45.992 --> 00:21:50.698 go through to get away from this person name-driven thinking to, okay, 00:21:50.778 --> 00:21:52.739 what are the actual steps? 00:21:52.819 --> 00:21:57.525 A second straightforward sign is if information gets typed more than twice. 00:21:57.605 --> 00:21:59.565 A third is, shadow systems. 00:21:59.645 --> 00:22:04.351 Shadow systems—shadow IT—are a very strong indicator that the process isn't working. 00:22:04.431 --> 00:22:08.745 And if you automate that process without integrating the shadow systems 00:22:08.825 --> 00:22:13.139 or without looking at why people had to create shadow systems, 00:22:13.219 --> 00:22:17.140 then you'll create a tool that nobody's going to use. 00:22:17.220 --> 00:22:21.534 I'm talking about private Excel files where the real numbers are, 00:22:21.614 --> 00:22:25.536 or the WhatsApp group that's the actual project management tool. 00:22:25.616 --> 00:22:29.929 So these are indicators where it's very clear that you have 00:22:30.009 --> 00:22:32.362 to redesign first and automate second. 00:22:32.442 --> 00:22:36.756 Yes, I was just smiling talking about shadow IT and WhatsApp 00:22:36.836 --> 00:22:39.973 groups that actually are used for project management. 00:22:40.053 --> 00:22:45.936 I have seen that and in my experience, startups tend to do this more frequently. 00:22:46.016 --> 00:22:50.722 There was a heyday for Slack, but I think it's partly gone. 00:22:50.802 --> 00:22:53.547 WhatsApp has remained popular in the Mittelstand. 00:22:53.627 --> 00:22:56.764 Yeah, maybe it just, it just moved there. 00:22:56.844 --> 00:22:58.244 That's also fine. 00:22:58.324 --> 00:23:03.814 Talk a little bit about more cases that I've seen from you for example, 00:23:03.894 --> 00:23:05.463 the Ostdeutscher Bankenverband case. 00:23:05.543 --> 00:23:09.856 The promise was source traceability and zero hallucination for an AI. 00:23:09.936 --> 00:23:15.819 What architectural choice prevents hallucination by design rather than by, yeah, you know, just hope? 00:23:15.899 --> 00:23:19.820 Yeah, hope and prayers are not something that you want, 00:23:19.900 --> 00:23:23.038 in your system design or in your IT. 00:23:23.118 --> 00:23:27.431 so what we did is, I think the key is obviously 00:23:27.511 --> 00:23:31.825 the key is not literally zero hallucinations, but lowering the risk 00:23:31.905 --> 00:23:35.042 and having a system that can catch hallucinations. 00:23:35.122 --> 00:23:39.436 Because in that case with the Ostdeutscher Bankenverband, they provide information 00:23:39.516 --> 00:23:44.222 to their members that is crucial that their members base decisions on. 00:23:44.302 --> 00:23:49.400 And if you have a hallucinated piece of information, that's reputation damaging, right? 00:23:49.480 --> 00:23:54.578 So what we did is obviously human in the loop at every step. 00:23:54.658 --> 00:24:00.148 Then second, we gave different, unconnected LLMs small tasks, because the smaller the task, 00:24:00.228 --> 00:24:05.326 the more concrete the information, the less the AI is forced to hallucinate. 00:24:05.406 --> 00:24:06.975 When does AI hallucinate? 00:24:07.055 --> 00:24:12.545 When models lack information but still have to come up with an answer, right? 00:24:12.625 --> 00:24:16.939 And so the answer was to cut the task very small 00:24:17.019 --> 00:24:20.940 and to include in the system a traceability of information. 00:24:21.020 --> 00:24:25.334 So every piece of information that I had, I could trace 00:24:25.414 --> 00:24:29.728 back to the original website, to the original source of information. 00:24:29.808 --> 00:24:32.945 And this was especially important for the humans. 00:24:33.025 --> 00:24:38.515 who were in the loop to be able to prove every piece of information. 00:24:38.595 --> 00:24:42.909 I would say that these two or three design decisions lowered 00:24:42.989 --> 00:24:45.342 the risk of hallucinations very much. 00:24:45.422 --> 00:24:49.736 So you give the AI a very small area of responsibility 00:24:49.816 --> 00:24:54.129 at each step and enough information to fulfill it in order 00:24:54.209 --> 00:24:56.562 that they Don't need to hallucinate. 00:24:56.642 --> 00:24:58.042 Great. 00:24:58.122 --> 00:25:01.260 Delia, it was a pleasure talking to you. 00:25:01.340 --> 00:25:06.045 We agreed that we're going to cut this interview into 2 pieces. 00:25:06.125 --> 00:25:10.831 We'll be back with number 2 for you in a few days, 00:25:10.911 --> 00:25:13.264 for us in a few minutes. 00:25:13.344 --> 00:25:14.913 Talk to you soon. 00:25:14.993 --> 00:25:16.393 Thank you. 00:25:16.473 --> 00:25:17.873 Bye. 00:25:17.953 --> 00:25:19.353 Bye.