#8 Alexander Oelling (Founder & CEO, INXM) on Building Deterministic AI for Enterprise Processes Reliably

The prevailing narrative in AI suggests that spending more tokens guarantees more intelligence and reliability. But what if the key to dependable enterprise AI is actually spending fewer, or even zero, tokens on repeated tasks?

In this episode, we speak with Alexander Oelling, founder and CEO of INXM, a company building a reliable execution layer for complex business processes.

Alexander shares how INXM’s "Compiled AI" solves the core hurdles of enterprise LLM adoption: hallucinations, high token costs, and a lack of repeatability. He explains how their system initially uses advanced AI agents to learn and master a workflow—such as processing complex SAP invoices—and then compiles those success vectors into a simple, deterministic executable artifact. This approach makes routine enterprise process execution auditable, mathematically correct, and up to 100 times cheaper than relying on continuous model inference.

The conversation then expands into the broader vision of the "autonomous company" and the rise of "invisible AI". Alexander discusses how human roles will shift toward outcome-based work, where employees are valued for asking the right questions and controlling results while AI seamlessly executes tasks in the background. We also explore why the next major leap in AI might not come from massive cloud models, but from highly efficient, localized systems interfacing directly with legacy enterprise software to eventually create digital twins of entire organizations.

If you’re interested in how compiled AI is overcoming the limits of current LLMs, bringing deterministic reliability to the enterprise, and laying the groundwork for autonomous organizations: this episode is worth a listen.

Transcript

[00:00:00] Alexander Oelling: So far, the narrative was clear. You had always to spend more tokens to get more intelligence. And hopefully, it will bring you to more reliability. I think from the test bet we've created, that's not true. You can have more reliability in spending less to zero tokens.

[00:00:33] Robin Harbort: This is Cutting Edge AI brought to you by Angel Invest with your hosts Jens Lapinski and Robin Harbort. Our guest today is Alexander Oelling, founder and CEO of INXM and they are creating a reliable execution layer for business processes. At the core of INXM is its AI process execution engine, a system which orchestrates AI to automate and execute complex enterprise workflows. In this episode, we will explore a new approach to AI that Alex and his teams are pioneering in. Instead of relying on expensive reasoning at every step. The system of INXM initially uses a large amount of compute and tokens to learn how a process should be executed. Once mastered, that knowledge is compiled into a plan and can be executed much simpler and more efficient. The result is process execution that can be up to 100 times cheaper, more reliable, and fully audible for enterprises. INXM just emerged from stealth with a 5.7 million euro Seed round by Cherry Ventures, Redstone, Angel Invest, Linton Capital, and others. Alex has so much more to talk about. The good thing is he also has a podcast, the Compiled AI podcast. So, if you want to know more about Alex as a person and what INXM is doing, I can really recommend the podcast. It's launching soon. But today, this is the Cutting Edge AI podcast by Angel Invest. Let's go. 

Hello Alex, welcome at Cutting Edge AI.

[00:02:17] Alexander Oelling: Thank you. Yeah, great to be here.

[00:02:17] Robin Harbort: And also hello to Jens.

[00:02:18] Jens Lapinski: Likewise, also happy to be here.

[00:02:18] Robin Harbort: So, Alex, you have been a founder already several times. But what are you doing right now exactly?

[00:02:18] Alexander Oelling: We uh created NXM and uh there's an interesting story around that since we um spent the last years in the aerospace industry um building rocket launch software, electric helicopter remote control software and so on and EIP systems and those kind of things. We figured out that uh we need better AI, we need more secure AI, more transparent AI. Yeah. And so higher the practicality of the AI integration and we created INXM to address all of those issues.

[00:02:57] Robin Harbort: For those who who don't know you, can you quickly describe what kind of hardware did you build over the last like five years be before you started in INXM?

[00:03:05] Alexander Oelling: Yeah, we actually we were always building the software for the hardware as you know modern complex products always involve software um to get functional and that's what we did. So I was previously at Isa Aerospace building the or helping to build the spectrum rocket and the launch system around that, and my team was responsible for building the launch control software and the yeah the the entire software you need on a shop floor to produce actually the rocket you whether have in-house developments, or external um software that you have to integrate and roll out over the organization and before that I was um part of Volocopter building an electric helicopter EV tall which was later then sold to a Chinese company. I was helping there building the the full integration suite called Volo IQ which is a kind of an ERP system from an end to end from booking solution for the end customer um to book flights via an electric helicopter to um yeah remote control of the aircraft, airspace integration, crew management, all those kind of things and so it was also a huge project with a a couple of hundred developers involved and it was pre-AI so that was one of the last really big software implementation projects in my opinion.

[00:04:24] Robin Harbort: I think every of those projects you just mentioned would deserve an own podcast, but we want to focus today on NXM and and what you're building in AI next time. Um, but what happened or at which moment did you realize, oh, I want to found my own company. I want to focus on AI right now. What happened?

[00:04:46] Alexander Oelling: I I already created two companies before. Um, the first company was focusing on next generation database systems. Um, the second company was focusing on on property technology, so door lock system and for industrial areas and those kind of things, and after that I've joined um ISA Volocopter and um I think for me I it was clear that I wanted to create one day another company. As I saw that there was actually a gap of mature AI companies that actually deliver and as a cure and sound solutions for the enterprise sector. It was clear that some of the developments we did in-house in the AI sector were pretty useful also for other companies, for other manufacturing companies also but um also for banking and finance sectors and those kind of things where you have a lot of data, a lot of unstructured data that you need to structure. 

It was also clear that a pure LLM is not enough in order to ship an enterprise use case properly, so secure that you can repeat him the process in the enterprise and um get always the same results. And we call this determinism. And when you have an SAP transaction for example about €10 million euros you send from A to B um you rely actually on the perfect outcome right so any mistake cost you 10 million and that's actually where the gap is of the AI adoption in enterprises.

[00:06:17] Robin Harbort: maybe one one simple question in between what does INXM stands for because you described uh this compile AI just just right now but how is where's the connection to INXM?

[00:06:29] Alexander Oelling: yeah so INXM stands for In X machines or intelligence X machina. We played around with it and actually it's pretty hard to to find a short new company name with AI because a lot of AI um URLs are already gone, right? Our idea here is that we deliver a system that brings back the control into the computer of the user. So it's not just in the cloud or a SaaS solution. Um it's actually running on your own hardware if you want in your own stack and in X machines. It stands for in x machine. So it runs in X machines, right? What we've created here is actually not just a wrapper around another LLM. Um we created we created a new form of AI that um is another way of doing reliable AI transactions without um the yeah well-known ways of reinforcement learning for example or other learning methodologies um for large language models.

[00:07:30] Jens Lapinski: INXM is obviously it's enterprise ready AI system. Yeah. Now there are a couple of fundamental insights that you've had where you said look we can't just take an LLM we have to and put it into enterprise and let it run loose because there are a number of real problems with that. Yeah. And the problems that as far as I understand them are number one if the LLM just goes there you know it's intelligent it hallucinates it makes things up it's not repeatable. It's always different. Minor differences, but it's still different. And when it's always running, it's also expensive. Meaning you have always have to do a, you know, burn tokens and and the tokens cost money and eventually the thing cost so much money that that you can't do it. And then the fourth problem is that you're always pulling back to Redmond or to wherever the servers are. They have all sorts of other issues around that because God knows where your data goes and you really don't know what's happening. Yeah. So all sorts of problems that where you say, "Well, I would really love to adopt AI but I've got problems left, right and center and I don't know how to get them done. Can you just explain to the people who are listening how exactly INXM is fixing that because I think what you're doing is like a paradigm shift in AI for enterprise. Yeah. So if you if you explain how it works so that they can they can start to understand how significantly different it is will be super useful.

[00:08:54] Alexander Oelling: Yeah. So the first point here that you you actually mentioned is reliability. So you need reliable um AI transactions. We've created a way in getting or forcing an AI to create repeatable results over time and we call it compiled AI. We are actually tracing the path of a result vector of the result vectors of your answer and um and making actually another form of AI that we call compiled AI that um is able to um yeah to reproduce a certain result. For example, you have a transaction um in SAP or in another enterprise system and you want to repeat that transaction. You actually don't want to have hallucinations, right? You want to have a mathematical correct way. Once said, here's a new invoice. Book that invoice through the ERP system. And in order to get there, you need you can use an AI, you can use several agents, AI agents. Each agent is then creating a lot of tokens and um the results um might be not correct over time. That means you will have hallucinations in each process step and what you want is you want it actually in an AI engine that um once you told the engine how the correct way looks like it does always the same result for you and and tracks through the system clicks through the system always the same path. And we found another way of doing that which is kind of secret source on our side, but um we actually created an agent system that um is able to create pre-compiled compiled AI sub vector space and put it into an actual executable um as you would normally execute on a computer processor. 

And that actually transaction um or this this this actually workflow takes 80 to 100 times less tokens for repeatable transactions than compared to use another agent for each process step. And that's actually something you can quote because we're working on releasing some benchmarks for that also with some well-known players in the industry. 80x tokens on repeatable transactions in enterprise use cases and enterprise processes is actually more than a game changer, right? It changes the the fundamentals of the industry where right now tokens is margin is revenue. That's what Nvidia set on stage in Taiwan, right? So, um right now everyone thinks um I get better results if I spend more tokens. But the reality here is there are technologies out there and we have one of them developed and invented in a form of more or less we I think we are we've created more or less an AI frontier lab for applied AI and we've created a new way um with compile AI to get there and I had an example for that. Right? Imagine you have a robot on the floor like your vacuum robot, right? In the past, you always had to repair this vacuum robot, right? Then agents came and you could say, "Please repair it for me, AI agents, right? Please repair it for me." And the robot will stand up again when it was fallen down. But if you transact that the next time if it's fallen down, you actually spend the same amount of tokens again. But the the paths of success for getting the robot on again online again might be the same all the time. You just turn it around, put it on its feet, and it starts walking or driving again. And that's what we call compiled AI. All of that success vectors in the LLM get be compiled in another executable. And that's what we put in our runtime environment on the robot. And yeah, in the future, the robot not even needs to have an internet connection in order to get on the feeds again, right? Because and we not spending tokens on that, right? And I think that's a very important fact here.

[00:13:13] Jens Lapinski: So basically, you take the LLM, the LLM does something complex, and you say, "Yeah, that actually worked. Yeah, that was a good way of doing that." And this then gets compiled into something that's really, really simple that runs over and over and over again. That's 80 to 100 times cheaper. It's repeatable, therefore it's auditable. Yeah. You can also run it in your own personal environment. You don't have to put it onto the outside. You can also share it inside the organization saying look 10,000 people here is the way of doing that. It's always the same. This is the way that that has been approved internally or whatever. This is the way to to do it. This has been certified secure, safe, reliable and so forth. Yeah. And then effectively what you then do is you have created a really really dumb mini AI system or it's not even a mini AI system. It's it's like a It's like an artifact if you like. Yeah,

[00:14:00] Alexander Oelling: that's correct.

[00:14:00] Jens Lapinski: And from the AI that's like, I think artifact is actually the right word. Yeah. So it's really it just knows how to do one thing which is do this thing pretty much.

[00:14:12]  Alexander Oelling: Yeah. Which actually is not some people and we have some early customers right and there was a procurement guy that asked like oh it's this I don't get it. It's actually the same like an enterprise process automation. He said yeah the difference is you actually prompt your process automation, right? And it's clever. It it heals itself if it's got broken, right? It has also a ton of functionality you will never get with... in other ways, right? Or with an enterprise process automation and um it's actually super fine granular and it's kind of scaled, right? We had a conversation Y couple of months ago where you told me where you you got the point and said, "Oh, wow. You don't need any consultants any more to to set up uh an SAP migration for example or that kind of things, right? Because you actually take things that usually take months for implementation or like you usually in the enterprise you connect a lot of enterprise systems together and you need a lot of consultants in order to do that. And now you can just have one guy that can prompt something and actually it works right. Um you can create an enterprise transaction within seconds. instead of minutes or weeks or months um which usually is attached to hundred thousands of euro spent for external consultants and that's actually just not just a shift for the for the um for the consulting industry but for the entire, yeah AI industry since so far the narrative was clear you had always to spend more tokens to get more intelligence and hopefully it will bring you to more reliability and I think from the test bet we've created that's not true. You can have more reliability in spending less to zero tokens actually. And that's kind of cool, right?

[00:16:11] Jens Lapinski: Yeah. I think it's like a total paradigm shift because you're basically saying you for the inventive step trying to figure out how to do it or if you then it stops working curing that you need to spend money but then repeating it is yeah almost free. I think that and the vast majority of the cost is is effectively so in the same way in which you train an an AI system and then you have inference and you just ask it but in your case the the inference and doing it all over again repeatability that becomes which is the vast majority of the cost

[00:16:46] Alexander Oelling: that becomes almost free I have more things um as said um sometimes I have the feeling we've created an an AI uh laboratory right a frontier lab for for applied AI right um since we've learned so much from our own experience but also um talking to enterprise customers there will be more forms of AI and also the features and functionality. Let me describe a bit the the components you need or which we which we released and first component is of course something we call orchestrator. It's a highly sophisticated AI model and um combined um actually with an agent and um that's actually the AI you are talking to right so that's actually where you prompt your stuff where you prompt your enterprise processes... That's where the creativity happens. That's not so special. It's just an LLM fine-tuned on enterprise processes and a lot of data um that we had from previous projects um but also that are available um through partnerships through some of the larger companies in the industry um that we also in the background did. That's one component. The second component in that is actually we call something that is the Compiled AI itself we call Plans and a Plan is actually a very highly sophisticated JSON uh document which can be very large hundred thousands of project steps. We did that because we wanted something that we can audit that we can trace that we can as a human understand. It's too big usually um to create it manually by hand by someone so you need the agent to create it because otherwise it's too complex right that's where actually LLMs are really good at right now; on the third component you we have the execution engine. So we we we we call the entire new sector that involves here enterprise process execution. So we all know um RPAs, robot process automations, we all know ERP systems, enterprise process tools and so on. And I think that enterprise process execution is actually that what happens. It's actually a new way where AI actually does the work. And that's the the overall meme here. So actually AI that finishes the job and that means you have hundred thousands of invoices per month you need to process, market research, you have procurement deals tons of that stuff that need actually human manpower, but also where human manpower lacks is understanding high complex environments and relationships in data. So you can create here workflows and Plans that then you can even give away to someone else. A plan actually is not that much different to like a file format like it's an executable, right? It's an artifact. It's like a PDF that you can send someone that does the invoicing and the months end planning um on the on the taxation side for example. And if you execute it in the execution engine that we are delivering, it's actually running in the in the context of the user. So, you have actually work thanks to AI, thanks to compiled AI that you can actually deliver to someone else even so it's portable, it's insane there's a ton of that technology stuff that we've developed in the in the past years and we're happy to to share that stuff very soon.

[00:20:19] ens Lapinski: I think it sounds to me as if you think about it what that means is that or at least partially means is that there are the work that humans do is now in adjacent file Yeah, I mean yeah or the JSON file is is is the person who does the that's effectively the description of the work that is being done that gets executed by the by the execution engine, right?

[00:20:44] Alexander Oelling: Look, we as humans have a small mind. We have a small memory, right? We can only remember three to seven things at the same time. It's really hard. If you have a thousand or more, sometimes 10,000 gyra tickets, zero tickets at the end of the month to um kind of close a certain milestone in an engineering company for example, you have to decide what of those um you can put in the backlog which was already fulfilled. How are the requirements on each ticket? It's a huge job. You never have enough people to do that. I have I mean right now maybe 20 different use cases of things like that where not enough employees, humans actually can put their brain in in order to solve a problem in time before a certain deadline. And that is in procurement, that's in engineering, that's in sales. If you have a public offering which you want to attend, and you and only have 10 days time to deliver a certain offer for example, you have to um you know scroll through thousands of pages of requirements. There are a lot of problems which are right now larger than a team of people can solve in a certain amount of time and that's actually the sweet spot for a technology like that in the first place. But I think it has the potential in the on the long run to transform the enterprise at all since once you have technology like that you can have different processes. It also has an impact of course on the workforce on the mid to long run. But you can't do all of that without the um practicality, reliability, transparency of that technology. That was actually the yeah the optimization curve that we've been on since a while now. It's not about creating the best AI. It's about creating the best reliable AI, and the best dynamic harness system around it in order to uh really fit into enterprise use cases.

[00:22:36] Jens Lapinski: Can you give us a couple of use cases that you can talk about publicly?

[00:22:40] Alexander Oelling: I think the use case actually everyone has is um invoice processing. It sounds super stupid. There are tons of solutions out there, but you know what one CIO of a big aerospace company said to me, "Hey, Alex, um what should I tell my internal people when you get over 95% occurrency? on this invoice processing stuff while the best technologies I can buy in the market and we can build oursel above above 60%? And we have a million invoices per months that we are processing and hundreds of people that do manual work on them to put them into SAP and to into the ERP systems right. And I think it's a super easy use case everyone has and it's much better resolvable with Compile AI. Why? So first of all: Traditional machine learning only tries to understand the invoice. That technology here combines and splits the large use case of understanding that invoice into several atomic yeah processing steps each sandboxed in an part of the compiled AI process step means first of all you have to understand the invoice. Okay, there might be 400 different versions or options of that in certain invoice right every invoice number can can be put it on different areas in the invoice whatever it's like in the text understanding the OCR text is a complex thing at all right so there are good LLMs out there that can do that job very well but then you usually get not above 60% currency. What you really want to do is understanding the entire process behind of it every SAP trans transaction every requirements document and every order number, every orders in the past, comparable invoices on each invoice. And you wanted to use not one AI system LLM to do that. You wanted to put each processing step into a separated secure transparent sandbox and make sure that there's no hallucination around that. So means that there's only a limited result vectors acceptable. Right? And that's actually what compiled AI does. It divides the big problem into thousands of little steps, individually, and um executes them very well through. Therefore, you come over this higher currency at the end because now you have literally a whole LLM step, encoded in thousand into thousand simple steps that only have you can force that those simple steps into result vectors. So and at the end of the day, you get a perfect result because it's not just the invoice it also it's also every transaction around that invoice that's understood and can be put into context including the right transaction in the ERP system that it belongs to. Right? And if you think about that you you normally you need a certified SAP uh specialist um to put that invoice into PP it takes 20 to 30 minutes per invoice. Think about that how much time you spend on that stuff. And if there's a new invoice that no one saw before, the system will reason and find a way to solve it. And if it's over a certain likelihood, it will process it automatically or it's under the likelihood, it will escalate it to a user that then does the manual job. But this doesn't stop here because now you you've trained literally the AI on solving that kind of invoices again automatically, right? And the whole ontology creation which was previously on those other systems a manual job is done automatically um for example. So there's a lot of things here um and it's actually a very complex prompt um to get this working but that's only one time and it's literally done in half an hour and it's only one use case, right?

[00:26:56] Jens Lapinski: So let's just assume that there are quite a few of those kind of use cases, use cases and companies and that eventually they will all get resolved um and then a lot of the manual work shifting bits from one screen to the other from one application to the other will get automated. Yeah. If we lift our gaze up beyond that multi- tens of billions of problem and say what's actually behind that and how do you think about that? I think the first time I heard the phrase autonomous company was in 2017 or something like that, when people started talking about it with the machine learning applications that were around at the time. And it felt like that it might potentially be achievable to to get there but it obviously it was a Fata Morgana, that wasn't that wasn't uh wasn't doable and also with the LLMs same story but how do you think about beyond solving these kind of use cases what does the future look like from from where you are sitting?

[00:27:59] Alexander Oelling: I think um that you um we have to talk about the time frame, right? Right now there's still a lot of politics in larger companies, right? There's still um separate um yeah areas of the company which are little kingdoms, right? Where people don't talk to each other properly. Decision processes are very slow. So in the past you've tend to pay people by their experience, right? And the longer someone is in a larger company um the more usually important the guy is or the girl is. And, now and that's actually from talks with with some chief people officers on on very large automotive companies and in the the past months. You pay people for the outcome. It's an outcome based payment means you pay people for two things. You pay people for asking the right questions and delivering the high quality outcomes. And in between there's AI, and AI agents that. You don't have hundreds of agents. You have a couple of specialized agents using enterprise plans with a technology like compiled AI that do reliable and produce deterministic results which repeatable results. You pay someone for still being in charge that those transactions have been done correctly. Right? So asking the right questions means creating the right workflows that are that are running in the context of a certain user that's responsible for it and second controlling those results is actually the task that someone has now in a future corporation and there might be special roles for that. I call it on the third pillar invisible AI since this kind of those kind of systems are not delivered to hundred thousands of people inside a large company most of the processes like this invoice use case as stupid as it sounds right, it's a complex one, but there's not that much people anymore that actually see that process. It's it's kind of indistinguishable for magic as Arthur C. Clark says, right? So, when technology is far enough developed, you can't separate it from magic. So, it runs in the background. And I think it's not the god AI that knows everything and can do everything. Here it's more the Star Trek vision, right? Where you have an AI that runs in the background is always is there for you and does things automatically which we don't have to care anymore about since it's reliable. It's a repeatable, reliable and to be honest stupid process and and boring process like an invoice processing that's then done automatically, right? But it can be also backlog uh solving in in Jira for um for milestones. Um there's so many things, right? Um in procurement um supply negotiations pre-automized um things like that that are now with enterprise process execution yeah possible.

[00:31:07] Robin Harbort: Every time you you speak about a specific process like with a ERP system or in Jira you're always working with the systems which are already in place and so far as I have understood you don't want to replace those systems right?

[00:31:23] Alexander Oelling: yeah

[00:31:23] Robin Harbort: if I assume this right why why did you the decision to do it in that way?

[00:31:29] Alexander Oelling: I think that there is an enterprise uh systems of record and that's actually the systems that are actually there like the ERP systems, BLM systems, MEES systems and so on in production and then there's the systems of action on top which we see right now right so that enterprise process execution is actually a systems of action and it takes actually actions in the systems of record. I think that there is probably hundred thousands of micro-processes we call plans to be implemented first, before you can exchange the systems of record layer underneath. 

So if you have reliable transactions and let's say you have the digital twin of your organization in plants, in compile AI then you can start to exchange ERP systems underneath right? Because then you're literally interfacing to those systems through the systems of action through prompts through okay let me release one other thing we've created it's called the generative UI and the deterministic generative UI. So one of those plans um can create you the interface itself on the fly and it's deterministic. It puts exactly the the right fields in the system underneath under the ERP system. So you are interfacing and getting away as a user from that those kind of systems of record. Um so you probably have never to log into ERP again. So at one point you have the digital twin running of your entire organization then you can start to exchange the systems underneath and it's then more or less a data platform underneath needed. Some of the enterprise we're talking to have already created those data platforms and we'll be so happy soon when I can share the use cases and what we're doing there. It's so cool. I mean some of them really did a great job in creating those data platforms over the past years and now they are really kind of earning the gains right so they they've done the right job the right decisions and now they can implement those kind of AI on top of it and uh and be very successful.

[00:33:31] Robin Harbort: So, but then you're saying like in the in the long term a company could run on the compiled AI of an INXM to store the records and do actions?

[00:33:40] Alexander Oelling: I think you never do an innovation alone, right? There might be other people doing having the same idea at the same time, right? But Compile AI is actually not our term. It's the a term of some researchers in Stanford. There's a paper around that. The fun fact here was that we've created started not writing papers. We've created the system a while ago. And then figured out that there might be um some other people working on the same things, right? 

So I assume there will be more people doing that and um enterprise process execution is definitely another category here. By the way, fun fact, how we get recognized of those? We have a 24/7 running agent, with plans that those plans are understanding every website, every news article, every competitor and put them into to a structured output. We have collected more than 180 gigabytes of of uh pure data and it's not MD files, right? So it's executed in compile AI and it runs 24/7 costs a couple of hundred euros per months. It actually runs through the entire web and it found in a report Compile AI understood what it does and u recommended us to have a look into it because it looks similar to that stuff that we have in our documentation and in our codebase. Yeah, you can build something like that. And it's by the way was just a prompt with two sentences. 

[00:35:08] Jens Lapinski: We typically ask what's cutting edge about what the company is really doing and the most cutting edge. I think in your case sounds like it's like a whole combo of cut after cut after cut. So when we first talked in April last year...

[00:35:25] Alexander Oelling: Yeah.

[00:35:26] Jens Lapinski: Yeah. There was no I mean this was uh a few months before you even started the company. How has your vision since then changed? Like what was this... What was the whether like did you have all of those thoughts already at that point? Yeah. Or did they come over the course of the next three to six months or nine months? How did this actually evolve?

[00:35:45] Alexander Oelling: Of course, most of the credits go to Matthias, our CTO. He's a genius. Um he worked on that stuff since years and um and the team around him. We have a world-class team, which could be also at Meta or at Antrophic or Open AI and those kind of guys and people have so much experience over the years, and also some great researchers. Most of the stuff is on their mind right so I I was talking to them a lot and understanding the things but when we met I think I understood that there will be an AI orchestration, so there will there's an orchestrational layer that kind of can steer enterprise processes and that's when Matthias's ideas with Compile AI um mixed up with I think my ideas for the for the enterprise orchestration. So funny that we created you know the pitch decks from last summer right and then you see for example SAP stepping on the Sapphire conference on stage and announcing the agentic AI enterprise right enterprise AI and it's so funny that you you know have pitch decks that are more than a year old having the same claims in there right and I think it's always you never have alone a certain idea there's always like other companies other people doing the idea... And at the end it's execution I think you came up initially when when I was in the audience we not had created a company and um you had a presentation right at this conference you had some slides in there where you were talking about the enterprise layers right? The UI layer um and the the systems of record and underneath like all the the other systems where you said that there will be players on the interface side and this you know enterprise software layer gets thinner and thinner because it's getting eaten up from the AI systems underneath right. And that matched 100% to to that what we think as well I think over time um and you see that now with Salesforce that are going agentic and changing their entire product um SAS offering in that direction so they're getting smaller and smaller and the new interface actually is prompting its agents you see that also on the laptops now with the sandboxing stuff that Microsoft released on Windows a couple of days ago. So, you get your agent, and there's just the fundamental gap and I I'm not sure why no one wants to really hear that or wants to accept that that spending more tokens will not bring us to more intelligent systems at the end of the day. We are reaching a we've reached a kind of plateau. Even Antrophic was releasing that in a in a a blog article today or yesterday. Okay. And they understood that as well. Everyone understands that. But the enterprise itself won't adopt those technologies if they are not as reliable as mathematical possible. Right? And you only get there if you force clever AI and it's math at the end of the day, right? It's vector uh uh operations that you do, right? And and and there's a mathematical solution which we found in order to do that problem.

[00:39:01] Jens Lapinski: Yes. And I think it's actually logical if you think about it that the asymptotic that that we've now reached a point where the AI if you spend more tokens cannot improve the end result. And this is completely logical because the AI is right now probably at the point where in these kind of environments it acts at the human performance level.

[00:39:21] Alexander Oelling: Yeah.

[00:39:23] Jens Lapinski: Yeah. So therefore if it was 10 times better than the human but it does a human process how can it deliver a better result? It can't because it's just more expensive, right? The system that's been designed for the humans effectively caps the ability of the AI to perform because it can't outperform in in that environment or in that context. It's just not possible. Yeah. So therefore, what you need to do is you say, okay, we're now at this point. So now what we need to do is basically optimize what is happening. Um improve the speed, improve the reliability, transparency, auditability, security or whatever all these other elements and the cost more than anything else. But spend paying more money on that will not make it any better. And I think this is definitely true for the phase of the next few years where where effectively processes done by or workflows done by humans will get partially replaced by AI. Yeah. But then the question is then when it's transcends when it gets to the next level up or or five levels up or whatever and then these systems are no longer made for humans but they're made for the AI and at that point the the the performance comes into play again. And I think it's just very very hard to envision what that would actually look like in real life, right? Because we haven't got anything like that right now.

[00:40:38] Alexander Oelling: Yeah. I think that um we have to accept that um there's a European approach here, right? And no one stops us in doing so, right? Just because of the entire industry thinks more tokens are better um doesn't mean you can think the other way around. And there is always a technology that makes something more efficient. Why did we stop in developing technology? that makes things more efficient. That doesn't make sense, right? So the approach here is use fewer tokens, have more rules, have also things like on prem functionality as serenity and use it for the actual production manufacturing use cases. Make deterministic controlled and more practical, reliable, auditable and by the way when the cloud is the end word of everything, right? And the we are creating here in the cloud the next god it gets faster and faster and faster. Why can I create with some open source models from China actually on my 10k, 10,000 euro hardware stuff here at home right um on prem right with specialized AI hardware? Why can I have similar results like a year ago on the cloud? Because it gets more efficient and there's not the the one answer to that which means the cloud will produce god um at the end or the best AI or AGI name it as you want. But the thing here is we can have some of the stuff pre-produced in executables in plans you have a lot of storage that doesn't that does that stuff does not consume a lot of storage in four terabytes I can put thousands hundred thousands of that plans and you use your local agent that runs locally on your machine using that plans as it uses tools and databases right now to execute things. So it can be everything from reading your email to calendar to more complex things which involve more people. Put other colleagues in the loop, human in the loop, decisions in the loop, um use other enterprise systems and so on. There's no there's nothing that stops a certain plan and doing complex things, right? And it's self healable. If there's a problem, it will heat itself through the agent around. So whatever, it can run on prem because you don't need unlimited tokens in order to do a certain process reliable. And that's actually Maybe also the reason why um Nvidia's for example putting a lot of effort in new processors and and stuff that run on prem that run in a laptop that run in a workstation and so on. There's more out there than the stuff that you can actually see and I'm really happy um to to share now. We were in stealth mode for a couple of months, right? I'm really happy actually that we are now able to to share that technology and over the the couple of next months we release also more and more of that stuff. There's also some PhD guys on our side that are working on new papers and so on. So, it's really cool and um I have to thank you guys for the trust, right, um that you've put into us um in order to make this a reality.

[00:43:44] Jens Lapinski: Yeah, I'm super happy.

[00:43:45] Robin Harbort: Thanks for being here. If you enjoyed this episode, support us by leaving a follow and share the Cutting Edge AI podcast. See you next time.