Latest / The Yield Doctor with James Deaker / What Happens When AI Agents Run Ad Ops? Burt CEO, Shay Brog

Transcript

What Happens When AI Agents Run Ad Ops? Burt CEO, Shay Brog · The Yield Doctor with James Deaker

0:00
I am James Deika. I am the yield doctor.
0:02
And today's discussion is part of a
0:06
project that I'm embarking on to talk to
0:08
companies that are driving change in the
0:10
new AI landscape for advertising,
0:14
specifically those who are using it to
0:16
drive solutions for publishers and
0:18
adtech providers. Today, my guest is the
0:22
CEO of BERT, Shy Brogue. Shai has been
0:25
in the industry for many years,
0:27
including companies like AOL and Gum.
0:31
And so with that as background, here we
0:33
go. Shai, I am so glad that we made this
0:35
happen.
0:36
>> Me, too. Thanks, James. I appreciate you
0:38
having me on.
0:39
>> Uh, I was going to start by asking you
0:41
about the origin of Bert, but I actually
0:43
want to start with your own origin
0:44
because as you and I were getting ready
0:45
to hit record, you told me and I was
0:48
surprised that you spent a lot of time
0:49
in Australia. So, what's the story
0:50
there?
0:51
>> Oh, yeah. Yeah. I'm uh I'm a globe
0:53
trotder for sure. I'm uh I'm Israeli as
0:57
uh the name indicates. I was born in
0:58
Jerusalem and then when I was 10 months
1:01
old, we moved to Sydney uh and I spent 5
1:04
and a half years there. So like I told
1:06
you, this this used to be a very heavy
1:07
Aussie accent. Uh there's still some
1:11
words there. I get Canadian a lot. Um I
1:14
definitely don't sound uh fully
1:16
American, but I've lived in New York
1:17
since 2009 and been in adtech since
1:21
before that. So, I'm dating myself here
1:23
a little bit.
1:24
>> Excellent. All right, let's get now
1:25
let's get really get into it. What is
1:26
the origin story of BERT? How did it get
1:29
started and how did you end up there?
1:32
>> Um, so BERT's been around for a while
1:34
actually. Um, surprisingly so, cuz when
1:37
I'll tell you about how I joined, I I
1:40
actually hadn't heard of BERT before I
1:41
joined BERT. Fun fact. Um, but BERT's
1:44
been around since like 2008, maybe 2009.
1:48
Uh, it's a Swedish company. It still has
1:51
uh um a strong presence in Sweden. We
1:54
have a large office in Goththingberg. Uh
1:57
and it was founded as a startup studio
2:00
of sorts in kind of martekch adtech and
2:04
BERT intelligence which is actually the
2:06
the name of the company now. um was an
2:10
internal analytics tool that was used
2:13
for internal stuff that then got sort of
2:15
externalized and spun off and eventually
2:18
sort of swallowed the larger startup
2:21
studio. So, it's gone through a lot of
2:24
uh evolutions and changes over the
2:26
years, but we're doing what we're doing
2:28
now since I guess about 2017. Um, and
2:31
we'll talk probably about what that is,
2:34
how I joined. Um, so like I said, I I
2:37
was recruited. I uh I'm not the founder.
2:39
I joined BERT in December 22 as head of
2:43
revenue. Uh and I became CEO in August
2:47
23. Um so it's been over 3 years now. Um
2:53
yeah, I I kind of fell in love with the
2:55
company pretty quickly when uh I don't
2:58
mind name dropping them. Uh uh when uh
3:02
the the recruiters reached out to me,
3:04
Three Pillars, um they're like, "Hey,
3:07
there's this Swedish company. They're
3:09
looking [snorts] for a new head of
3:10
revenue. Like, look, I haven't heard of
3:12
them, but I what I want is something
3:15
that is one publisher facing." Um
3:19
companies in our industry are generally
3:22
either oriented around publisher success
3:25
or advertiser success. It's pretty rare
3:27
to successfully do both.
3:30
I've been publisher facing my whole
3:32
career and I like being at companies
3:35
where that is their focus as well. Um,
3:38
so one, I want something that's
3:39
publisher facing and two, I want
3:42
something that isn't smoke and mirrors.
3:45
Um, you've been in adtech long enough as
3:48
have I. We all know there's a lot of
3:50
smoke and mirrors in adtech and I wanted
3:52
something that was uh real and material.
3:55
And so while I hadn't heard of BERT, um
3:59
I understood it was a SAS company. And
4:01
when I heard of some of BERT's
4:03
customers, frankly, um and and they
4:06
named dropped them, but it's like at the
4:08
time, you know, New York Times and
4:11
Paramount and Reuters, and I'm like,
4:13
wait, hold on. How have I not heard of
4:14
these guys? Um it's like uh ADTE's
4:18
bestkept secret or something. And as I
4:21
learned more and more what it is that we
4:23
do and how much value we generate for
4:26
customers and again we're we're
4:28
commercial off-the-shelf SAS business.
4:30
So um folks are paying us. It's not the
4:33
usual adtech of here's one line of
4:35
JavaScript and I'm actually paying you a
4:38
check at the end of the month which is
4:40
pretty typical when you're selling to
4:41
publishers. That that isn't the case for
4:43
us. So, we really need to earn that
4:45
value uh with our customers and just got
4:51
convinced the more and more I learned
4:53
about what we were doing and uh how much
4:55
our customers love us and the value that
4:58
we're driving for them, I just I feel
5:01
like I found my place.
5:02
>> Excellent. So, let's get into what the
5:05
company does because you've mentioned
5:07
you work with publishers, you mentioned
5:09
some of the partners. What is the
5:11
solution that you provide to those
5:13
partners?
5:14
>> So, BERT is the operational intelligence
5:17
layer for adup supported media
5:19
businesses. We unify fragmented systems.
5:23
We automate high friction workflows and
5:25
we help commercial teams make faster,
5:28
better revenue decisions. Um, it's a bit
5:31
of a mouthful. So in practice, what we
5:35
do is we connect firstparty ad servers,
5:38
third party ad servers, SSPs, DSPs, MS,
5:42
CRM, all of our delightful three-letter
5:44
acronyms that we love in our industry.
5:48
Um we bring it all together, we
5:51
normalize it, um we make it make sense,
5:54
we make it speak coherently and then we
5:56
address a lot of different use cases uh
5:58
for our customers around everything from
6:02
uh revenue reconciliation towards end of
6:05
month billing automation, client-f
6:07
facing reporting automation,
6:10
um partner pacing on the programmatic
6:13
side. Uh we also uh acquired a company a
6:17
few years ago maybe we'll talk about
6:19
this that uh math olympiades there
6:22
developed uh ad inventory forecasting
6:25
algorithms and this is something that
6:27
we've discussed in the past. So we have
6:30
uh forecasting as an offering. We have
6:33
um a very powerful aggregate data
6:35
network so we can address use cases such
6:38
as benchmarking and sales intelligence.
6:40
And so there's a lot there. Um but
6:43
basically serving as the single source
6:46
of truth, the foundational governed data
6:49
layer for all of the revenue operation
6:51
side.
6:52
>> Yeah. Let let me um ask a question both
6:55
that will help me understand the model
6:56
better and I think help the audience
6:58
because a lot of the things that you
7:00
said if you hadn't already said SAS
7:03
model twice when you described what you
7:06
do I would have put you in a bucket of
7:08
almost consulting and advisory services
7:11
or custom software development because
7:14
you talked about going in and solving a
7:16
range of different problems and so it
7:19
sounded very custom and ad hoc but then
7:22
You you said it's a SAS model. So help
7:24
me understand how you engage with the
7:25
customers a little bit better.
7:27
>> I think you actually hit the nail on the
7:29
head. We are probably a lot more managed
7:32
serviceoriented
7:34
I'll call it than anyone else out there
7:37
that sort of does what we do and there
7:39
there's definitely a few out there but
7:41
we are more custom. We are more
7:44
flexible. Um we will connect the last
7:47
mile. So it's not just oh here's SAS
7:52
commercial offtheshelf but you know
7:54
we'll do some customization but
7:55
basically you need to run the operation.
7:58
It does come with a pretty significant
8:02
um support and solutioning uh layer on
8:05
top of it. So onboarding and training
8:08
and stuff like that is one piece of it
8:10
which does involve a lot of
8:12
customization. Almost every media
8:14
business out there is their own thing
8:18
and looks at the world a little bit
8:19
differently. Um the tool itself, the
8:22
platform itself is built for that
8:24
flexibility inherently. Um in a way that
8:28
you can imagine like some CRM are built
8:31
to inherently support a lot of different
8:33
uh potential workflows and and
8:36
operations. uh but then there is a layer
8:39
of support that comes built in with that
8:42
SAS where for example we uh we have a
8:46
team we call them IOD integrations on
8:49
duty where we're m monitoring all of the
8:52
connections for all of our customers
8:55
across all of the ecosystem so that you
8:58
don't have to. So we are owning all of
9:02
that. We're setting up we're monitoring
9:04
we're setting up the data backfill
9:05
tasks. doing all that work behind the
9:07
scenes. So ideally you woke up in the
9:10
morning and you looked at your dashboard
9:12
and it was all just solved before you
9:13
had to look at it. So it is a little bit
9:16
of a model that sits somewhere in
9:18
between. It is in practice commercial
9:20
off-the-shelf SAS. You can do everything
9:22
on your own in the platform but we sell
9:26
it together uh to our enterprise
9:28
customers with a [snorts]
9:30
um layer of service and support that is
9:32
inherent in the model. So, um, as I
9:37
discussed, this is part of a project I'm
9:39
doing to learn what companies do. And I
9:40
know you've listened to a couple of my
9:42
other guests who have emphasized that
9:44
they think their competitive
9:46
differentiation is in the semantic layer
9:49
and how they know the data, how they
9:51
structure it, how they've worked with
9:53
different vendors in the past. And so,
9:55
when they go into an environment, they
9:56
kind of already have a structure in
9:58
place. Uh it sounds like when you go in,
10:01
you're solving potentially variety of
10:03
different problems on a variety of
10:06
different systems. How do you confront
10:09
the need to sort of balance that need
10:11
for structure with the need to solve the
10:13
custom problem?
10:14
>> Yeah. Yeah. I mean, we we do both,
10:16
right? So probably about 90% of the
10:21
magic that happens in the normalization
10:23
layer is out of the box because there's
10:26
stuff that is pretty standard in the
10:28
market. You know, uh um device mapping.
10:31
These guys call it a mobile phone, these
10:33
guys call it mobile, but ultimately
10:36
we're all talking about the same uh kind
10:38
of uh devices and we all sort of agree
10:41
on it. um or geom mapping or you know
10:44
there's a lot of those types of things
10:46
that happen um uh entity resolution
10:50
proctor and gamble png p amperand gp p
10:54
underscore amperand underscore g etc etc
10:57
all of those types of mappings are
10:59
really out of the box and folks don't
11:01
generally change that uh what is custom
11:04
to each individual's business is the
11:07
part that is that is the 10% extra so we
11:10
sort of do both in that we have our own
11:12
canonical layer and what we understand
11:15
how the universe works and how this deep
11:17
domain knowledge that we've built over
11:19
the last few decades around adtech works
11:22
but then if your individual business
11:25
I don't know you don't want to separate
11:28
between tablets and phones for example
11:31
making this up um yeah we can do that
11:34
for you and by the way if two years down
11:36
the road you change your mind about that
11:38
we can split that back out we have it so
11:40
there's A there's a layer of support
11:43
there around like here's how you look at
11:46
the world and then there's also the
11:48
underlying data that gets maintained. So
11:51
we can always roll back to the truth of
11:53
what it is uh of the signal that we
11:55
received and how we look at it and
11:57
normalize it. Um and we do the other way
12:00
around as well, right? If you want to
12:02
carve out [snorts] gaming consoles from
12:04
CTV, sure you can do that. That makes
12:06
sense to your business. Um, but there's
12:09
also just a lot of like naming
12:11
conventions and stuff like that, you
12:13
know, where it wouldn't make sense to
12:15
normalize them to normalize them across
12:18
uh the whole landscape. We call our 720
12:22
by 90 on the front page the leaderboard,
12:24
but somebody else might call it
12:25
something else, right? So every business
12:28
wants to look at the business in the
12:29
language that they understand their
12:32
business but then we also have the
12:34
capacity to take that and understand how
12:37
that then maps against the whole
12:39
universe when we're looking at our uh
12:41
aggregate data network for example h
12:44
>> how have you changed how you invest over
12:48
in terms of in innovation over the
12:50
course of the last let's say year and
12:52
you know I had a previous guest on who
12:54
said that they had changed their whole
12:56
development profile and putting 90% of
12:58
their development into AI. I don't need
13:01
to I don't assume that that's the right
13:03
answer. That's that's a a potential
13:05
play. But as you think about responding
13:08
to the current environment that we're
13:09
in, how has it changed how you think
13:11
about investing particularly in
13:14
different technology components?
13:16
>> Yeah, absolutely. And and I think we're
13:18
we're definitely more uh leaned in on
13:21
that front. So Bert was very early to
13:24
market on the AI stuff and and even
13:27
before that in ideiation. I presented to
13:31
my board something that we called at the
13:33
time BERT GPT u years ago at this point
13:37
and we even tried it in market with a
13:40
few of our alpha um testers and I don't
13:43
think the market was ready for it yet.
13:45
Folks were using the little chatbot
13:47
thing as um support basically. they they
13:51
they looked at it as oh this is a this
13:53
is a support widget primarily. Um so it
13:55
took a while. I mean this is even before
13:57
MCP was released as a as a standard by
14:00
anthropic. So we lean very heavily into
14:03
that once it was released. I think we're
14:06
one of the first I don't know five or 10
14:09
MCPS in the IAB's um agent and MCP
14:13
registry. Um so we've definitely leaned
14:15
in on that. It's probably one of the
14:17
fastest adopted products that um that
14:20
we've ever released and we're we keep
14:23
investing in it. We just released a new
14:24
version with some new tools the other
14:26
day. Uh so we're definitely heavily
14:28
invested in that future not just on the
14:32
customer side of being where the
14:33
customer is and how they're going to
14:35
operate but also internally to your
14:37
point about our own development and our
14:39
own investment in product. we are
14:42
building our own next layer of products
14:44
and services on the same technological
14:48
uh foundations on the same platforms and
14:51
so we ourselves are on this journey from
14:55
uh what I guess AI assisted is what we
14:58
call it internally to AI native and part
15:01
of that journey as we see our role in
15:04
the industry is to also help our
15:06
customers move from being AI assisted
15:10
you know uploading a spreadsheet and
15:12
then downloading the result like this uh
15:14
sort of meat proxy thing to more
15:17
automated outcomes where these things
15:20
can just fly through um and naturally
15:23
land in the process that they're looking
15:24
for. So we ourselves are in this process
15:27
and journey and I don't know that
15:29
anyone's done yet with that um as well
15:32
as sort of the overarching goal of
15:34
helping our customers get there as well.
15:37
And as you think about, you know, one of
15:39
my themes is always h the role of the
15:42
human in the loop. As you think about
15:44
the independence of these agents working
15:47
for your clients, how do they currently
15:50
and how do you in the future see them
15:52
engaging both with one of your clients
15:55
humans who is responsible for a certain
15:57
task and also the the BERT person
15:59
potentially behind the scenes who has to
16:01
be monitoring the guard rails. What does
16:04
that process look like? Yeah. So, it's a
16:07
bit of a governance question, but behind
16:09
that lies almost a philosophical
16:11
question, right? Um, there's almost
16:14
these two philosophies of technology out
16:16
there where one is
16:20
Google's are you feeling lucky today
16:22
button which kind of disappeared a few
16:24
years ago. But the notion that the
16:27
machine can just get to the answer for
16:29
you uh and can do the work for you and
16:32
is quite typical for uh Google or or
16:35
Meta's mode of thinking. And there's
16:38
another philosophy of technology which
16:40
is more maybe Microsoft and Apple kind
16:42
of uh uh signify it which is that the
16:45
role of the machine or the software is
16:48
to help humans be more productive. Um
16:52
and these are very kind of orthogonal I
16:56
guess philosophies. Um personally I
16:59
definitely tend towards more of the
17:01
human assisting humans to be productive.
17:05
Having said that, um, a big part of
17:08
BERT's value prop is automation and is
17:10
part of what we've always done for our
17:12
customers. And there is a lot of kind of
17:15
busy work, I guess, today that perhaps
17:18
doesn't need humans anymore. And part of
17:21
the opportunity that some of the
17:23
technology uh and some of the
17:25
technological advancements that have
17:26
happened over the last few years, it
17:28
kind of enables us to uh broaden the
17:31
swath of what are the areas that humans
17:34
really do need to be part of. Um and
17:38
some of that is one of the things that
17:40
I'm seeing is that uh especially in the
17:42
realm that we operate in and enabling
17:45
more sort of data democratization
17:49
is becoming curious has become a lot
17:52
cheaper uh and that's really beneficial
17:54
for organizations. There's a lot of
17:56
value to curiosity and what that can do
17:59
for orgs. So in some sense, let's
18:02
automate the easy stuff and leave the
18:05
hard stuff for humans in terms of
18:09
policies
18:10
uh defining policies and exception
18:13
handling and whatever. I do see a future
18:17
in this industry where a lot of uh a lot
18:21
of the manual work today is automated.
18:24
uh we might not even have a choice if
18:27
each and every campaign out there has
18:29
you know a 100,000 creative variants and
18:32
who who's supposed to traffic all of
18:33
those things. um that all needs to be
18:36
then automated downstream as well and
18:39
measurement and all of these gaps around
18:42
um the process uh you know the
18:43
nitty-gritty the the way the sausage is
18:46
made uh does need to be more and more
18:49
automated I think um but ultimately
18:52
there are still areas of policy and
18:55
decision- making and ultimately humans
18:57
are always going to be accountable for
18:59
those pieces um especially actually in
19:03
the world that you generally in I guess
19:05
around yield yield optimization
19:08
um is one of those topics that I don't
19:12
think the best of science can tell us
19:14
what is the single optimum point for
19:17
yield optimization it's um it's a moving
19:20
target it's a policy decision there
19:22
there's a lot of factors out there do I
19:25
care more about lifetime value do you
19:27
know am I going to turn on the greedy
19:29
optimization if it's a you know maybe
19:31
it's a bad quarter and I I need the
19:33
juice right So those are always going to
19:36
be decisions that at some level uh need
19:40
a human accountable for them. Uh but
19:43
hopefully at that point we've provided
19:46
uh a lot of tools that help make those
19:49
decisions um easier to get to. I I I
19:52
love the example because you know one of
19:54
my pet peeves is often people try to
19:56
find a solution without asking
19:58
themselves actually what is the question
20:00
as you and I have discussed in the past
20:03
are you solving for revenue growth or
20:05
profitability short-term versus
20:08
long-term um you know the the you can uh
20:11
optimize for your advertiser seeing more
20:14
return on ad spend and many publishers
20:17
never even stop to ask themselves what
20:19
they what their ideal outcome come is
20:21
from implementing these solutions. So I
20:24
I love I love that example for sort of
20:26
when you still need that human judgment.
20:28
Let me ask you the flip side of that
20:31
right now. What are the two or three
20:32
areas where your clients are seeing the
20:36
most benefit from automation either just
20:39
through traditional automation or
20:41
through the implementation of some kind
20:43
of agent to sort of put more color on
20:45
your the interaction that you have with
20:46
your clients.
20:48
>> Yeah. Yeah. Um so some of the stuff that
20:50
we touch upon probably a few of the more
20:52
popular ones are around um a revenue
20:57
reconciliation end of month billing. Um
21:00
in North America at least folks tend to
21:03
be build on third party numbers. Uh so
21:06
that's immediately a uh discrepancy that
21:10
needs to be tracked right between your
21:12
first party ad server numbers and the
21:14
third party whether it's you know an
21:16
innovate or a GCM or or wherever it's
21:18
coming from however the billing is going
21:20
to happen uh you immediately need to
21:22
track that there's a lot of noise around
21:25
that there's a lot of motions and stuff
21:29
um consolidating that at the end of the
21:32
month can be a multi-day project for
21:35
even very well uh uh equipped companies.
21:40
And so we one of the things that folks
21:42
do with BERT is automate that and make
21:45
that a lot easier to handle and turn
21:48
processes that could have taken you know
21:50
3 days or a week into something that
21:52
takes you know maybe an hour. Um, so
21:55
that's one good example. And I don't
21:58
think any of our customers have ever
22:00
said, you know, I really miss those
22:01
three days at the end of the month of
22:03
revenue reconciliation. It's really not
22:05
something that humans should be doing. I
22:08
think um, another piece of automation
22:12
that folks see a lot of return from is
22:16
client-facing reporting. Um this is uh
22:19
this is a piece where basically we help
22:21
automate the generation of campaign wrap
22:24
decks or or QPRs or or whatever the case
22:27
may be. But let's focus on campaign wrap
22:29
dickex the whole PCA process. Um maybe
22:32
you've been doing it manually for your
22:35
you know top five 10 50 clients. What
22:39
does everybody else get? Do they get
22:40
nothing? Do they get a spreadsheet? and
22:43
you know automating that process and
22:46
keeping it consistent as well telling
22:49
the same story in market that the sales
22:52
leadership wants to tell especially with
22:54
large sales forces it sometimes becomes
22:56
a little tricky hurting in all the cats.
22:58
Um getting that out there and getting at
23:01
with the right story beats that you want
23:04
to be telling on clients on how you
23:06
perform for them uh is really
23:08
meaningful. It drives up um net
23:11
retention rates for customers like
23:12
renewal rates. Basically, it's a it's a
23:15
meaningful thing that you can do to
23:17
drive outcomes. And again, you there's
23:21
no reason to manually produce uh uh
23:24
PowerPoint decks for uh you know a
23:26
thousand clients. It it doesn't make
23:28
sense. It doesn't scale well. Um so
23:31
that's something that can be automated
23:33
with the right guard rails in place, the
23:35
right data governance, etc. Um, and I
23:38
guess the last bit I'll mention maybe is
23:41
around uh ad inventory forecasting. So I
23:45
mentioned BERT acquired a company a few
23:47
years ago and had since introduced uh uh
23:51
BERT forecasting as a product um where
23:55
basically we do a better job of the
23:58
typical ad server forecaster. We handle
24:01
use cases that are more complex and
24:04
that's really money on the table. So
24:06
getting good at forecasting
24:09
um if you are also able to do a good job
24:12
with the analysis that follows that um
24:15
is really just money on the table
24:16
because otherwise folks are taking a lot
24:19
of buffers you know this working on kind
24:21
of pricing and yield forecasting is the
24:23
basis for a lot of that decisioning
24:25
right like what are the contending line
24:27
items
24:29
what do I think is available for this
24:30
product for that product what does the
24:32
package look like um and being able to
24:34
get as much uh out of as much juice out
24:39
of the those lemons that you can uh I
24:42
mean is a fundamental problem and being
24:44
a and having more sophisticated
24:46
forecasts uh that are able to handle you
24:49
know seasonality and and just better
24:52
address use cases that a lot of uh
24:54
enterprise publishers and retail media
24:56
networks see um is very very powerful.
25:00
Okay, let me uh you and I having this
25:04
conversation in Q4 of 2026 and I think
25:08
it's fair to say that for all of this
25:10
year, publishers, particularly
25:12
traditional media publishers, have been
25:14
under pressure. They're under pressure
25:16
because inventory can be trending down
25:19
as a result of AI enabling consumers to
25:22
get information in different ways.
25:24
Search revenues are certainly down. And
25:27
so when you think about the ways to help
25:30
publishers right now, do you do you
25:34
think there's more benefit on helping
25:36
them try to grow revenue and what could
25:38
be a very challenging revenue
25:40
environment or help them to take costs
25:43
out of the business through more
25:44
efficiencies? And obviously if you
25:47
choose to answer the question in a in a
25:48
different way, that's that's fine. But
25:51
it strikes me that it's a very tough
25:53
environment. And so the motivation for
25:55
the question is how can you help the
25:57
publishers in this tough environment?
26:00
>> Yeah, I mean we try to do both obviously
26:03
bringing up the top line is always uh a
26:05
good idea. Uh and again client facing
26:08
reporting automation is an example of
26:10
something that helps with that. But um I
26:13
think you're right that from a kind of
26:15
macro perspective when you're looking at
26:18
this market with exception of certain
26:21
niches like uh uh streaming media and
26:24
and retail media and maybe digital out
26:26
of home there there's definitely a few
26:28
places that are still uh growing uh
26:31
quite significantly. Uh for the most
26:33
part this is not a market that is
26:35
growing. there are not a lot of um
26:38
analog dollars still being translated to
26:41
digital dollars. And so the market is
26:44
growing in the sense that the economy is
26:45
growing, right? There's more advertising
26:47
dollars every year, but are they being
26:49
allocated to this market? And and how
26:52
much? And it does become more
26:54
challenging to grow in this environment.
26:58
AI is just a component of that
26:59
environment. I' I'd argue it's was
27:02
probably true even before that. Um, so
27:05
I'd say in that environment,
27:09
companies stop being measured on topline
27:11
growth and they start being measured on
27:12
bottom line. Um, where that does lend
27:16
itself to outsourcing and automation.
27:20
Yeah, for sure. Uh, and that definitely
27:23
sits in a bucket that we play well in.
27:26
Uh, we we are in that automation bucket
27:29
and we can help reduce costs and make
27:32
efficiencies there. It's not my favorite
27:34
way of presenting how BERT contributes
27:36
and produces value, but but it is true.
27:39
It is in some ways it can be a CFO sale.
27:42
Um because there are efficiencies to be
27:45
had by automating some of the grunt
27:47
work.
27:48
>> Um so publishers are under pressure, ade
27:52
companies are also under pressure. And
27:55
you mentioned that you have
27:56
traditionally had a SAS model. You can
27:59
choose to answer this either about BERT
28:01
or about the industry at large, but I've
28:04
seen companies start to ask themselves,
28:07
should we be changing our business model
28:10
to charge for tokens, move away from
28:12
SAS, charge for value ad? How do I I'll
28:16
ask it in the broadest context if you
28:18
choose to ask it, but for adtech at the
28:21
moment, how do you see people responding
28:24
to the challenge of incurring these AI
28:26
costs and basically having to work out
28:29
how to pass that value on to the end
28:31
customer?
28:32
>> Yeah. And and I think a lot of folks
28:35
have this challenge. There's definitely
28:36
a few companies out there that are
28:39
figuring that oh I can just arbitrage
28:41
the cost of tokens to uh you know a flat
28:44
fee and for now that's going to work for
28:46
me. What happens down the road? What
28:48
happens when the cost of tokens change?
28:51
Um it's also just how folks are
28:53
operating. We are thinking about this as
28:55
well. Um so today there's not really any
28:58
restrictions on how many queries you can
29:01
run with BERT. um because it's sort of
29:04
constrained by the friction of humans
29:06
running queries. But what happens if
29:09
that grows by four, five, seven orders
29:12
of magnitude because it's run by a
29:14
machine?
29:15
That has a cost to us, right? So there's
29:18
a level where we also need to think
29:20
about okay, but what if usage here
29:23
changed the patterns of usage changed
29:25
dramatically?
29:27
um do we still arbitrage that uh cost or
29:31
do we also need to be thinking around uh
29:34
usage and stuff like that? It it's not
29:36
my favorite way to build I'd say and
29:39
we've managed to avoid it to date. Um
29:42
but it is on our mind because we do
29:45
expect the usage patterns at our
29:47
customers to change. Um from I guess the
29:52
larger perspective on that uh you know I
29:56
I see SAS as a sign of a mature market.
30:00
I I think and there's been attempts to
30:02
this in the past and I you know I'm not
30:05
naming names or calling anyone but I do
30:08
think you know SSPs
30:11
maybe they should be SAS businesses. I'm
30:13
I'm not convinced that the model that
30:16
has run for the last 20 years has to be
30:18
the model um as it continues. Maybe it
30:22
makes sense and you know there's some
30:24
principal agent problems that are
30:26
playing out right now in the market that
30:28
could be resolved if um if you change
30:31
the model and how folks got paid or who
30:34
paid them uh to that end. And and I
30:37
think there's there's perhaps more
30:40
openness now to even think about some of
30:42
these things as the current model starts
30:44
to uh deteriorate. I think it is
30:47
worthwhile for folks out there that
30:50
normally have just oh let me tack on and
30:52
I'll just take another half a cent out
30:54
of the CPM out of the bidst stream and
30:56
I'll take another quarter of a cent and
30:58
you know folks were it was very easy to
31:00
tack on to this stream of revenue but
31:04
you know what as the market matures
31:06
maybe that isn't the easiest way or the
31:07
m or the way that makes the most sense
31:09
going forward uh both from a
31:13
business predictability perspective but
31:15
also aligning incentives and um the
31:18
structure of the market as a whole.
31:20
>> I I didn't think we'd end up there, but
31:22
I 100% agree with you and it still blows
31:25
my mind that SSPs charge a percentage of
31:28
the the revenue on the way through. And
31:31
frankly, the publishers continue to pay
31:33
it. Um at a time when I was responsible
31:36
for pricing an SSP, I was like, "This is
31:39
great." And so, of course, we snapped to
31:41
the model because it is very profitable
31:43
for the SSP owner. Um I am surprised
31:47
nobody has tried to use leverage against
31:50
the SSPs to make them uh have a slightly
31:53
different business model. Maybe it's a
31:55
SAS model, maybe it's something
31:56
different, but I I agree with you that
31:59
there are and and times like this when
32:01
you get a seismic change due to
32:03
something like AI are the times that
32:05
those sorts of things get put back on
32:07
the table. And so I'm sure those
32:09
conversations are occurring somewhere.
32:11
Uh let me ask you where what is
32:15
something that you believe that is
32:19
different from the majority of people in
32:21
adtech.
32:23
>> Uh your contrarian view question.
32:25
>> Correct.
32:25
>> Uh I think maybe just that I'm an
32:28
optimist. Uh [laughter] I think at this
32:31
at this point that's almost a radical
32:33
position. Um, especially, you know, I
32:36
talk to publishers a lot and we're uh
32:40
we're sort of in a privileged position
32:42
at BERT in that, you know, I I can see
32:44
my customers revenue. Um, and there is
32:47
pain. It's not evenly distributed, but
32:49
there's certainly pain there. Um, but
32:52
I've also been an adtech long enough as
32:54
of you. you know, we
32:57
we've had the cookie apocalypse and, you
33:00
know, we we are now in this AR uh
33:03
Armageddon and everything seems I don't
33:07
know, can feel very doom and gloom, but
33:10
again, I've just been around adtech long
33:12
enough. I I remember prior
33:14
esquetological events, you know, I
33:16
remember when viewability was dramatic.
33:18
I remember I mean, ads DXD shut down a
33:21
few uh uh adtech companies for sure.
33:23
like we've been through a lot of these
33:25
things. So, I don't want to sound
33:27
dismissive. There's
33:29
going to be winners and losers and some
33:31
of it might not be pretty. Uh, but I do
33:35
think that eventually the cards will
33:37
fall and we'll reach some kind of Nash
33:39
equilibrium and there will still be
33:42
publishers and there will still be
33:44
advertisers and they will be exchanging
33:47
dollars on the internet. Um, and yeah, I
33:50
think I'm also just generally optimistic
33:53
about, you know, what the people running
33:55
these businesses can do, uh, with with
33:58
better tools, not just BERT, but, you
34:00
know, all all over better tools. There's
34:03
a lot of expertise currently occupied
34:06
uh, with some pretty tedious work that I
34:08
think if you can unleash that talent
34:11
elsewhere,
34:13
you know, there's no end to human
34:16
ingenuity and what we can do. So I I'm
34:18
an optimist I guess is my contrarian
34:20
view.
34:21
>> I am also an optimist and I think part
34:23
of what you're talking about is almost
34:25
expectations of some of these companies
34:27
have been so high for growth. I I spoke
34:29
to someone this morning who had recently
34:32
left a commerce network and it was
34:35
because the network was not growing fast
34:37
enough. And this person pointed out that
34:39
they were still profitable. In fact,
34:40
much more profitable than mainline
34:42
business, but they'd been sold a story
34:45
around how fast this was going to grow
34:47
that was out of line for for all
34:49
reality. And I think there are a lot of
34:52
solid businesses in this space. And I
34:54
like you, I'm an optimist. Shai, thank
34:57
you so much for taking the time to talk
34:59
today. Where can people learn more about
35:02
you and about Bert?
35:04
>> Uh, I'm on LinkedIn. I'm pretty active.
35:06
Uh my in address is shib sh s sh s sh s
35:09
sh s sh s sh s sh s sh s sh s sh s h h h
35:09
h h h h h h h h h h h h h h h h h h y y
35:10
b. Uh, and BERT is uh either bert.ai,
35:13
which is probably the easiest one to
35:15
remember, or bertintelligence.com.
35:18
>> Brilliant. Thank you for the audience
35:20
for staying to the end. I am James
35:22
Deika. I am the yield