Latest / The Yield Doctor with James Deaker / What Happens When an Ad Tech Company Goes All-In on AI? | CEO Nicolas Schueller

Transcript

What Happens When an Ad Tech Company Goes All-In on AI? | CEO Nicolas Schueller · The Yield Doctor with James Deaker

• 0:00
  I am James Dika. I am the Yield Doctor,
• 0:03
  and today's discussion is part of a
• 0:05
  project that I am embarking on to talk
• 0:08
  to companies in the digital media and ad
• 0:11
  tech space that are driving all of the
• 0:15
  change in the AI landscape in
• 0:18
  advertising. And specifically, those who
• 0:21
  are doing it to drive change for
• 0:23
  publishers or other ad tech providers.
• 0:26
  And I'm very transparent about the fact
• 0:28
  that half of the reason I'm doing this
• 0:30
  is to make sure that I am well educated,
• 0:33
  but I also have a belief that a lot of
• 0:36
  other people have the same desire to
• 0:38
  learn about these companies in the
• 0:40
  space. Maybe they've heard of company,
• 0:42
  maybe they've seen it in the Lumascape,
• 0:44
  or maybe they did what know only what
• 0:46
  the company did 5 years ago. And so
• 0:49
  today, with that in mind, I'm talking to
• 0:51
  Nicolas Schueller, who goes by Nico. He
• 0:55
  is the CEO and co-founder of Adomik, and
• 0:59
  Adomik was founded in 2012 and is
• 1:03
  headquartered in France, but they have
• 1:05
  significant operations around the world,
• 1:08
  and especially in the United States.
• 1:10
  Nico, thank you so much for taking the
• 1:13
  time to talk today.
• 1:14
  >> Hi James. Um, thank you so much for
• 1:17
  having me.
• 1:19
  Uh, when accepting your invitation, I
• 1:21
  didn't realize that would be part of
• 1:23
  your own education, so I'm very honored,
• 1:26
  and I'll try to do my best.
• 1:29
  >> Brilliant. So, let's uh let's start at
• 1:31
  the beginning. What is the origin story
• 1:34
  of Adomik, and how is it evolved?
• 1:37
  >> Yeah, I mean, as you mentioned, you
• 1:39
  know, uh we started back in 2012, you
• 1:42
  know, the the classic, uh you know, two
• 1:45
  friends in a garage story. That was us.
• 1:48
  Uh
• 1:49
  literally.
• 1:50
  Um and we started uh
• 1:53
  with the uh idea that um
• 1:57
  programmatic, especially, but digital
• 1:59
  advertising as a whole
• 2:02
  was generating so much data
• 2:05
  um that we could build products
• 2:08
  um on the
• 2:10
  um the basis that
• 2:12
  by leveraging the data, uh we could do
• 2:16
  we could help
• 2:17
  uh our clients uh do a better job. So,
• 2:21
  that really was the
• 2:23
  you know, in a this is like you would
• 2:25
  call that the mission statement.
• 2:29
  And so, by the time
• 2:30
  uh very few uh platforms
• 2:34
  uh were providing uh
• 2:36
  you know, granular data that could be
• 2:38
  leveraged.
• 2:39
  Um therefore
• 2:41
  um
• 2:42
  the early days of the company
• 2:45
  were all about uh AppNexus.
• 2:48
  Uh because um AppNexus was at the time
• 2:51
  the only platform providing log-level
• 2:54
  data. Um
• 2:56
  both at the level
• 2:58
  and even more interesting uh at the bid
• 3:01
  level.
• 3:02
  >> [snorts]
• 3:03
  >> And um
• 3:04
  based on that, we um
• 3:07
  we developed uh two products.
• 3:10
  One was more about, you know, uh
• 3:13
  analyzing the bid landscape.
• 3:15
  Uh that's if I recall very well that
• 3:18
  well that was uh analytics, the the name
• 3:20
  of the of the product.
• 3:22
  And the second product that was way more
• 3:24
  successful was uh optimizing
• 3:27
  uh
• 3:28
  um
• 3:29
  uh price floors
• 3:31
  uh [snorts] that were
• 3:33
  uh you know, second auction price based
• 3:35
  at the moment.
• 3:37
  Um and as you know, you know,
• 3:40
  this is not something you will be
• 3:41
  educated with, James. But uh second
• 3:44
  auction price uh provides way more
• 3:47
  opportunities of
• 3:48
  uh optimization than uh first price
• 3:50
  auction. so
• 3:52
  um
• 3:53
  then
• 3:54
  um um
• 3:55
  uh with the arising of uh header
• 3:57
  bidding,
• 3:59
  um we shifted um the uh analytics
• 4:03
  line of product to revenue management
• 4:07
  because uh
• 4:08
  with header bidding,
• 4:10
  um
• 4:12
  arose the
• 4:14
  heavy problem of data fragmentation.
• 4:17
  And since uh you know, we had developed
• 4:19
  a strong expertise in uh
• 4:21
  digital advertising and programmatic
• 4:23
  data, I mean, this was a
• 4:25
  you know,
• 4:26
  an easy an easy move
• 4:29
  uh to make.
• 4:31
  Um this is a
• 4:33
  product line that
• 4:35
  is still part of our portfolio,
• 4:38
  and I think we'll we'll we'll get to
• 4:39
  that.
• 4:41
  Um maybe the other notable
• 4:45
  um
• 4:46
  uh milestone in our history
• 4:49
  um is the creation of market data sets.
• 4:54
  Um video Leon,
• 4:57
  uh
• 4:57
  while the revenue management business
• 4:59
  was growing,
• 5:01
  um you know, we found out that we were
• 5:05
  sitting on a treasure of data.
• 5:08
  Um that when aggregated and anonymized
• 5:12
  could be very useful uh to our clients.
• 5:16
  And uh this was the birth of the market
• 5:20
  data set,
• 5:22
  on top of which
• 5:23
  uh we have created
• 5:25
  a new
• 5:27
  um
• 5:28
  product line,
• 5:30
  which is all about sales enablement.
• 5:35
  Um and uh last major change is obviously
• 5:40
  AI,
• 5:42
  um because uh
• 5:44
  literally I mean, we've been working on
• 5:45
  the eye since uh
• 5:48
  Q1 2025
• 5:50
  both
• 5:52
  uh I would say for
• 5:53
  uh our own purposes as a company
• 5:57
  but also for the products and our
• 5:59
  clients.
• 6:00
  Um which has uh led to the launch of uh
• 6:05
  Jean Pierre which is our uh
• 6:07
  In the US it's more called JP.
• 6:09
  Um
• 6:11
  uh which was launched
• 6:13
  uh in January uh this year
• 6:17
  and um which is now
• 6:20
  you know
• 6:21
  95% of uh our development bandwidth is
• 6:26
  on JP right now.
• 6:28
  >> Excellent. Well, thank you. You uh
• 6:30
  you've covered a lot of stuff that I
• 6:32
  want to get into and I'll come back to
• 6:34
  JP here in a moment. But uh I'm I've
• 6:37
  missed mentioning in in the my lead-in
• 6:40
  that you and I met uh over a decade ago
• 6:44
  uh when I was still at Yahoo and at the
• 6:46
  time I think you were you still you had
• 6:48
  the tools that you mentioned around the
• 6:50
  bid landscape and uh changing the
• 6:53
  pricing floors.
• 6:55
  Uh what I have not been as familiar with
• 6:57
  but I've heard of second hand is the
• 7:00
  market data set. And if you could just
• 7:03
  take a moment to describe what is in the
• 7:06
  market data set? Is it pricing
• 7:08
  information? Is it volume? Is it shares
• 7:12
  of different publishers? And the related
• 7:15
  question is is it directional or do you
• 7:19
  have um some sense of credibility?
• 7:21
  Because often we have people who have a
• 7:24
  very small subset of information about
• 7:27
  the market who then extrapolate. And so
• 7:29
  I'd like to to get us if you could talk
• 7:31
  as much about the data market data set
• 7:33
  as you could.
• 7:34
  >> Sure.
• 7:36
  Um
• 7:38
  So at first what I need to mention is
• 7:40
  that um we were able to build a market
• 7:44
  data set
• 7:45
  uh because of the way that we process
• 7:49
  our clients' data.
• 7:50
  So, maybe uh
• 7:52
  we'll get that We'll get to that, sorry,
• 7:55
  later, James. But, what is very very
• 7:58
  important is that you cannot build a
• 8:00
  reliable market data set
• 8:03
  if
• 8:04
  the
• 8:05
  uh the data that nurture and that is
• 8:08
  used to build the market data set is not
• 8:11
  standardized, okay? So, there's a heavy
• 8:15
  work of standardization that has to be
• 8:17
  done at the source.
• 8:20
  And in this instance, the source is not
• 8:22
  the data that we collect from SSPs or ad
• 8:26
  servers. It's the data that has been
• 8:29
  already processed, right? To become
• 8:34
  uh what we have
• 8:35
  uh define as the Adomic Data Engine.
• 8:40
  Okay, so
• 8:41
  we we we we are processing a huge number
• 8:47
  of publishers' data
• 8:50
  that is standardized, okay?
• 8:53
  And then
• 8:55
  we had um
• 8:57
  we had the
• 8:59
  really the the
• 9:01
  You know, we really wanted to make sure
• 9:02
  that our clients
• 9:04
  were comfortable with this idea
• 9:07
  of using their data in an anonymous way
• 9:11
  and aggregated way.
• 9:13
  And we had to explain to them that they
• 9:16
  would get more value
• 9:19
  out of participating, right?
• 9:23
  Then
• 9:26
  to not participate at the end of the
• 9:28
  day. And and and we need really to to to
• 9:30
  protect.
• 9:31
  So, this comes with a few uh carve-outs.
• 9:34
  So, what we guarantee to our clients is
• 9:38
  that no individual business relationship
• 9:42
  can be re-engineered from the market
• 9:44
  data.
• 9:45
  Okay?
• 9:47
  We guarantee that
• 9:49
  um any given views of the data and and
• 9:52
  I'll get to your question of what's in
• 9:54
  there, right?
• 9:56
  Um any given views of the data
• 9:59
  uh
• 9:59
  for any given views, sorry, of the data,
• 10:02
  no publisher shall represent more
• 10:05
  than 15%
• 10:08
  of the revenue scoped by the said view.
• 10:13
  Okay?
• 10:14
  Obviously, uh
• 10:16
  uh we are using
• 10:18
  you know, an infrastructure that ensures
• 10:21
  that everything is highly isolated, so
• 10:24
  the market data set has its own data
• 10:27
  processing.
• 10:29
  Okay?
• 10:30
  So, we cannot mix
• 10:32
  uh any uh
• 10:34
  you know, individual publisher data
• 10:37
  uh with the market
• 10:39
  uh
• 10:40
  uh data set.
• 10:41
  >> [clears throat]
• 10:42
  >> And uh obviously, as I said, but very
• 10:46
  important,
• 10:47
  all the data is anonymized, okay?
• 10:52
  Which means before being aggregated,
• 10:54
  which means that um if you look at the
• 10:57
  dimensions that are available in the
• 11:00
  market data set, the further you can go
• 11:02
  in on the sell side is the SSP.
• 11:07
  Okay?
• 11:08
  Everything beyond is obfuscated, so uh
• 11:11
  no obviously, no publishers know ad slot
• 11:14
  uh
• 11:14
  uh whatever.
• 11:16
  So, with all of these
• 11:19
  um
• 11:21
  how can I say, you know, with all of
• 11:24
  these rules in place,
• 11:27
  um we are we then had to ensure
• 11:32
  that we
• 11:33
  a common set of dimensions and metrics,
• 11:36
  right?
• 11:37
  That were representative enough
• 11:41
  of all the SSP data that we collect.
• 11:45
  Okay, so if you know, if we only had 10%
• 11:48
  of the SSPs providing buyers, you
• 11:50
  wouldn't find buyer in the market data
• 11:52
  set.
• 11:54
  Actually, you do because
• 11:55
  all the SSPs now provide
• 11:59
  buyer information.
• 12:00
  >> [clears throat]
• 12:01
  >> So on the buy side, you have
• 12:03
  um
• 12:05
  informations on
• 12:07
  advertisers,
• 12:09
  buyers,
• 12:11
  creators, and I'll get back to that
• 12:13
  because this is very unique
• 12:15
  to us.
• 12:16
  Um
• 12:18
  DSPs,
• 12:20
  SSPs,
• 12:21
  okay?
• 12:24
  And then you have a
• 12:25
  everything that is not
• 12:27
  um
• 12:28
  identifiable such as
• 12:31
  ad formats, transaction type,
• 12:34
  impression
• 12:36
  share of voice, and impression share of
• 12:39
  voice, revenue share of voice, and
• 12:41
  impression share of voice.
• 12:43
  Um and then you you can also break down
• 12:46
  the data by um
• 12:49
  formats
• 12:51
  and by media type, plus a few additional
• 12:54
  dimensions that are
• 12:56
  ours. Typically, we provide also a
• 12:58
  benchmark on web uh versus app, which is
• 13:02
  very very useful to to some of our
• 13:04
  clients.
• 13:05
  And because I know this is very
• 13:07
  important to you, James,
• 13:09
  uh
• 13:09
  >> [snorts]
• 13:10
  >> we also provide CPMs.
• 13:12
  So
• 13:14
  if you want to know how much Geico is
• 13:16
  paying for
• 13:19
  you know,
• 13:21
  in-stream video
• 13:22
  uh in the US
• 13:24
  when it buys through
• 13:27
  Google,
• 13:28
  uh this is a this is an information that
• 13:30
  we have
• 13:31
  in our database.
• 13:33
  So, then coming to your question of the
• 13:35
  representativeness and uh
• 13:38
  you know, the directionability.
• 13:41
  Sorry for my French.
• 13:43
  Uh of the of the data.
• 13:45
  Um it's a very good point that you are
• 13:47
  making. So, we always try not to
• 13:49
  oversell
• 13:50
  >> [snorts]
• 13:50
  >> uh this data.
• 13:53
  Basically, the the point is to say
• 13:55
  we don't have all the data, right?
• 13:58
  And uh
• 13:59
  this is a starting point.
• 14:01
  But,
• 14:03
  all the data that we have
• 14:05
  is 100% accurate and you can rely on it.
• 14:10
  Okay?
• 14:11
  Going back to my GEICO CPM, the GEICO
• 14:13
  CPM is actually the GEICO CPM, right?
• 14:15
  It's not extrapolated, it's not
• 14:18
  re-engineered, it's not calculated. It
• 14:20
  comes from the source, right? It comes
• 14:22
  from
• 14:23
  >> [clears throat]
• 14:24
  >> SSP reporting.
• 14:27
  So, to give you a more uh
• 14:29
  insights,
• 14:30
  today in the US, um
• 14:33
  our market data
• 14:35
  is around
• 14:37
  is around 5.5
• 14:39
  uh billion dollars of annual
• 14:42
  advertiser spent.
• 14:44
  Okay?
• 14:45
  So, depending on
• 14:48
  uh how you
• 14:50
  compute the size of the market,
• 14:52
  we are at about 15 to
• 14:55
  uh maybe 17%
• 14:58
  of the uh open web uh market. Okay?
• 15:03
  So, I will never
• 15:04
  um
• 15:05
  encourage [snorts] you to
• 15:07
  you know, build strong analysis uh of
• 15:10
  the market.
• 15:12
  But, if you want insights, right?
• 15:15
  On what is the CPM GEICO is spending
• 15:18
  before you meet with GEICO,
• 15:22
  that's a huge benefit for your sales
• 15:25
  pitch.
• 15:27
  >> Yeah, absolutely.
• 15:28
  >> If you want to understand
• 15:31
  what is the sales strategy of Nike on
• 15:34
  the open web,
• 15:35
  the buying strategy, sorry, of Nike on
• 15:37
  the open web,
• 15:39
  this is something that you can figure
• 15:40
  out.
• 15:41
  If you want to move one of your clients
• 15:43
  away from curation, so that it conti- it
• 15:46
  continues spending with you on PMPs,
• 15:50
  that's the kind of use cases that you
• 15:51
  can leverage thanks to the dealer.
• 15:54
  So, it's a whole trade-off between
• 15:57
  completeness
• 15:58
  and accuracy.
• 16:01
  >> Thank you. Now, let's get into JP. So,
• 16:05
  Jean-Pierre,
• 16:06
  uh you launched it recently, and tell me
• 16:10
  about the scope of what it does, because
• 16:12
  a lot of people when they hear an AI
• 16:15
  agent, maybe they think it's a chatbot,
• 16:18
  maybe they think it helps with
• 16:20
  analytics, maybe they think it's
• 16:22
  integrated into the operational systems.
• 16:26
  Everybody's investing in it. What What
• 16:28
  does JP do, and how is it different
• 16:31
  than, say, your competitors?
• 16:36
  >> Well, I would say
• 16:37
  your first question is very hard because
• 16:40
  the the the answer to the second is that
• 16:42
  JP is not an agent, JP is not a chatbot.
• 16:46
  JP is an infrastructure, okay?
• 16:49
  JP is sold to our clients as your AI
• 16:52
  infrastructure.
• 16:53
  It It doesn't have to be the only one
• 16:56
  within
• 16:58
  you know, a client ecosystem.
• 17:00
  But JP is an AI infrastructure or an AI
• 17:04
  toolbox, okay?
• 17:06
  So,
• 17:08
  what is very important What was very
• 17:10
  important to us when
• 17:12
  launching JP
• 17:13
  was really
• 17:16
  to create the conditions and the basis
• 17:20
  of
• 17:21
  that that that help and that
• 17:24
  are required
• 17:26
  to build efficient agent and efficient
• 17:30
  AI for our clients.
• 17:33
  As you know, James,
• 17:35
  the first
• 17:36
  thing that is absolutely critical is the
• 17:40
  data, right?
• 17:42
  Because
• 17:44
  while
• 17:46
  you know, LLMs and agent in general are
• 17:49
  renowned to be able to
• 17:51
  cope with very unstructured data.
• 17:54
  Um
• 17:55
  when it comes to numbers,
• 17:57
  when it comes to
• 17:59
  transactional data,
• 18:01
  um there has to be
• 18:04
  um
• 18:05
  some
• 18:07
  teaching
• 18:08
  uh to your LLMs
• 18:11
  that has to be done.
• 18:13
  If you want to
• 18:16
  make sure
• 18:18
  that
• 18:20
  the reasoning of the LLM is done on data
• 18:24
  that the LLM understands, right?
• 18:27
  That the LLM can interpret
• 18:32
  and then that the LLM can analyze to
• 18:35
  provide outcomes. So,
• 18:37
  materially,
• 18:39
  this has led us to work for
• 18:44
  almost 9 months
• 18:47
  on the JP semantic model.
• 18:51
  Okay?
• 18:53
  And here you get the idea now of, you
• 18:55
  know, this is an infrastructure because
• 18:56
  obviously and I'll come to that later,
• 18:59
  the semantic model is not used
• 19:01
  exclusively for for chatbot. It's used
• 19:04
  for whatever we want to build with JP.
• 19:07
  So, just one example, go to your clone
• 19:10
  and ask
• 19:12
  uh by uploading, you know, a few reports
• 19:16
  of SSP
• 19:17
  to create a view of your revenue with
• 19:19
  one of the old codes.
• 19:22
  The concept of a old code
• 19:24
  is totally understood by generic LLM.
• 19:28
  But to map it with your actual data,
• 19:34
  you know, at the beginning we try.
• 19:36
  This this this does not end well. This
• 19:39
  does not end well.
• 19:40
  Uh same thing for transaction types.
• 19:43
  Same thing for the different formats.
• 19:45
  Uh well, the differences between CTV and
• 19:48
  display.
• 19:49
  The concept the concept of of
• 19:52
  of app
• 19:53
  versus web. This all seems super
• 19:56
  obvious, but what is expected from us
• 20:00
  what is expected from you
• 20:02
  when you consult James
• 20:04
  is to be 100% accurate, right? Is to be
• 20:08
  accurate and reliable, right?
• 20:11
  So, you cannot leave any
• 20:15
  open
• 20:16
  corner for
• 20:18
  whatever interpretation to an LLM
• 20:21
  because at the end of the day this is
• 20:23
  going to come back through your clients
• 20:26
  that will say, "Hmm, this is not exactly
• 20:28
  what what I expected. And by the way,
• 20:31
  your data is not saying the the the same
• 20:33
  thing, right?"
• 20:34
  So, coming back to your question, at
• 20:36
  first what we wanted to achieve with GP
• 20:39
  is to create a semantic model on top
• 20:43
  of the data that we process.
• 20:46
  And there I'll come back to, you know,
• 20:49
  the standardized aspect of the atomic
• 20:52
  data engine.
• 20:53
  You know,
• 20:54
  if all of our clients add,
• 20:57
  you know, their own data schema. If all
• 20:59
  of our clients add their own
• 21:02
  differences
• 21:04
  in their different fields,
• 21:06
  it's impossible to to
• 21:08
  you know, a semantic model that is
• 21:10
  applicable
• 21:11
  um
• 21:13
  it it you know, it becomes custom work
• 21:15
  every time.
• 21:17
  And as a product company, that's
• 21:18
  obviously something we want to avoid,
• 21:20
  right?
• 21:21
  >> [snorts]
• 21:21
  >> So,
• 21:22
  the the semantic model was key. So, we
• 21:25
  we have
• 21:26
  based on the semantic model, we have
• 21:28
  built a chatbot, right? Which is called
• 21:30
  Riddle Analyst.
• 21:32
  And then on the side, you have Market
• 21:33
  Analyst that heavily relies
• 21:36
  uh on both our data and the
• 21:41
  the semantic model.
• 21:43
  >> [clears throat]
• 21:45
  >> But,
• 21:46
  the whole purpose of JP
• 21:50
  is to be able to address problems
• 21:53
  that our legacy products cannot.
• 21:57
  Right?
• 21:59
  So, today we are creating agent agents,
• 22:01
  sorry, that marry our data with Google
• 22:04
  Analytics data.
• 22:07
  We are basically using our AI
• 22:10
  infrastructure,
• 22:11
  right? AKA JP,
• 22:14
  to basically expand
• 22:18
  the value proposition that we are able
• 22:20
  to provide to our client by
• 22:24
  incorporating new sources of data.
• 22:29
  And also providing JP as an MCP server,
• 22:33
  so that our clients can build their own
• 22:36
  agents, right? That orchestrate the
• 22:39
  agents that are
• 22:41
  running within JP.
• 22:45
  >> Okay. Let let mini mini follow-up
• 22:47
  questions. Um but but I guess the the
• 22:51
  first one is if you now have JP
• 22:56
  and it has learned and has the semantic
• 22:58
  model,
• 23:00
  what does the deployment at a new client
• 23:02
  look like? So you're you start working
• 23:05
  with someone that you haven't worked
• 23:06
  with before. Does that mean now the
• 23:08
  process is much faster than it used to
• 23:11
  be in the old SaaS software days because
• 23:15
  it's much more seamless and you don't
• 23:17
  need to do too much training or is it a
• 23:21
  situation where it is still iterative?
• 23:23
  You need to go in there and have your
• 23:26
  agent review their model and do some
• 23:28
  learning. Is that what we're talking
• 23:30
  days, weeks, months or more for a
• 23:33
  deployment at a new customer?
• 23:34
  >> So
• 23:36
  let's let's start with, you know, the
• 23:39
  the pre-GP era and
• 23:42
  what we are used to do.
• 23:44
  Thanks to our standardized approach, we
• 23:47
  have always been able to onboard a a
• 23:50
  client, you know, as fast as in 2 weeks,
• 23:53
  okay?
• 23:54
  In full honesty, it never happens
• 23:57
  because there is some work on the on the
• 23:59
  client side who has to
• 24:02
  provide, you know, access and
• 24:04
  credentials to all the
• 24:07
  monetization partners that they want us
• 24:10
  to integrate, okay?
• 24:12
  Um
• 24:14
  Now I would say
• 24:16
  it's more AI than GP that has
• 24:20
  helped
• 24:22
  improve
• 24:24
  on this
• 24:25
  in this area.
• 24:27
  Um at first
• 24:29
  um we are identifying
• 24:33
  the data onboarding process. So what
• 24:35
  does that mean?
• 24:37
  I told you a lot about the Atomic Data
• 24:41
  Engine being standardized, but
• 24:44
  it's not a one-size-fits-all,
• 24:47
  right?
• 24:48
  And we always have the objective to
• 24:51
  ensure
• 24:52
  that the data that we structure for our
• 24:55
  clients
• 24:57
  speaks the client's language.
• 25:00
  Okay?
• 25:02
  Let me give you one example. We have
• 25:05
  some clients that don't want to see
• 25:07
  curation as an independent individual
• 25:11
  transaction type.
• 25:13
  So, they would come to us and say,
• 25:15
  "Okay. For me, curation
• 25:18
  uh you know, uh shall be accounted as
• 25:20
  PMPs."
• 25:21
  Fine.
• 25:22
  That's not a problem.
• 25:24
  That's the layer of you know,
• 25:26
  customization that comes on top of the
• 25:28
  standardization.
• 25:30
  Because again, what is super important
• 25:32
  to us, especially in revenue management,
• 25:35
  is that the data speaks the client's
• 25:38
  language. Okay?
• 25:40
  So, we are currently identifying this
• 25:43
  process of identifying what kind of
• 25:45
  customization we have to do
• 25:48
  in order to uh speak the client's
• 25:50
  language.
• 25:51
  We believe that this should uh I'm using
• 25:56
  conditional because in full candor, this
• 25:58
  is not yet uh
• 26:00
  available, but
• 26:02
  should be uh next month.
• 26:04
  Um we anticipate, you know, uh
• 26:08
  uh reduction by 30% of the uh onboarding
• 26:12
  time uh for
• 26:15
  our client.
• 26:17
  Then, we onboard
• 26:18
  uh users. Okay?
• 26:21
  And this gives me the opportunity to
• 26:24
  to anticipate maybe one of your question
• 26:26
  and to talk about uh the second version
• 26:29
  of GP or GP's new new newly born uh
• 26:33
  brother.
• 26:34
  Uh we are launching in December a new
• 26:37
  major version of GP
• 26:39
  >> [snorts]
• 26:39
  >> where
• 26:41
  that differentiates from
• 26:43
  uh
• 26:44
  the first one
• 26:45
  um by really creating a peer of a user.
• 26:52
  Basically, rendering the GP best
• 26:55
  teammate promise that we made initially.
• 26:59
  And this will
• 27:01
  um
• 27:02
  be translated
• 27:04
  in
• 27:05
  as many iterations of GP as needed with
• 27:09
  some templates such as GP for adults, GP
• 27:12
  for sales, GP for yield, okay?
• 27:16
  >> [snorts]
• 27:16
  >> And within the frame of
• 27:19
  so Jim she would have
• 27:21
  your own GP for yield obviously,
• 27:24
  but you again, this is not going to be a
• 27:27
  one-size-fits-all.
• 27:28
  When you are the first time you are
• 27:30
  logging in to GP,
• 27:32
  you will be asked tons of questions. You
• 27:34
  answer to as many as you want.
• 27:37
  That will create, you know,
• 27:40
  which is obviously teaching GP of who
• 27:43
  you are,
• 27:45
  how you speak,
• 27:46
  what are your objectives, what are your
• 27:48
  targets,
• 27:50
  and how you want
• 27:51
  to achieve them.
• 27:53
  So this is really a revolution in how we
• 27:56
  onboard users, and this should be
• 27:58
  available by December.
• 28:01
  >> Excellent. Um one of the challenges that
• 28:04
  I've seen in the industry
• 28:06
  since I've been in it 15, 20 years, is
• 28:10
  the complexity keeps going up with all
• 28:13
  of the different partners, all of the
• 28:15
  different demand sources, and so it's
• 28:18
  been really tough for uh publishers and
• 28:23
  anyone in the ad tech space to trust
• 28:27
  their data.
• 28:28
  So I guess a question given all of what
• 28:31
  you talked about in terms of the
• 28:33
  investment in the semantic layer, the
• 28:36
  other AI agents, the rollout of GP,
• 28:40
  are we getting closer to a world where
• 28:42
  publishers are going to be able to have
• 28:45
  the same sort of confidence that they
• 28:47
  had in their data back when they only
• 28:50
  had one ad server, you know, one site,
• 28:54
  you know, direct sold. Are we are we
• 28:57
  getting back to that world now?
• 29:00
  >> That's an excellent question,
• 29:02
  James, really. And I'm going to be very
• 29:06
  very honest with you.
• 29:08
  Uh this still remains a day-to-day
• 29:11
  challenge.
• 29:12
  Because
• 29:14
  um
• 29:14
  you know, you are talking about
• 29:17
  um
• 29:18
  how the industry itself has
• 29:22
  um the industry has made itself more
• 29:24
  complex.
• 29:26
  You know, new transaction types, you
• 29:27
  know.
• 29:29
  When you and I met, there was no
• 29:30
  curation. Uh maybe there was only one
• 29:32
  flavor of PMPs, all right? Um at the
• 29:35
  time. Um
• 29:37
  then
• 29:38
  you had new integration types. Header
• 29:42
  bidding, now server-side header bidding.
• 29:44
  Rappers.
• 29:46
  And and so on and so on.
• 29:49
  So
• 29:51
  it really depends also on
• 29:55
  um
• 29:56
  I would say the client's rigor.
• 29:59
  Right?
• 30:00
  Uh it has to be
• 30:02
  uh day-to-day objective
• 30:06
  uh because not everything can rely on
• 30:08
  the vendor. Let me give you an example.
• 30:10
  Um you add a
• 30:12
  a new uh SSP for testing purposes and
• 30:16
  there are three seats, one in prebid,
• 30:18
  one in whatever wrapper, and and one in
• 30:21
  your server-side prebid, right?
• 30:24
  And let's say you miss one seat that you
• 30:26
  share with us.
• 30:28
  There is going to be a gap. Okay?
• 30:30
  So
• 30:32
  we are building tools and we have tools
• 30:35
  on board
• 30:36
  that typically checks, you know, uh
• 30:40
  client configuration with us against
• 30:41
  their ads.txt.
• 30:44
  And we try to be proactive in
• 30:46
  identifying gaps. But same thing,
• 30:49
  not all
• 30:51
  ads.txt, you know, are uh are filled
• 30:54
  with the same seriousness, with the same
• 30:57
  rigor. And this is not always a a good
• 31:00
  source of of of comparison, right?
• 31:03
  Um we are always trying to identify gaps
• 31:07
  in our own data.
• 31:09
  Um
• 31:10
  Sometimes, you know, you have
• 31:11
  credentials that expire. Sometimes you
• 31:13
  have
• 31:15
  So, yeah, I mean, you know, one uh thing
• 31:17
  I can share with you,
• 31:19
  um
• 31:22
  our most fragmented client, if I may put
• 31:25
  that so,
• 31:27
  has 107
• 31:30
  monetization partners
• 31:34
  that we process.
• 31:36
  And that we have to uh
• 31:39
  basically
• 31:41
  and whose data we have to
• 31:43
  inject in its data engine.
• 31:45
  And this client wants to have dashboards
• 31:48
  that work. He wants to run GP on top of
• 31:52
  the data. So, we have this duty, you
• 31:55
  know, of doing the best job possible
• 31:58
  so that the as I said, data speaks the
• 32:01
  client's language and the data is as
• 32:03
  accurate as possible.
• 32:05
  That said, you know, sometimes we have a
• 32:08
  uh use cases um
• 32:10
  for finance teams.
• 32:12
  This is the times where we have no
• 32:14
  choice of being 100% accurate, you know,
• 32:16
  on D plus four at the beginning of each
• 32:19
  month. So,
• 32:21
  So, it's not undoable, but I would say
• 32:23
  it's a teamwork.
• 32:24
  The client and us.
• 32:25
  >> Sure.
• 32:27
  So, so let me ask for close with a
• 32:29
  couple questions I'm trying to ask
• 32:31
  everybody to see the different
• 32:33
  approaches that companies are taking.
• 32:35
  One of the challenges as AI becomes more
• 32:39
  and more powerful is how to put the
• 32:40
  right controls or the right guide rails
• 32:43
  in place to oversee the AI as it works
• 32:46
  with a human. How are you approaching
• 32:49
  the question of when you need the human
• 32:52
  in the loop versus letting the agent run
• 32:55
  on its own?
• 32:59
  >> Very, very good question but uh
• 33:02
  Do you have another one? Uh
• 33:04
  it's it's very uh
• 33:06
  It's [snorts] a very tough one. Um what
• 33:08
  I would like to say is that before even
• 33:11
  answering that question
• 33:13
  when designing an agent the the
• 33:16
  you know obviously with um so much
• 33:19
  ambition on JP uh we have built
• 33:21
  our own expertise and skills on you know
• 33:24
  what it takes to build a reliable ad
• 33:26
  tech agent or advertising agent.
• 33:29
  And the very first question that you
• 33:32
  will ask yourself is
• 33:34
  you know what kind I mean how should I
• 33:37
  get the data
• 33:38
  that I need?
• 33:39
  Okay?
• 33:41
  Should I call an agent? Should I make a
• 33:43
  SQL request? Should I run a Python
• 33:46
  routine?
• 33:47
  And basically by with
• 33:50
  you know I'm I'm I'm I'm sharing this
• 33:51
  story because
• 33:53
  it helps a lot limiting the guard rails
• 33:56
  that you have to apply
• 33:58
  um to your agent to make it safe.
• 34:03
  Basically you can avoid most of the
• 34:05
  guard rail. Check if the data is right.
• 34:08
  Approve the data. Things like that.
• 34:11
  We have many, many instances
• 34:15
  where for customized agents that we
• 34:17
  build for our clients or purpose-built
• 34:20
  agents that are generically
• 34:22
  uh available within the the the the JP
• 34:26
  framework
• 34:28
  we do a ton of Python and and SQL code.
• 34:33
  So this idea of defining
• 34:36
  you know, what food you give to your
• 34:38
  agent, is it deterministic data or is it
• 34:41
  probabilistic data?
• 34:43
  If you start by asking, you know,
• 34:46
  properly this question, you know,
• 34:48
  you are avoiding yourself a loads of
• 34:52
  headaches with guardrails.
• 34:55
  Because you you you you you have a clear
• 34:57
  idea, you know, of the reliability of
• 34:59
  the data
• 35:00
  the agent is reasoning on.
• 35:03
  Then, it's all about
• 35:05
  what you want to achieve. If
• 35:09
  you want to
• 35:10
  uh apply changes through an MCP server
• 35:13
  to one of your SSP platform,
• 35:17
  very often it's the company policy,
• 35:19
  right? We
• 35:21
  for the moment
• 35:22
  always always recommend 100% of the time
• 35:25
  a human in the loop for when
• 35:28
  things are are changed to an external
• 35:30
  system.
• 35:31
  We do believe that we are not there yet
• 35:33
  for full automation.
• 35:36
  But, uh we have built the infrastructure
• 35:38
  of change logs, you know, uh being able
• 35:41
  to track uh what has been changed by a
• 35:44
  human, what has been changed by an AI
• 35:46
  approved by a human, and things like
• 35:48
  that, because that's setting the stage
• 35:50
  for full autonomous agent
• 35:53
  which will arrive obviously in a very
• 35:55
  short future.
• 35:58
  >> All right. Thank you. Uh
• 36:01
  and and a similar question or a
• 36:03
  different question. What is something
• 36:05
  that you believe that is different from
• 36:09
  the majority of people either in ad tech
• 36:12
  or in AI development about the way we're
• 36:15
  going with AI?
• 36:19
  >> Uh that's another tough one.
• 36:21
  Uh Is is it okay to be a little bit
• 36:23
  controversial here? Yeah?
• 36:25
  >> Oh, that's the whole point of the
• 36:26
  question.
• 36:27
  >> [laughter]
• 36:28
  >> Okay.
• 36:31
  >> Look, you know, I think that at first
• 36:35
  I'm going to say something obvious, but
• 36:37
  that's always something I I love to
• 36:39
  recall
• 36:40
  especially to my kids.
• 36:43
  I I truly think that, you know, the AI
• 36:46
  revolution
• 36:48
  is on par with uh
• 36:51
  the industrial revolution, with Fordism,
• 36:54
  with you know, the
• 36:56
  the invention of the transistor or the
• 36:59
  rise of internet.
• 37:01
  As in the same way, it's going to
• 37:04
  totally reshape business forever, okay?
• 37:08
  So, that's really setting the the the
• 37:11
  the stage. And to add to that, we are
• 37:13
  still, you know, in the infancy of AI
• 37:16
  development, you know?
• 37:18
  I think definitely not due to the lack
• 37:20
  of
• 37:22
  sophistication of LLMs, not at all, but
• 37:25
  there are very few industries. There are
• 37:26
  some, but there are very few industries.
• 37:28
  Certainly not I think that leverage, you
• 37:31
  know, the capabilities of models like
• 37:33
  Opus or
• 37:36
  Fables.
• 37:37
  That's that that's totally for sure.
• 37:41
  My controversial point is to say that
• 37:44
  even if there is a turn in front of us,
• 37:47
  right?
• 37:50
  Even if live coding is fun, right?
• 37:55
  Even if everyone has
• 37:58
  you know, done some cooking with a
• 38:00
  closed recipe or has found some nice
• 38:04
  spots in their last vacation thanks to a
• 38:07
  few
• 38:08
  questions to uh
• 38:10
  ChatGPT or Gemini or
• 38:13
  whoever.
• 38:15
  My strong belief
• 38:17
  is that enterprise AI
• 38:20
  should remain
• 38:22
  in enterprise software companies.
• 38:26
  I see a lot
• 38:29
  So
• 38:31
  you can challenge my point that that
• 38:33
  as biased obviously given the position
• 38:36
  that we have in the
• 38:38
  in the value chain. That said, you know,
• 38:41
  I see a lot of
• 38:43
  clients, partners, or even friends that
• 38:45
  say, "Oh, that's easy. I'm going to
• 38:47
  write code that for myself."
• 38:50
  And I think yes,
• 38:53
  maybe. And I always praise, you know,
• 38:55
  you and I know how hard it is to develop
• 38:58
  a product, right? So I always praise,
• 39:00
  you know, people that
• 39:03
  into it to to develop their own stuff.
• 39:06
  But
• 39:07
  as easy as it seems
• 39:10
  you know, developing AI software
• 39:13
  remains developing software, right?
• 39:17
  So what about maintenance?
• 39:19
  What happens to your company when you
• 39:21
  leave the company? Who is going to be in
• 39:22
  charge of the agent that you've built,
• 39:25
  right?
• 39:26
  What
• 39:29
  security? Permission?
• 39:31
  Where is the data that your agent, you
• 39:34
  know, has produced stored?
• 39:37
  It's on your It's on the duck DB on your
• 39:39
  Mac.
• 39:41
  That's not going to fly very far.
• 39:43
  So
• 39:46
  I get it, you know, many many companies
• 39:49
  are building things and they are
• 39:51
  encouraged, you know, in our
• 39:54
  in our
• 39:56
  publisher industry mostly by Google and
• 39:58
  they are generally pushed.
• 40:00
  I would say
• 40:03
  do absolutely fine.
• 40:06
  But never forget that enterprise AI
• 40:10
  is enterprise software.
• 40:13
  >> I think that's a great place to stop. So
• 40:16
  thank you, Nico, and thank you for
• 40:18
  taking the time to talk today. As people
• 40:21
  digest this content, where can they
• 40:23
  learn more about you and more about
• 40:26
  Adomi?
• 40:27
  >> Adomi, I mean, we have just totally
• 40:30
  reshaped the website with tons of JP on
• 40:33
  it. So, I strongly recommend to uh
• 40:36
  to to check it out and uh feel free to
• 40:40
  send me a LinkedIn message. I would be
• 40:42
  absolutely delighted to uh
• 40:45
  have a follow-up conversation on a
• 40:47
  one-on-one basis through LinkedIn.
• 40:50
  >> Brilliant. So, thank you to you and
• 40:52
  thank you to the audience who stayed
• 40:53
  this far. I am James Dica. I am the
• 40:56
  Yield Doctor.