Latest / Deep Tech Germany – AI, Robotics & Frontier Innovation by Startuprad.io™ / Agentic AI Employees as a Systems-Design Problem — ADA AI

Transcript

Agentic AI Employees as a Systems-Design Problem — ADA AI · Deep Tech Germany – AI, Robotics & Frontier Innovation by Startuprad.io™

• 0:00
  Welcome to part two. In part one, we uncovered in
• 0:04
  our interview with Oliver, founder of ADA AI, why the
• 0:07
  traditional back office is breaking under the weight of modern startup
• 0:11
  operations and how AI employees are already
• 0:15
  taking over full workflows at companies adopting
• 0:18
  agentic AI. But today we go even deeper
• 0:22
  in part two of this episode. Originally planners one episode,
• 0:26
  but Oliver was. Oliver was giving so good and long
• 0:29
  answers, we spontaneously decided to break it into two.
• 0:33
  In this episode, Dr. Oliver Lugos reveals the
• 0:36
  technical breakthroughs that separates simple chatbots from real
• 0:40
  autonomous agents. The industries that are secretly becoming
• 0:44
  agent ready. And why building an AI workforce might
• 0:47
  be the defining strategic advantage over the next decade.
• 0:51
  If you ever wondered how autonomous an AI should
• 0:55
  be, what God rays matter, or how far you
• 0:59
  can push automation without losing control, this is the episode
• 1:03
  you cannot miss. Let's jump into part two.
• 1:12
  Welcome to Startuprad IO,
• 1:16
  your podcast and YouTube blog covering the German
• 1:19
  startup scene with news, interviews and
• 1:23
  live events.
• 1:27
  Welcome to part two of our conversation with Dr. Oliver
• 1:31
  Lugosch, co founder of ADA AI and
• 1:34
  previously co founder of the Razer Group based in Berlin, the E commerce company
• 1:38
  that scaled beyond 700 million euros in annual
• 1:42
  revenue. In part one, Oliver walked us through the origins
• 1:45
  of ADACOP, the rise of AI employees and
• 1:49
  why Agentic AI is fundamentally different from
• 1:53
  every wave of automation that came before it.
• 1:56
  Today we shift gears. In the second segment, Oliver
• 2:00
  breaks down the technical and operational mechanics that
• 2:04
  allow AI employees to execute multi step
• 2:07
  workflows, make decisions, coordinate across tools and
• 2:11
  operate at levels that were considered impossible just a few years
• 2:15
  ago. We explored the real breakthroughs behind autonomous
• 2:19
  agents. How companies move from one agent to AI
• 2:22
  workforce, would guardrails prevent runaway autonomy,
• 2:25
  emerging industry leading the adoption curve, and
• 2:29
  how agentic AI will reshape high rank robos
• 2:33
  and organizational design. In part one
• 2:36
  showed us why AI employees matter and
• 2:40
  part two shows us what's coming next and how fast.
• 2:44
  Oliver, welcome back. Thank you. Great to be here.
• 2:48
  Totally my pleasure. For everybody listening, it was a few days or even
• 2:52
  longer for us, it has been just five minutes. So
• 2:57
  let us talk about what's the real technical
• 3:00
  breakthrough that makes AI employees
• 3:03
  fundamentally different from chatbots or those
• 3:07
  RPAs? Why we have chosen to
• 3:12
  call them AI employees is because that
• 3:15
  wording should show that
• 3:19
  the usage, the application, goes beyond these very
• 3:23
  singular purposes. What do I mean by that? A
• 3:26
  chatbot is there to chat, Right? It's mostly
• 3:30
  confined to the exchange of communication. Right?
• 3:34
  You can have small LLM helpers that for
• 3:38
  example, summarize documents. Right. And give you
• 3:41
  quick notes around it. But the thought about AI
• 3:45
  employees is that it can go much, much further, it can
• 3:49
  go beyond. It can actually act on a longer process
• 3:52
  scale, so to speak, and connect many different dots.
• 3:57
  And what you need for that is many different
• 4:00
  things. You need the capability of using LLMs in
• 4:04
  exactly the right way, that they produce the outcome that you want
• 4:07
  human like communication. You need integrations into many
• 4:11
  different systems. You need an integration into your communication system,
• 4:15
  into your Outlook or whatever email provider you use,
• 4:19
  into your maybe direct messaging service. You need integration
• 4:22
  integrations into your system of records, into your erp, your CRM,
• 4:26
  you know, for the logistics folks out there, into the tms, the wms,
• 4:31
  you need a human interface, you need a front
• 4:34
  end that the team can actually use, right? And
• 4:38
  there are all of these different elements that you need to combine in order
• 4:42
  to really think an entire process. Because only if you
• 4:45
  think an entire process, you can also have
• 4:49
  impact that is really sizable, that is really substantial.
• 4:53
  And it's more than a little helper here, a little
• 4:57
  helper there, where yes, it might be interesting,
• 5:01
  it might be helpful in everyday life, but it doesn't create
• 5:04
  this measurable meaningful impact for organization
• 5:08
  that truly, you know, gives teams the room to
• 5:11
  breathe and feel like, wow, something has really changed.
• 5:15
  And because you can combine all of these different things now, and
• 5:19
  because you can tie all of these ties together, you
• 5:23
  can think about a broader automation and more impact
• 5:27
  that creates benefit for the organization.
• 5:34
  As AI grows quickly,
• 5:38
  what patterns are you noticing in the demand
• 5:41
  of your customers? Which industries are or
• 5:46
  which processes are agent ready? Now,
• 5:50
  as of today, the pattern that. Emerges is
• 5:53
  that companies often have one or
• 5:57
  two use cases that is on top of everybody's minds because
• 6:01
  it seems to be present, it seems to be that people discuss these
• 6:06
  and we often start there. And they
• 6:10
  most frequently lie in the procurement process,
• 6:14
  sometimes in the distribution process, but often in procurement.
• 6:18
  And then when you go deeper and you start to work on the project and
• 6:21
  you start to build, then more and more people hear
• 6:25
  about, you know, what is possible and what can be done nowadays
• 6:29
  with the help of LLMs and other technology. And then
• 6:33
  more ideas emerge and people start to see more use cases
• 6:37
  from different parts of the organization and start to discuss
• 6:41
  and also initiate further project, further
• 6:44
  use cases across the org. So I would say
• 6:48
  classically your procure to pay and your order to cash
• 6:52
  flows are agent ready or automation
• 6:55
  ready? Really? I think the word AI
• 6:59
  agent is something that I find, I
• 7:02
  don't know, difficult in how to interpret it. Right.
• 7:06
  What does an agent Actually do. What does an agent mean? I feel
• 7:10
  like it's a word that everybody uses. But what is really behind
• 7:14
  that, behind that curtain? What, what, what is an agent?
• 7:18
  I would bet if you ask 10 people, you get 12 answers.
• 7:25
  That always reminds me, in the, in the mid-1990s,
• 7:29
  everybody was trying to sell multi multimedia
• 7:33
  computers and they had experts from like five new
• 7:36
  newspapers doing the elevator pitch and everybody talked about something
• 7:40
  different. Yeah, I fully, fully agree. Right. And
• 7:44
  by the way, I'm not claiming that AI employee is the best wording, right. In
• 7:47
  any way. I'm not claiming that. I'm just saying that, you know, these AI agents
• 7:51
  have, have popped up and I find it difficult to grasp what is really,
• 7:55
  really meant by it. In, in any case, I think
• 8:00
  these use cases pop up throughout the classical again, procure
• 8:04
  to pay, order to cash cycles also on the money flow,
• 8:08
  right. Speaking about finance, order invoice processing,
• 8:12
  issuing of invoices, reconciliation accounting, also these
• 8:15
  processes are ready. What you shouldn't forget
• 8:19
  is that this human in the loop process is something that
• 8:23
  you typically start from, right? If you want to automate a process,
• 8:27
  do it, but still have a human confirm certain actions,
• 8:30
  confirm certain communication that goes out. So you can already
• 8:36
  get 80% of the efficiency of
• 8:39
  the time saving, maybe even 90%. But you still have that
• 8:43
  component of a human double checking and ultimately signing off
• 8:47
  on certain actions. And that is typically the starting point. And then when you
• 8:51
  want to move to full automation, get these last 20 to
• 8:54
  10% of, of, you know,
• 8:58
  capacity of efficiency. That's what you should do under certain
• 9:01
  circumstances, right? If you really have the trust, if you really have seen
• 9:05
  a lot of positive examples, a lot of perfect
• 9:09
  drafts and proposals from the AI, that's when you can move to
• 9:13
  full automation. But that's a very, very organic process.
• 9:16
  That's a very trust driven process that just comes over
• 9:20
  time and again. Even if you are stuck with these, you know,
• 9:24
  human in the loop processes, you still have most of the value
• 9:28
  that will, that you will get ultimately.
• 9:36
  We've been always talking about the employees who get rid of very boring
• 9:39
  stuff. We've been talking about the entrepreneurs who get more
• 9:43
  efficient with a few employees. But where are the
• 9:46
  customer wins? What's the most surprising outcome a
• 9:50
  customer has seen after deploying an AI
• 9:54
  employee? The first thing that customers think about
• 9:57
  is the efficiency, right? And saving capacity,
• 10:01
  saving resources. But what also comes out
• 10:05
  of, of all of these projects is
• 10:09
  an improved speed of processing, data processing
• 10:13
  input, responding to customers, responding to suppliers,
• 10:17
  reacting to any kind of signal that comes in really. So
• 10:20
  you get rid of this backlog, right, that often piles up and it
• 10:24
  can again slow down other parts of the company. Other
• 10:28
  teams will notice, will notice that
• 10:31
  processes run quicker and more reliably
• 10:35
  with the use of these automations. And the
• 10:38
  quality of the output data is typically much better.
• 10:42
  If anything, it's certainly of higher consistency,
• 10:46
  right? Because if you use an AI based automation,
• 10:49
  there's consistency of how things are interpreted, of how things
• 10:53
  are put into a certain place in the system, for example, and
• 10:57
  that consistency drives usability of that data
• 11:00
  because, you know, it's no longer, I would say a human spread
• 11:04
  of how things are being done, but it's typically very clean and
• 11:08
  very standardized. And that is
• 11:11
  a benefit that many companies don't foresee, but
• 11:15
  that they experience as they use the AI employees because
• 11:19
  they make the data more consistent and more usable going
• 11:23
  forward. Also for other teams.
• 11:28
  I see you do that because you want to scale.
• 11:31
  But what are some scaling challenges with
• 11:35
  AI employees? What's the most
• 11:37
  underestimated challenge that companies, your
• 11:41
  customers face when scaling from one agent
• 11:45
  to an AI workforce, to many agents?
• 11:49
  When we think about these automations and changing
• 11:52
  processes in companies, ultimately that's what we're doing, we're changing processes.
• 11:56
  Processes. Then you should never underestimate the human
• 11:59
  factor, right? You will always have those teams that are
• 12:03
  responsible for the processes, that oversee the processes, that
• 12:06
  further develop and advance the processes. And
• 12:10
  you cannot go faster than the team can go,
• 12:14
  right? You always have to take everybody along and make sure
• 12:18
  that these changes happen at a pace that is digestible
• 12:22
  for the organization. And so the limitation
• 12:26
  in scaling is really not a technical limitation.
• 12:30
  It is not a limitation of, you know, being able to produce these
• 12:33
  automations. It's a limitation of how quickly can
• 12:37
  teams and organizations adopt to these changes. And I think that's a very,
• 12:41
  very natural limitation. And it's something that you
• 12:44
  should always keep in mind when you think about how quickly do I want to
• 12:47
  scale from 1 to 2 to 5 to 10 use
• 12:51
  cases or AI employees.
• 12:55
  I see in the founders world we will be talking about
• 12:59
  what's the workflow. No founder realizes can be
• 13:02
  automated, but every founder should automate using
• 13:06
  AI employee. Let's do a little bit
• 13:10
  outlook into the future. How do you see the balance
• 13:14
  between human teams and AI employees
• 13:18
  evolving over let's say the next five years? I know
• 13:22
  it's pretty, pretty early, but givens a positive vision,
• 13:27
  I think this. View is pretty clear.
• 13:31
  I think there will be less and less and less of these
• 13:35
  painful Manual things that you have to do, the repetitive ones.
• 13:39
  I think we will spend more time on really figuring out stuff,
• 13:42
  figuring out how to do things and then having your AI
• 13:46
  employee, your assistant, your agent, whatever you want to call it, take
• 13:50
  over what you have defined. It
• 13:54
  will also help us become more creative.
• 13:57
  You can use LLMs to get new ideas, get new
• 14:01
  impulses and then turn them into great ideas, turn them into great
• 14:04
  execution and really move your business,
• 14:08
  your startup, to the next level. So I think you will
• 14:12
  have less of this boring, painful piece and you will have
• 14:15
  more of the creativity. Do things differently
• 14:19
  and, you know, come into a state where it's, it's, it's
• 14:23
  exciting to advance your business because you no longer have to take
• 14:27
  care of the very, very basic stuff.
• 14:31
  I was just wondering about AI employees. The
• 14:35
  very, very boring stuff. I was wondering, would it be a dream
• 14:39
  for you to automate or start
• 14:42
  automation? German tax authorities but
• 14:46
  because I do believe there will be quite a lot of potential
• 14:50
  and I also do believe there are a lot of people who
• 14:54
  have jobs that are very repetitive and I'm sure
• 14:58
  they would love to get rid of at least those steps
• 15:01
  of the work. I'm not a
• 15:05
  tax expert, but what I can say as, you know,
• 15:09
  being subject to tax, I fully agree with you. I think
• 15:13
  there are probably many, many process steps in, you know,
• 15:16
  filing taxes, anything around taxes that
• 15:20
  could be at least simplified or
• 15:23
  accelerated with AI and LLMs. And
• 15:27
  I think there is a huge space that is to be tackled in that area.
• 15:31
  I fully agree now around the regulation
• 15:35
  aspect of it. I cannot speak to it, but I'm sure
• 15:39
  we'll figure this out and make sure that, you know,
• 15:43
  we can also use AI in these, in these topics
• 15:47
  and these areas of businesses agree.
• 15:52
  Going back into the, from government into the
• 15:55
  real industries. We've been talking about
• 16:01
  AI agents helping employees to
• 16:04
  get rid of boring work. But I was
• 16:08
  wondering, how do you imagine
• 16:12
  agentic AI reshaping the org
• 16:15
  charts of larger companies? I think
• 16:19
  AI will start to become part of the org charts.
• 16:23
  Right. Ultimately, that's where we're headed.
• 16:27
  And what does it resemble? Org charts basically resemble
• 16:31
  or mirror, you know, certain resources
• 16:34
  capacities. I'm sorry, I'm not trying to be, you know, inhumane or anything.
• 16:38
  I'm just trying to abstract it. Right. What is mirrored in an
• 16:42
  org chart, it is, you know, human resources
• 16:45
  executing certain processes, certain tasks, taking over certain
• 16:48
  responsibilities. As we see AI applications take
• 16:52
  over that role, they move into that capacity. Right. And
• 16:56
  you will see that a person will be completely
• 16:59
  occupied with running many, many, many
• 17:03
  operational or a large volume of operations through
• 17:06
  the help of AI, AI employees, agents, whatever you want to call them,
• 17:10
  and being that person that keeps an eye on all of
• 17:14
  these processes, right, keeps an eye on quality, further
• 17:18
  advances the process, further advances the model, the
• 17:21
  application. And really being able to
• 17:25
  have such a huge leverage, right? One person, imagine one
• 17:29
  person handling, I don't know, thousands of customer requests every day.
• 17:32
  Why? Because they have LLMs and AI at their fingertips,
• 17:36
  right? And they can do this at a speed and a
• 17:40
  quality that was unprecedented, that you could never have had with
• 17:44
  just humans. Because you can further improve
• 17:47
  these algorithms and these automations more and more and more. And
• 17:51
  I think that operational leverage that you can have will result in org
• 17:55
  charts looking very different from today. Do you know
• 17:59
  when you've been talking about the difference in org
• 18:02
  charts, what do you say triggered one idea in
• 18:06
  my mind because when I was working in a lot
• 18:10
  of different companies as a consultant, if you're a good consultant, you get
• 18:14
  invited to the Christmas party of the company and
• 18:17
  then you do have the retirees, the people who come in, they
• 18:21
  retired and they talk about what was work like 5, 10,
• 18:25
  20 years ago. And I was wondering if at one point
• 18:29
  you'll have like one room in a company. I
• 18:32
  had in mind a lot of old screens where you can basically talk to old
• 18:35
  AI agents and, and pick their brains
• 18:39
  on knowledge from the past or how it was done in the past.
• 18:43
  Why not? I think it would be enriching, right, if you could tap into
• 18:47
  that knowledge. It's still fresh because it will always be fresh, right? It
• 18:51
  will remember and we'll be able to help us and learn, right.
• 18:55
  Obviously you will never be able to fully take over
• 18:58
  that human element, right? You will always have the core
• 19:02
  team, will be humans at your Christmas party and you will probably
• 19:06
  not celebrate with AI employees. I don't know, maybe we do,
• 19:09
  I assume we will not. But having that,
• 19:13
  you know, this, these capacities of remembering stuff very
• 19:17
  clearly giving you insights, you can ask any questions it will
• 19:20
  hopefully answer in the correct way, in a true way.
• 19:25
  I think this is a value add that we are just starting
• 19:28
  to understand, just starting to grasp really
• 19:32
  what I. Had in mind when I was talking about this was having a digital
• 19:35
  room, not necessarily to be entered during Christmas party. We all know that
• 19:39
  can get out of hands. A digital room where can
• 19:43
  basically look into different screens and have
• 19:46
  all the old AI agent archived there, which
• 19:50
  of course will be a big job for knowledge management. I'm sure you're working on
• 19:54
  IDs for that as well.
• 19:57
  Let's pick your brain a little bit about a contrarian view.
• 20:01
  What is one unpopular truth about autonomous
• 20:05
  agents that nobody in the AI industry wants to
• 20:08
  admit? I think it comes down to this
• 20:12
  topic of autonomy that you mentioned before. I
• 20:16
  think the unpopular truth is that we are
• 20:19
  still far away from this true, true fundamental
• 20:23
  autonomy of really running human processes.
• 20:27
  People sometimes claim and continue to claim that we are there
• 20:31
  already. I don't see it. I have not seen it in
• 20:34
  application and I think it's still a little bit of
• 20:38
  a way to go to really have that live
• 20:42
  and ready and in production.
• 20:47
  For our founders out there, for our audience. DM us on
• 20:50
  LinkedIn, me, your manager.
• 20:54
  What would be the one process you'd
• 20:57
  automate first? We'll share the most creative ones
• 21:01
  anonymously.
• 21:05
  Looking forward to ideas there. Let's get a
• 21:08
  little bit towards the end and this pencil a little bit at end. The
• 21:12
  advice, what's
• 21:16
  your advice to early stage founders who want to build
• 21:20
  AI ready operational structures when they're
• 21:24
  starting out like today or.
• 21:27
  Well, Today is late November 2025, but this will
• 21:31
  air in January 2026. A lot of activity is going on in
• 21:35
  January. Everybody wants to start something new. What advice would you give
• 21:38
  to entrepreneurs now thinking about setting up the company?
• 21:44
  Great question. I think yes, it should be
• 21:48
  AI native, I think because not starting something
• 21:52
  AI native or with AI in mind would be
• 21:55
  foolish given all of the technology that we have and the
• 21:58
  capabilities the technology brings. But I won't, I wouldn't
• 22:02
  overdo it. I wouldn't say, you know, first day you start
• 22:06
  thinking about AI. No, first day think what is your business? What do you want
• 22:09
  to do? What's the value that you will bring to your
• 22:13
  customers, to, you know, society, to the world. And since post
• 22:17
  corona, how do you make money? What's your unit?
• 22:20
  Economics, your mix of scale. And then think about
• 22:24
  how I could raise that. That would
• 22:28
  be something I will be thinking about. How about you, Oliver? Exactly 100%.
• 22:31
  So first start with, you know, what are you doing? How do you create app
• 22:34
  value? How do you monetize? All of these are very, very important questions. And
• 22:38
  once you figure that out, or once you have an hypothesis and you want to
• 22:42
  test it, use AI as a tool, right? Use LLMs to then
• 22:45
  scale yourself, scale your time. But
• 22:50
  if you're not a business, you know that, that,
• 22:54
  that sells AI where AI is really the core product.
• 22:58
  Don't start with AI the very first day. Start with the fundamentals of the
• 23:02
  business. I think this is true. This will always be true. And once you've started
• 23:05
  to figure out things and you have your hypotheses or something that's even
• 23:09
  proven, use AI to scale. Use AI to
• 23:13
  create that leverage that you want to have.
• 23:20
  I'm curious for one thing, because we go
• 23:23
  into the last standard question, and then we have the two mandatory. But
• 23:27
  before we go into that, can you see a future
• 23:31
  where the AI employees are targeted
• 23:35
  by hackers and used for nefarious purposes?
• 23:40
  I think that time is already here. I'm very sure that
• 23:43
  hackers are already trying to attack LLM systems,
• 23:47
  AI systems. And it's something that you have to stay on
• 23:51
  top of and that you have to build
• 23:54
  countermeasures against. So I think that's absolutely critical.
• 23:58
  Whenever you design any kind of AI system that you have that in mind,
• 24:02
  people will always try to get to your system. They will
• 24:05
  always try to hack you. So security is
• 24:09
  number one priority, particularly if you think in broad applications.
• 24:13
  Right. For a small, small company, small startup, they might not
• 24:17
  be the first target, but when you move to larger scales, then you're
• 24:20
  certainly targeted day in, day out. And that must be very core
• 24:24
  to how you approach the entire thing. I
• 24:28
  see, I see. And the
• 24:31
  last question I prepared for this interview, if every company
• 24:35
  in the world would deploy AI employees tomorrow,
• 24:39
  what do you guess would be the one that breaks first?
• 24:44
  Like what function? What capability?
• 24:48
  What would break first? Great question,
• 24:52
  great question. I mean, looking at what LLMs are great
• 24:55
  at, I would assume some of the numerical stuff would
• 24:59
  probably break. I don't know, maybe the forecasting piece or
• 25:03
  the optimization piece might be something that is
• 25:06
  most fragile because, you know, LLMs,
• 25:10
  large language models, are not designed for
• 25:14
  numerical analytical approaches. But that's just my guess.
• 25:18
  I hope we will not see that happen. But if I had to guess that,
• 25:21
  that would be my take.
• 25:25
  Our standard questions are pretty, pretty normal. Are you open to talk
• 25:29
  to new investors? We are always open to talk to new
• 25:33
  investors. We are not actively fundraising, but we are always
• 25:36
  open to talk. And
• 25:40
  this is a funny question for you, because are you also open
• 25:44
  to look for talented employees that you cannot yet,
• 25:49
  that you cannot completely replace with AI employees?
• 25:53
  We are also looking for great people. Absolutely. At any point in
• 25:57
  time. So, yes, we are hiring. So
• 26:01
  basically, we'll link down here your career website and your personal
• 26:04
  LinkedIn profile so investors can reach out to you. Perfect. Let's do that.
• 26:09
  Oliver, thank you very much. Was such a pleasure to have you here twice.
• 26:13
  Thank you. Thank you, Johan. It was a real pleasure. Was great talking to
• 26:16
  you. Same here. Have a good day. Bye bye. Thanks.
• 26:25
  That's all folks. Find more news streams,
• 26:29
  events and
• 26:30
  interviews@www.startuprad.IO.
• 26:35
  remember, sharing is car.