Latest / The In-Between Tech and Trust Podcast / The Agentic AI Gap: When Tech is Used Before its Architecture is Ready - Anthony Alcaraz, Agentic AI Architect (EP 25)

Transcript

The Agentic AI Gap: When Tech is Used Before its Architecture is Ready - Anthony Alcaraz, Agentic AI Architect (EP 25) · The In-Between Tech and Trust Podcast

• 0:14
  Hello, this is Eva and you're listening to the In Between Tech
• 0:18
  and Trust podcast. And in this week's episode, I
• 0:21
  get quite exciting because we talk about one of the most
• 0:24
  pressing topics in the tech and the ice era.
• 0:27
  And these are AI agents. And if I had to give it a title,
• 0:33
  then it's like the one O 1 of agentic systems for
• 0:37
  organizations. And for that, I have Anthony
• 0:40
  Alcatraz, author of the O'Reilly book on agentic graph rag, a
• 0:45
  senior figure of AWS pushing agentic AI adoption at
• 0:49
  enterprise scale. And he's also a business Angel
• 0:53
  working with early stage startups that are actually
• 0:56
  building, scaling and incorporating layers of the AI
• 1:01
  economy. And so with his background and
• 1:04
  his daily operations, Anthony sits Anon quite unusual in
• 1:09
  between an intersection because he did deep enough in the
• 1:13
  technical architecture to write a book and to co-author it, but
• 1:17
  he's also close enough to the business reality to know where
• 1:22
  organizations, corporations and teams currently are struggling.
• 1:27
  And so in our conversation, we get like so knee deep into what
• 1:32
  actually changes when it becomes more of a systems perspective
• 1:38
  and when it also starts making decisions and taking actions on
• 1:42
  your behalf when agents take over and do the research and
• 1:47
  feed it back into an organizational output.
• 1:50
  And so the tension that we looked into is that a genetic
• 1:53
  systems, if they are already, you know, incorporated into the
• 1:58
  organizational structures in a way that users cannot see them.
• 2:02
  The biggest question that we need to discuss is how we build
• 2:06
  trust into something that is invisible and that is yet to be
• 2:11
  defined and to be educated. Welcome Anthony to the In
• 2:17
  Between Tech and Trust podcast. Thank you for the invite.
• 2:19
  I want to start with you at the intersection of technology and
• 2:23
  trust, and particularly when you think about it and from a
• 2:27
  context of the Gentek AI, but also from your collaboration
• 2:31
  with start-ups, what's the first thing that comes to your mind?
• 2:34
  In term of the technology and trust that is to say especially
• 2:37
  for for I think currently it's a very clear change, OK.
• 2:41
  Before we had system that were mostly reactive, I would say.
• 2:46
  So you, you would ask a question, you would get an
• 2:48
  answer. Now, which change considerably
• 2:51
  with agent CKI, we can't define agent CKI, but but system
• 2:54
  basically that's observe, reason and act.
• 2:57
  So when will you have such system?
• 2:59
  I would say the trust, you need much more trust than before
• 3:03
  because basically you, you are delegating what we used to do as
• 3:07
  I mean, people human to a system that is automated that will make
• 3:12
  that, that which reason for us and that will take decision.
• 3:16
  So if you think about implementing that in regulated
• 3:20
  industries like, I don't know, healthcare or even legal, legal
• 3:24
  system or insurance and that it says you, you need to, to have
• 3:28
  a, a clear trust in what's the system is deciding.
• 3:32
  And I saw to put, I would say the human expertise that is
• 3:35
  still needed at the right place. So it changed considerably.
• 3:39
  I would say the data governance I wanted and even the governance
• 3:43
  I wanted. I think there is a problem
• 3:44
  currently because we don't have, I would say currently there is
• 3:49
  no clear framework or governance that has emerged to manage at
• 3:54
  large scale such system. I mean currently we can
• 3:58
  generate, we can like build, build the system, scale it.
• 4:01
  This is what we are doing at Entropic, open air etcetera.
• 4:05
  But I would say the governance infrastructure of it is not
• 4:09
  quite there. We, we, we struggle for example,
• 4:12
  to secure the security risk etcetera, so.
• 4:14
  Particularly when it comes to building a genetic systems, you
• 4:18
  work with startups, but you also help them scale into like
• 4:22
  environments that then also enable them to grow and to
• 4:25
  transition into this enterprise context.
• 4:28
  What fundamentally changes when AI shifts from a single use case
• 4:33
  and maybe also in an early stage company into a system setup that
• 4:37
  then also need to be robust, need to be trusted and need to
• 4:41
  be able to also be, you know, as flexible as the market evolves,
• 4:45
  but as stable as the company or organization requires?
• 4:49
  Yes, yes, yes. So, so it's complicated, OK,
• 4:53
  frankly to to put into production agents.
• 4:56
  So you need to understand that before we, we had system that
• 4:59
  were stateless, OK. So basically between
• 5:02
  interaction, it didn't have to remember, didn't have to
• 5:05
  forcibly connect with internal data, etcetera.
• 5:08
  Agent 6 needs to be stateful to work.
• 5:11
  So you need to build a a system that's basically maintain
• 5:14
  context. So to maintain context, you need
• 5:16
  to access a context, you need to store it, you need to maybe
• 5:20
  classify it to understand it. On top of that, most of this
• 5:23
  system are building up as some memory capabilities.
• 5:26
  So you need a memory system, you need to manage this memory.
• 5:30
  OK, currently I will cite an example from Munich Cogni, for
• 5:33
  example, this it's a graph databases building memory system
• 5:37
  for for agents. And this agent need also to
• 5:40
  learn from VZ. So basically you need to build
• 5:42
  an evaluation situation system. You need to make sure that the
• 5:45
  agent that you are putting into production are are performing
• 5:48
  well. And to build such evaluation
• 5:50
  system is still is still difficult.
• 5:52
  I will set you one problem with the complexity of evaluating
• 5:56
  agent is that agent are multi step.
• 5:58
  OK? It's not it's not only I give
• 6:00
  you an answer. Is it right or wrong?
• 6:02
  You have multiple dimension in this.
• 6:04
  You need to to to measure as it come with the right step.
• 6:08
  Are these steps secure? Are these steps efficient?
• 6:11
  Did it consider enough source when he when he looked into your
• 6:14
  data, internal databases, etcetera.
• 6:17
  So it's it's much more complex and This is why I think so.
• 6:20
  This is my thesis and this is what I'm writing about with my
• 6:24
  book in my O'Reilly agency graph rag.
• 6:26
  I think graph architecture are pretty useful to manage this
• 6:30
  complexity. But before we like hint there
• 6:33
  and also dive a bit into the context of it, I want to linger
• 6:38
  with the agentic systems because you've now explored on how the
• 6:43
  impact is in applying it. But when we come back to an
• 6:46
  architectural point of view, what is the big shift for
• 6:50
  organizations that they need to do, that they need to apply and
• 6:53
  that they need to evolve into when they wanna translate their
• 6:57
  structures into energetic flow or Yeah.
• 7:02
  This is big OK, because basically what is happening is
• 7:05
  that everything that can be accessed, OK is automobile.
• 7:10
  You can automate everything that can be accessed.
• 7:12
  So enterprises need to think about this accessibility of our
• 7:16
  current databases, tools, etcetera, because everything
• 7:21
  that can be accessed, consumed by agent is like food for it.
• 7:26
  So basically you need to have like a clear strategy about what
• 7:30
  does my agenting system need to access.
• 7:34
  So you need to build like some sort of gateway.
• 7:37
  For example, I know publicly companies like like big banks,
• 7:41
  Morgan Stanley built a huge gateway of all of their tool,
• 7:46
  all of their software are are accessible by their agents.
• 7:48
  There is a a public like an MCP gateway, OK.
• 7:52
  And there is this concept of course of MCP.
• 7:55
  I think we don't need to go in into detail technically, but
• 7:58
  basically MCP it's like a USB key for for agent.
• 8:02
  That is to say it's a protocol that allow agents to access a
• 8:06
  certain tools or a certain piece of software.
• 8:09
  So you can transform your existing software like
• 8:11
  Salesforce, for example, MCP, Salesforce allow an agent to
• 8:14
  access basically Salesforce. So the current opportunities,
• 8:18
  the, the customer notes, etcetera, extra.
• 8:20
  So if you think about it at large scale, you need to map out
• 8:23
  what are your current capabilities, what's your, your,
• 8:26
  your, your, your user, your, your, your, your employee
• 8:29
  accessing their work. To rationalize this into I would
• 8:33
  say a main gateway. And when you think about this,
• 8:36
  there is complexity because you need to manage your identity.
• 8:39
  Maybe when I work in a share, I got access to tools related to a
• 8:44
  share, but I don't want my sales agent to access this, a share
• 8:48
  tool. So you need to build this
• 8:51
  identity layer that is to say preventing agent to accessing
• 8:55
  tools that are not authorized for them.
• 8:57
  You have many complexity around it.
• 8:59
  And for example, so I don't speak as an AWS employee, but
• 9:04
  AWS for example, has deployed agent core OK, that help builder
• 9:08
  build agentic system. And here in agent core, we have
• 9:12
  agent core identity that's manage this credential identity
• 9:15
  layer for for agents. So, yeah, to sum up is access is
• 9:19
  agentic and it creates problem because you have many system in
• 9:24
  enterprises currently that are not accessible by IPI or MCP.
• 9:28
  So you have this integration layer that get that can get
• 9:32
  complexified by by this old system.
• 9:35
  And you will need to have and I think it's already the case and
• 9:39
  the engineering of this old system to be accessible by
• 9:42
  agents if you want to identify such processes.
• 9:46
  And interestingly, you now talk about this unifying or
• 9:49
  interoperability of those layers of integrating it with a tech
• 9:53
  infrastructure that combines the scattered use case that you
• 9:56
  currently find in heritage organizations, right?
• 9:59
  Yeah, I assume that this is a challenge, but what are the
• 10:03
  three biggest challenges that you currently see for
• 10:06
  organizations when they want to deploy a genetic systems or also
• 10:10
  transition from experimenting with AI to fully operationalized
• 10:15
  systematic collaboration? Yeah, yeah, yeah.
• 10:17
  So first one that come into mind and it's it's always the same
• 10:21
  for AI. OK, Well, whatever AI you you
• 10:25
  consider like machine learning or or no modern LLM surgency
• 10:29
  system, it is that architecture of fragmentation and it's data
• 10:33
  quality basically. One problem is that the system
• 10:35
  of record that we built for example, it doesn't contain and
• 10:39
  it was a recent thesis from Foundation Capital in the US
• 10:43
  around context graph. What we are saying is that in
• 10:46
  the current system of record, you don't have the why of the
• 10:49
  decision you have you have the what happened, but you don't
• 10:52
  know all such decision out such data was created.
• 10:55
  If you want to have autonomy agents that make decision, we
• 10:59
  need to access the why, the criteria, the factors that's
• 11:02
  that's that's allow such data to be produced etcetera.
• 11:06
  So this is the first problem. Second problem is that we
• 11:10
  studies demonstrate that agents like particularly for example
• 11:16
  graph architecture, because graph architecture allow to
• 11:18
  connect. So once you have an anthology,
• 11:21
  you you archive rationalize your processes to to the facts.
• 11:25
  To have it within a graph architecture diminish diminish
• 11:29
  complexity to access this data. It's reduced for example the SQL
• 11:33
  query that you would make if you were only to use tabular a beta.
• 11:39
  If you take the example of vector search only, the problem
• 11:42
  with vector search is that not everything can be retrieved with
• 11:46
  only semantic similarity. So basically the first problem
• 11:49
  is that you need to think of that architecture to be agency
• 11:52
  friendly. And this would be my my first
• 11:54
  problem. The second problem is basically
• 11:57
  the governance that I mentioned earlier here, here there is
• 12:00
  many, many thing I, I, I would demonstrate the the governance
• 12:04
  problem with the security risk recently, so there has been an
• 12:07
  explosion of this thing called clobot.
• 12:11
  We know for a fact that there are risk reversions.
• 12:14
  For example, there is this risk called tool poisoning.
• 12:17
  Tool poisoning is is that if see an agent access a tool that is
• 12:22
  online OK and that this tool contain for example malicious
• 12:26
  instruction that are hidden to human OK.
• 12:29
  This could be like instruction that can't be read by human.
• 12:32
  You could basically contaminate the agent, maybe modify his main
• 12:36
  behavior and have like a Trojan horses, that is to say an agent
• 12:40
  that's send your data or modify your system without you notice
• 12:45
  it or access databases etcetera. So basically agents are so
• 12:48
  powerful is that if you contaminate it, if you low
• 12:52
  external threat to act it, the consequences can be really big
• 12:57
  in terms of data exfiltration, data modification, ETC.
• 13:02
  So all of these governance gap that is to say is still
• 13:05
  emerging. So you you have a lot of startup
• 13:08
  in Israel, for example, the Israel Israelian are the best
• 13:11
  around this security risk and in governance I did not mention,
• 13:15
  but you have the evaluation so many of our of our problem.
• 13:19
  And the third because you ask for free is basically So coming
• 13:24
  up to your, your first point is a trust that is to say that is
• 13:27
  to say you need, I think to you need to demonstrate because the
• 13:32
  problem is that this system of course for sure will impact
• 13:37
  current occupation. I estimate myself that maybe
• 13:40
  1995% of knowledge worker job will be impacted.
• 13:44
  What will stay I think is I impact customer meetings this
• 13:48
  kind of stuff, but to say what is intrinsically human for the
• 13:52
  rest all of this, all of that is administrative task that
• 13:56
  analysis etcetera. All, all of the knowledge deep,
• 14:00
  I mean the routine knowledge work or even the deep one, OK
• 14:04
  can either be supported or automated by agentic system.
• 14:07
  So you will need, I saw in term of change management, I think
• 14:12
  rework around the current processes, what people are
• 14:15
  doing, what are their roles and will amplify.
• 14:19
  So I am not pessimistic about this I think.
• 14:21
  I think this system allowed to intensify the work.
• 14:25
  It's allow everyone to do much more, not less.
• 14:28
  But interestingly, I want to just briefly iterate on the
• 14:31
  trust topic that you touched upon because connected to your
• 14:35
  second challenge that you pointed out.
• 14:37
  Yeah, when it comes to the design of them, you explain to
• 14:42
  us that they can also be designed for misbehavior.
• 14:46
  So how would we, as appliers of an agent that has not been
• 14:53
  deployed or built by ourselves, how can we trust them that they
• 14:57
  act with our best intention in mind?
• 15:00
  Is there anything that you could advise us to either look for or
• 15:05
  make sure that once we apply it that we are prepared for the
• 15:09
  next or for the output? Yes.
• 15:10
  So basically there is like basic, I would say security
• 15:14
  tools that can be used to prevent, I would say most of the
• 15:17
  risks. For example, if you take the
• 15:18
  identity one, the fact that your agent is only allowed to access
• 15:23
  your own tools or exercise, it's it's reduce the, I would say the
• 15:27
  blast radius of of adacking basically.
• 15:30
  But you have like tools like for example sandbanks sandboxing,
• 15:34
  because before your agent is maybe doing a query or doing a
• 15:39
  coding activity, what you can do is to 1st sandbox it, observe
• 15:44
  what he's doing and honest is validated by an external system
• 15:48
  can do the actual coding. So there are many, I would say
• 15:51
  technologically, I think there will be more and more solution
• 15:54
  to prevent either the plus reduce of a of a problem or or
• 15:58
  even completely prevented. OK, I will set you an example
• 16:02
  that is this company called the TST, They work.
• 16:05
  They they work currently. So they are was a creator of of
• 16:09
  a library called outlines. It's it's structure output for
• 16:12
  LLMS basically. So it's very simple for everyone
• 16:15
  because I know our listener won't be forcibly technical.
• 16:19
  But what it does is that it forces the LLM to respect
• 16:22
  certain rules, basically when it's watching, when it produce
• 16:26
  prediction and and output. So this in itself, when it is
• 16:30
  placed at sensible points of the agentic system, prevents most of
• 16:35
  the risk because you make sure that the agent is respecting
• 16:38
  your schema, your rules, your procedure.
• 16:41
  And this company, by the way, is working on, on something very
• 16:45
  interesting. They are working basically on
• 16:47
  the control layer for agent. So basically with your question,
• 16:50
  I think trust will come from the emergence of this, this, this
• 16:55
  building up of this control layer for agent that I know many
• 16:59
  startup are are working, working on.
• 17:00
  But for enterprises, I mean, of course you need red teaming.
• 17:05
  So you need a team. So red teaming is 1/2 people
• 17:08
  internally that are testing for bad behavior.
• 17:10
  Basically you need to test for the bad usage of the agent.
• 17:15
  So to work against it basically, of course should you need to
• 17:18
  have that in mind, you start up are are also proposing that to
• 17:22
  to enterprises. But I'm sure many enterprises
• 17:24
  that have that have a strong data science team or data
• 17:27
  engineering team or data engineer team are doing doing
• 17:30
  that by by themselves. Yeah.
• 17:32
  So, yeah. Sorry, because you just explain
• 17:36
  to us the difference between the startups and also the enterprise
• 17:41
  setup. And I would be interested to
• 17:44
  dive a bit into the business Angel aspect of yourself and
• 17:47
  also how you collaborate with startups at early stage.
• 17:50
  Yes. If we look at them from an early
• 17:53
  stage perspective, what would be a differentiator for startups
• 17:57
  that understand how and trust architecture is being set up
• 18:01
  versus those who chase like the hype or also the progress of AI
• 18:06
  and itself? I mean, the best doctor that I
• 18:09
  met basically first, they they understand very, very precisely
• 18:14
  all the current processes are are are working within
• 18:18
  enterprises that are they're trying to improve or or
• 18:21
  disruptive. OK.
• 18:22
  So they they passed most of their first months or time to
• 18:27
  meet to meet basically the business owner of these
• 18:30
  processes. They speak to them.
• 18:32
  They have a clear understanding of all all it is going on
• 18:35
  currently within the auto prices.
• 18:37
  Most of the time one of the Co founder is coming also from
• 18:40
  these processes. So he has he has already a
• 18:42
  network within the auto prices to support his first go to
• 18:47
  market movement, this kind of stuff.
• 18:49
  So this is very important because you need to clearly
• 18:52
  understand, but what is your it's regulation landscape, all
• 18:55
  the processes are what will be the point of integration I saw
• 18:59
  about as mentioned earlier, maybe there will be such
• 19:02
  challenges to integrate with the system currently put in place.
• 19:05
  So you need to all of this inter standing at this background to
• 19:08
  make sure that after all you can have a clear go to market to
• 19:12
  disrupt or improve the current enterprises.
• 19:14
  So I, I have example recently, I mean, I must see this kind of
• 19:18
  start up every week. And when I evaluate them, I have
• 19:22
  this this clearly in mind. It's a social angle, I would
• 19:26
  say. Does this start up as design
• 19:28
  partner or as access to the enterprises they are trying to
• 19:31
  improve all the processes they are trying to improve.
• 19:33
  You can't work in isolation. Basically, what will be valuable
• 19:37
  is access to I I have this example of entropics Entropic
• 19:40
  recently did Nakatan and what is funny about the Sakaton is that
• 19:44
  the third so the three guys that were wanted were not developers,
• 19:48
  they were people. I think it was like a doctor.
• 19:51
  So they were basically the the business expert and and this
• 19:55
  tells something like basically I think the most of the creation
• 19:59
  as a value will come from business experts that were
• 20:02
  backward from virtual and processes virtual and problem
• 20:05
  that have no the mean with agent decoding to code it.
• 20:08
  We'll cut it and we'll come up with a start up ID and try to
• 20:12
  describe the current way things are done.
• 20:16
  So basically. Yeah.
• 20:17
  And now I guess when we look and also come back to your book that
• 20:22
  you've written and you just highlighted about the entrance
• 20:27
  of startups into, you know, markets, but also how they
• 20:30
  approach the orchestration from the infrastructure layer and so
• 20:34
  on. In your book like Agenda Graph
• 20:36
  Rack, you explore also how it advanced retrieval or general
• 20:40
  orchestration architecture is being set up.
• 20:43
  Yeah, writing to you and also of what you've written in the book.
• 20:47
  Why is retrieval architecture so central when you wanna build
• 20:51
  trust in AI systems? So basically if you have like a
• 20:55
  metaphor, LLM are just they are not databases.
• 20:59
  OK, LLM are used when an energetic framework only to
• 21:03
  reason open some data. So because and in most
• 21:07
  industries your data will evolve over time, OK, if you take the
• 21:10
  legal system every month, every week you have like a new
• 21:14
  regulation, you have courts, courts that are that are doing
• 21:18
  their business. So you need a way for agent to
• 21:21
  access freshness in the data. You need a way for agents to
• 21:25
  have an infrastructure that allow them to reason correctly
• 21:28
  on what is happening and I found myself by building such system.
• 21:32
  That's the best way to manage it long term.
• 21:35
  OK is to to build it with graph technology basically why because
• 21:39
  graph technology I lost something that is very powerful.
• 21:43
  It's called the multi op reasoning.
• 21:45
  That is to say you can start with a first inquiry about I
• 21:48
  don't know. So we you have maybe so I need
• 21:52
  to find a customer of my customer, OK.
• 21:56
  And I want to understand the customer of the customer of my
• 22:00
  customer. This is multi op reasoning.
• 22:02
  OK. And you can only do that with
• 22:04
  with rough basically. And agents are really good at
• 22:09
  navigating and all topic has demonstrated with cloud cuts for
• 22:13
  example, are really good at navigating this kind of multi op
• 22:17
  reasoning and trace back basically the information.
• 22:22
  So increase the access accessness, the assetsness of
• 22:27
  the data. I mean, it has been measured
• 22:29
  currently most of the industry is has been has been doing a
• 22:33
  vector search, OK, and only vector search because basically
• 22:37
  you can use a vector such as over graph.
• 22:39
  OK, if you measure so you you put a both by comparison, there
• 22:43
  is a 30% accuracy gap between only using vector search by
• 22:48
  itself or using graph rag solution.
• 22:51
  OK, and 30% is very big. And then so in the book what
• 22:55
  what we explained with my core for what we explained with my
• 22:59
  core for is that there is also this self evolution self
• 23:03
  evolution layer. So cognitive strays on the the
• 23:06
  memory, but memory system are very complex, OK, because when
• 23:10
  you have memory, you need to have a system, but forget on
• 23:13
  Casey information, OK, you need to choose what to remember.
• 23:17
  You need to and make sure that when you remember something it
• 23:22
  is attached to an existing memory etcetera, extra and what
• 23:25
  most of the of the memory builder are doing, so there is
• 23:28
  cognitive, there is memo etcetera.
• 23:30
  They are using graph to do that. OK, It's the only way.
• 23:33
  It is the only data structure that allowed to have such a
• 23:36
  complex external memory to achieve this.
• 23:40
  This basically. And in the industry we talk a
• 23:43
  lot about reducing hallucination or so you know, being precise on
• 23:49
  the output and very accurate on like the explainability of the
• 23:55
  output itself. What role does graph based
• 23:58
  reasoning play in? Yes, yes, yes.
• 24:00
  So it's really good because yes, so you, you increase the
• 24:03
  accuracy of the of the retrieval.
• 24:05
  But if you think logically, even if you retrieve correctly, maybe
• 24:09
  the LLM will interpret it wrongly or do approximation or
• 24:15
  over interpretation, etcetera. But what is great is that
• 24:18
  currently technically you can make sure that the LLM of the
• 24:22
  agents will respect the ontology of the, I would say the semantic
• 24:26
  layer of the graph that has been built.
• 24:28
  So what I am seeing emerging, and This is why I'm writing this
• 24:31
  book, is that the control layer for agents we come mostly and
• 24:35
  from a concrete and organized ontology and graph and to be
• 24:40
  sure that the agent that is interpreting it or using it is
• 24:44
  respecting it. OK at the token level.
• 24:47
  That's to say, currently the problem is that LLM are black
• 24:50
  box. But if you retrieve correctly
• 24:53
  from the graph, you no, no thanks to the possibility of the
• 24:57
  graph. And I mean, it is symbolic.
• 25:00
  So you know which, which nodes has been queried.
• 25:03
  So it is part of the graph that is related to the concept X
• 25:06
  star. So you can have a layer of
• 25:08
  explainability thanks to this graph.
• 25:11
  So you, you get it from SQL research.
• 25:13
  OK, but it's less less efficient and, and, and let's, let's,
• 25:16
  let's powerful relevant graph, but you don't get it with vector
• 25:19
  set. So basically it's increase.
• 25:21
  I saw the explain ability of it increase the traceability of the
• 25:25
  reasoning chain and the explainability of the of the
• 25:28
  graph. You know, while you're speaking
• 25:30
  and while I appreciate if we have this hands down very deep
• 25:34
  technical conversation right now, I'm just like, I, I'm so
• 25:38
  humbled of all the experience. And at the same time, I'm
• 25:41
  currently like, I'm asking myself how much of the knowledge
• 25:45
  that we just discussed and also, you know, translate into how
• 25:48
  it's applied and where it's also brought into the organization.
• 25:52
  How do we ensure that people within organizations feel
• 25:57
  empowered to understand of what we're actually talking about and
• 26:00
  how they also can shape it? Because in my concern that I
• 26:05
  have is that within organization says currently this friction
• 26:10
  between people using infrastructure technology to do
• 26:14
  their job and IT departments are like experts that are dedicated
• 26:18
  to provided. However, with AI and the genetic
• 26:21
  systems, we all come empowered to use do apply and develop the
• 26:26
  infrastructure of how we want to work with it ourselves.
• 26:29
  But a lot of employees do not yet have the knowledge of how to
• 26:33
  do it. So they also ask, you know, the
• 26:35
  experts in the organization to provide them with the respective
• 26:39
  infrastructure, which is like a huge gap that's currently not
• 26:43
  being moderated or also no solved.
• 26:46
  And so I will be interested in first, what's your take on how
• 26:49
  deep of a knowledge to employees going forward need to have an
• 26:52
  order for them to apply it and how to second solve this?
• 26:56
  I don't want to call it limbo, but at least I want to call it
• 26:59
  responsibility loop that is not solved yet.
• 27:02
  I think it's key, OK, what you are saying, I won't pretend to
• 27:06
  have a clear solution. It's a clear problem, OK?
• 27:09
  And it is even worse when you are considering that the fact
• 27:12
  that people that are not technical but have the business,
• 27:16
  business knowledge and the business insight will be going
• 27:20
  forward, the more valuable. Because basically you need this
• 27:23
  business knowledge to create powerful agentic system and and
• 27:28
  the tech guys won't be able to create it without them.
• 27:31
  So basically what you are saying this limbo as you call it, you
• 27:34
  need to solve this to create powerful with the sprayed
• 27:38
  agentic system all over the across the organization.
• 27:41
  So one advice I would have, so there's been this theme with the
• 27:45
  industry around forward deploy engineer.
• 27:47
  What are forward deploy engineer?
• 27:49
  They are engineer that are embedded within the business
• 27:53
  teams to understand what they are doing and to create a
• 27:55
  genetic system on top of it. I think this I would say
• 27:59
  intellectual framework to put tech guys within business units,
• 28:03
  so sales guys, marketing guys, whatever to understand what they
• 28:06
  are doing, maybe educate them, unpower them, OK, when why they
• 28:11
  are doing so is a very powerful idea.
• 28:13
  And I think it's This is why open AI, Entropic and all of
• 28:17
  these guys are basically are basically building up and
• 28:21
  scaling such, such such system. It has been the first deployed
• 28:25
  by Patanti, but it's a right now it's it is, it is more and more
• 28:30
  we desperate. But what is changing now there
• 28:32
  has been the release of cloud cohort by Entropic.
• 28:36
  What people need to understand is that agentic coding, so
• 28:39
  agents that allow you to code are accessible by anyone
• 28:42
  basically because because they are coding, you don't need to
• 28:46
  know I would say. I would say the art of coding
• 28:50
  will disappear, but what will stay is system design.
• 28:54
  Basically, you need to have a clear sense of what you are
• 28:57
  doing basically to, to make make the the best out of it.
• 29:00
  Myself, for example, on my personal usage, I must have, I
• 29:03
  don't know, daily, I must have a 50 agent that are working for me
• 29:06
  to code what I want. OK, I don't code anymore, but I
• 29:09
  understand what I'm doing. I, I can catch errors, I can
• 29:13
  Orient, reorient, I can come up with better ideas.
• 29:17
  So it is a mixture between me and the agents.
• 29:19
  And I think we need to educate, we need to arm power people to
• 29:22
  leverage this technology. Because what I am observing here
• 29:25
  on the field is that this combination of human plus agent
• 29:29
  or coding agent, it creates 5050 time your productivity
• 29:33
  basically, if not more so. So basically we need a clearer
• 29:38
  organization. We need to have a clear program
• 29:40
  of empowerment. Once we have the basis, that is
• 29:43
  to say the basis is, is what we mentioned earlier, you need to
• 29:46
  access. So you need to create gateway
• 29:48
  for the agents to work on your data to work on your.
• 29:52
  Once you have that, we will need to empower people to use it at
• 29:55
  large scale. And this is by the way.
• 29:57
  So what I'm saying is this is what I'm doing, OK?
• 30:00
  I'm pushing for agentic solution, a really structural 1
• 30:04
  and I try, I'm trying to empower people because what people need
• 30:08
  to understand is that this modernization is impacting
• 30:10
  everyone, OK? There is no, not in the world
• 30:13
  maybe except at entropic organization that I'm using it
• 30:16
  at large scale. So yeah, so it will transform.
• 30:19
  I don't want to be a prophet, but the way we know our
• 30:22
  workplace, so how we are doing business will change
• 30:25
  considerably in the coming months.
• 30:27
  OK. And not leaning into the profit
• 30:30
  side of it too much, but still looking in the future, yeah.
• 30:34
  What does enterprise AI look like for you in 2020, let's say
• 30:37
  8 or 9? I mean, whatever long you want
• 30:40
  to look into the future and how do we transition into this multi
• 30:44
  agent ecosystem, organization, collaboration, what do you want
• 30:48
  to call it? And also how specific, like how
• 30:52
  would you envision of how being productive would look like in
• 30:56
  that? Yeah, yeah, yeah.
• 30:58
  So I will give you an image. I think most people will do what
• 31:02
  I'm doing OK every morning you will have like a set of outcome
• 31:06
  coming from your agent because my agent are working by night.
• 31:09
  That's all OK. It's, it's practical and
• 31:10
  basically all of day long you will switch between high impact
• 31:15
  customer meetings because we, we, we will, it will stay OK as
• 31:19
  a fact of going to town, having a nice lunch with your high
• 31:23
  profile customer, etcetera. But what would be cool is that
• 31:26
  everything that is around such high impact meetings will be
• 31:29
  managed by agents. So before this meeting, I will
• 31:32
  have prepped my meeting, your customer meeting with all the
• 31:35
  data that I need. After the meeting, everything
• 31:38
  that is around follow-ups, connecting with design partner,
• 31:42
  connecting internally or externally will be managed by
• 31:45
  agents, Gatner says. So I, I, I fetched up of course
• 31:49
  for this podcast, they said that 50% of work decisions will be
• 31:52
  made autonomously by agency Ki by 2028.
• 31:57
  I think there is one condition is a governance layer is secure
• 32:02
  and we have security etcetera that this number is will be
• 32:05
  considerably more and maybe 50%. But would that end up in agents
• 32:10
  talking to agents? Because if we would now have
• 32:13
  lunch, let's say in Munich or your hometown Paris.
• 32:17
  And then we met up, we had this great conversation and then each
• 32:20
  of us, you know, returns to or the Skype translate it anyways.
• 32:28
  And then my agent would send your agent a follow up of the
• 32:31
  conversation and your agent would answer or what?
• 32:34
  How would that look like? Of course, I mean, yeah, if I
• 32:37
  it's already happening, OK, I can't I can't speak about the
• 32:41
  detail, but myself I already built system where yes, of
• 32:46
  course everything that is follow up understanding you like like
• 32:49
  following the state, OK and updating the state, it's managed
• 32:53
  by agents. So of course, I think everything
• 32:56
  that is related to schedule, everything that is related to
• 32:59
  updating the schedule maybe with new data or coordinating an
• 33:03
  event. For example, recently I work on
• 33:06
  such system, that is to say, maybe matching people with with
• 33:10
  an event, event during an event you have multiple meetings,
• 33:13
  multiple schedule and you want to perfectly personalized the
• 33:16
  invite to the people you want to invite to what is happening
• 33:19
  during the event. All of this can be managed by
• 33:21
  agent. So yeah, and more books we have,
• 33:23
  we had recently an experiment which is called the MUD.
• 33:27
  So there has been a lot of fuse around it.
• 33:29
  That is to say it's a social network for agent etcetera.
• 33:31
  A lot of so a false promise. That is to say most of his
• 33:35
  agents were prompted to do this. OK about the end of the world or
• 33:38
  we will take over extra. So it was funny, OK to watch.
• 33:41
  But what's demonstrate is that you can of course have like it's
• 33:45
  an event based system. That is to say an agent is is
• 33:49
  like subscribed to the to a flow of event and react to the change
• 33:54
  of this event. So you can think all of the
• 33:56
  enterprises system around this. That is to say myself, what I I
• 34:00
  do normally is what I would follow my e-mail and react to
• 34:03
  the e-mail I received by checking up the databases.
• 34:07
  The news can do that perfectly. So basically takes the example
• 34:11
  of emails and following up e-mail.
• 34:14
  Currently, that's the future. Currently all of this stuff can
• 34:18
  be automated. Once we stay, I think we will
• 34:20
  have an explosion of software, OK, basically an explosion of ID
• 34:24
  and an explosion of execution of those IDs.
• 34:26
  Because what's what's this technology empower people is to
• 34:29
  directly go from IDs to execution.
• 34:32
  So what will happen I think is an explosion of of personalized
• 34:37
  software, OK, that's a fit to your use cases, fit to your
• 34:42
  problems, adapt to your situation, etcetera, etcetera.
• 34:45
  If that's the case, doesn't then also the human side of the
• 34:49
  connection and interaction become even more valuable and
• 34:53
  also more appreciated? Yeah, yeah, yeah, yeah.
• 34:56
  I think, I think what will stay is this is I impact, I impact
• 35:01
  customer meeting or I impact partner meetings, but these
• 35:04
  meetings will be so empowered. This meeting will will be like
• 35:08
  the the last step of of a process where we will have
• 35:11
  everything to decide or or do business or exactly.
• 35:15
  So I think this technology will basically arm power to have much
• 35:19
  more quality time gathering with people like Star star.
• 35:23
  I don't navigate anymore on Internet.
• 35:25
  OK, So people will will find it funny.
• 35:27
  It's the only thing that I navigate are social media.
• 35:30
  OK, because you can scrap it. You can't scrap it at large
• 35:33
  scale, but everything that is related to website YouTube news,
• 35:37
  everything can be analysed by agents and and everyday I
• 35:41
  received like a newsletter of what I need to understand what I
• 35:46
  need baby to to look up. I have synthesis across
• 35:48
  analysis, etcetera. All of this time it is gain is
• 35:51
  gain on on on on on being being more outdoor meeting people,
• 35:56
  meeting customer extra. OK, so I think this
• 36:00
  paradoxically and this is my optimistic vision, this
• 36:03
  technology will allow to get more human less, less time, less
• 36:07
  time spent on front of a screen and most time spent with real
• 36:12
  human interaction where you create a trust where where you
• 36:15
  haven't saw all of this. I would say I'm that sick
• 36:18
  understanding everything that AI can do currently and I hope I
• 36:24
  hope won't be as powerful as us, because I think it will be
• 36:28
  really, really complicate to automate what we get.
• 36:32
  The intuition, the personal feeling, the energy feeling
• 36:36
  everything. All of this stuff is very would
• 36:38
  be really really located for the.
• 36:39
  Neuroscience side of. It exactly.
• 36:41
  Since we're now come at the end of our of our episode, I do want
• 36:46
  to like jump into some in between moments that I want to
• 36:49
  have. And the first one that I put
• 36:50
  towards you is what does Gentek AI feel like to you?
• 36:55
  It's funny because so when I was a kid, I used to play a game
• 36:59
  called StarCraft. So it's, it's really geeky.
• 37:02
  But basically it's a game where you are commander and you
• 37:05
  basically you lead units to gather resources to attacks the
• 37:10
  opponent, etcetera. An agent TKI is like this, OK?
• 37:12
  You have like units are managing, you are managing your
• 37:15
  workforce basically, and you are managing your workforce maybe of
• 37:18
  5040 agents. So it's it's it's required from
• 37:24
  you strategic thinking, some sort of management, also
• 37:27
  organization, etcetera. It's feel like this, but it's
• 37:29
  feel like a game. But it's very fun, OK?
• 37:31
  And what does trust in enterprise system require?
• 37:35
  I think transparent failure mods, yes, I think basically you
• 37:39
  need to have a transparency of what is happening and what it
• 37:43
  fail. You need to precisely understand
• 37:45
  why. And what is the biggest risk for
• 37:47
  enterprises? I will be provocative but
• 37:50
  staying at the current stage where you only have human in the
• 37:54
  loop and calling it transformation.
• 37:57
  And what? What should start-ups stop
• 37:58
  doing? Building technology in search of
• 38:01
  a problem. Start by with a problem.
• 38:03
  Well, thank you for pushing ideas and boundaries and also
• 38:08
  moving us into the future. Thank you.
• 38:11
  I appreciate the conversation and all the insights that you
• 38:14
  share. Thank you.
• 38:15
  Thank you, Eva.