Latest / AI DTC WTF / How to Build an AI Agent Team for Your Marketing Agency | Tristram Dyer

Transcript

How to Build an AI Agent Team for Your Marketing Agency | Tristram Dyer · AI DTC WTF

0:22
Welcome to another episode of AI D2C WTF.
0:27
We are excited today to
0:28
have Tris Dyer on the show.
0:30
So AI DTC WTF is a super practical
0:34
podcast, so we don't
0:35
have time for introductions.
0:36
So we'll just go jump
0:38
into the good stuff.
0:38
So thank you for joining us today, Tris,
0:42
and excited to talk a
0:43
little bit of AI and what
0:45
you guys been doing lately.
0:46
Yeah, lots.
0:48
What haven't we been doing?
0:49
I guess, yeah, look, I think overall, one
0:55
of the things we've seen
0:56
through AI is it started
0:58
off about a year to year and a half ago
1:00
where we saw a few things
1:02
on ChatGPT and we're plugging
1:04
it in, seeing what happens.
1:05
And we've gone all the way now to pretty
1:08
much semi-autonomous
1:09
humans doing what they're doing
1:11
with AI, which is amazing to
1:14
really see the development.
1:15
But so many people find it scary, but I'm
1:17
only enthused about
1:18
where we can go with it.
1:19
So yeah, I can go top to bottom if you
1:22
like where we're going.
1:22
Yeah, for sure.
1:23
And I mean, I agree.
1:25
There's just lots of opportunities, but
1:26
also, of course, kind of
1:28
lots of fair and stuff like
1:29
that.
1:29
So yes, just tell us a little bit about
1:33
kind of how things have
1:33
been evolving lately and
1:35
what you guys have been doing.
1:36
It looks like you guys are using agents
1:39
and are like really
1:40
diving into the whole thing.
1:42
Absolutely.
1:43
Yeah.
1:43
We base our theory on, I think it's
1:45
Daniel Kahneman's book,
1:47
Thinking Fast and Slow.
1:48
We talk a bit more about kind of, you
1:50
know, there's two
1:50
different ways of approaching AI.
1:52
So first of all,
1:53
especially when it comes to work.
1:55
So from a fast point of view, you're kind
1:56
of thinking of, you
1:57
know, humans, speeding humans
1:59
up.
1:59
So, you know, it's a person's work.
2:01
They take their grunt work, take all the
2:03
stuff that, you know,
2:04
was done by execs and
2:06
et cetera, and that's moved into AI.
2:08
And that's not just,
2:09
you know, automating work.
2:10
It takes things from
2:11
minutes, from hours into minutes.
2:13
It speeds things up significantly.
2:15
But then from a slow standpoint, we've
2:17
got kind of that
2:17
background work to make sure there
2:19
is, you know, scheduled ongoing work
2:21
that's continuously needed.
2:23
The kind of slips through any cracks,
2:25
anything that, you know, you might as
2:26
part of a team might
2:27
miss or whichever that
2:28
kind of holds underneath.
2:29
And then it kind of builds out a strategy
2:31
that it helps as well to inform.
2:33
So I'll go through in detail.
2:35
I suppose that's kind of
2:36
the overarching look at it.
2:37
But from a fast...
2:38
Yes, absolutely.
2:39
Fast AI and slow AI, kind of.
2:41
That's the overarching...
2:42
Yeah, yeah.
2:43
Well, it makes sense from the slow AI
2:46
standpoint, because so many
2:47
people think, oh, AI is just
2:49
going to speed everything up
2:50
faster and faster and faster.
2:51
But actually, there's stuff that you can
2:53
do to help you slow
2:53
your thinking down as well,
2:54
because you still need to think as a
2:56
human as what you're going to be doing.
2:57
Got it.
2:58
Okay, so take us maybe through then.
3:00
Let's start with the
3:01
kind of the fast AI thing.
3:02
What kind of tools are you using?
3:04
What are you using it for?
3:06
Kind of give us the rundown on that.
3:09
So, yeah, I suppose on a fast AI
3:10
standpoint, it's very MCP-based.
3:13
It's very focused on trying to get fast,
3:15
bite-sized data that
3:16
you can use to analyze.
3:17
But it's getting larger
3:18
data sets to analyze quicker.
3:20
So, we use things that we obviously use
3:22
the meta MCP, the
3:24
Kaizen MCP and Motion MCP.
3:26
I'll talk about Kaizen now in a second.
3:28
But bringing in the data really, really
3:30
quickly into different
3:31
dashboards that allows us to
3:33
understand what's
3:34
actually happening in the world.
3:35
So, if that's performance creative, if
3:37
it's performance of an actual company,
3:42
how it's actually working,
3:43
what's making profit, what's not, where
3:45
are we spending, all that stuff comes in
3:47
really, really quickly
3:48
and allows a user to
3:50
use the AI like a tool.
3:51
So, if you think about going from using a
3:53
manual screwdriver into an electric
3:54
screwdriver, that's
3:55
what we've done there.
3:56
So, we've split it up quite quickly.
3:58
Obviously, the sexy kind of making
4:01
creative has been a
4:02
huge jump for us as well.
4:03
So, we've gone from a couple of designers
4:05
making some kind of
4:06
craters here and there
4:07
to jumping up and using something like
4:09
Arcads or Claude Design.
4:11
So, the two of those connect.
4:12
They work together.
4:13
So, Arcads allows us to create the UGC
4:16
that is AI generated.
4:18
But we then take that and move that into
4:20
Claude Design and actually to have the
4:21
overlays of the different
4:22
brands that we have over the top.
4:24
So, there's a couple of different
4:25
connectors that have worked quite well.
4:27
Arcads has been really revolutionary for
4:29
us because it's allowed us to be able to
4:30
build out the type of
4:33
creative that we want
4:34
without having to go and ask the client
4:35
to come back and actually
4:36
make us more and more content.
4:38
So, it's taking out that roadblock that's
4:40
slowing things down.
4:46
We all hear a ton about AI.
4:48
We all hear about what it can
4:49
do and how amazing it all is.
4:51
And I don't think that there's anything
4:53
incorrect about that.
4:54
But this episode is brought to you by an
4:56
AI brain that is absolutely incredible
4:59
for your marketing team called Runneth.
5:01
It's from Motion and therefore, you know
5:03
it's awesome because it's from Motion.
5:05
But really what Runneth does is connects
5:08
every single function of your marketing
5:10
department into one place.
5:12
It can do so many different things.
5:14
It can build you ads.
5:15
It can verify that
5:16
your brand looks proper.
5:18
It can verify that what your top
5:20
creatives were and you can ask it.
5:22
You can have it live in Slack or you can
5:24
have it live as a CLI so you can interact
5:26
with it any way you want to.
5:27
The thing that's really insane to me
5:29
about Motion is how absolutely
5:31
intelligent it truly is and how it works
5:33
cross-functionally across the team.
5:35
There was recently a marketer that Reza
5:38
talks about in the landing page on
5:40
Runneth, which you
5:41
should definitely check out,
5:42
where basically she went on vacation, the
5:44
CMO did, and said, if Runneth says the ad
5:47
is okay to run, then it is okay to run.
5:50
So she uploaded all
5:51
the context for the team.
5:52
They ran it into Runneth and Runneth
5:55
said, yeah, that ad is totally on brand.
5:57
It's absolutely crazy the number of
5:58
things that this can do.
5:59
And I think the fact that we all live in
6:01
silos and the fact we all live separately
6:03
is really, really challenging and one of
6:04
the hard parts about AI.
6:06
So that's not what this does.
6:07
This brings it all together with shared
6:10
context, shared resources, and it also
6:13
looks at thousands of ads every day.
6:15
So it's always up to date on what
6:16
actually is the best performing ads that
6:19
it's seeing across Motion.
6:20
And so that's another
6:21
whole thing that it does.
6:22
It's really crazy because instead of one
6:24
more dashboard, you get a teammate in
6:26
your Slack that knows
6:26
your performance data cold.
6:28
As I said, you can ask it really anything
6:30
and answers instantly.
6:31
Weekly recaps, top creatives, which
6:33
creators your audience trusts.
6:35
It learns your brand, your playbooks, and
6:37
gets smarter every day.
6:38
E-commerce and D2C teams are using it to
6:41
turn weekly reporting from a two-day job
6:43
into a two-minute conversation.
6:45
Head to runneth.com
6:46
and see it for yourself.
6:47
And it also has a
6:49
30-day money-back guarantee.
6:50
So if you need a cheat code in e-commerce
6:53
and D2C with your ads and your ad
6:56
creative and your marketing
6:57
team, check out runneth.com.
6:59
30-day money-back guarantee.
7:01
It's absolutely incredible what this
7:03
thing can do, and I
7:04
encourage you to give it a shot.
7:10
Okay, so just to kind of, so I understand
7:13
because there's a lot of things.
7:14
So you have then MCP connections into
7:19
typically cloud, cloud code, cloud code
7:21
with meta, motion, and Kaizen.
7:25
And that's kind of where you do the
7:27
brainstorming, creating scripts or
7:30
working on strategy.
7:31
Yeah.
7:32
And then, yes.
7:33
Let me put you on Kaizen because that's
7:34
actually an interesting
7:35
one that I glazed over there.
7:36
So, I mean, it's really good to bring in
7:38
motion and meta information.
7:40
That's all the numbers, right?
7:41
And that's all the quantitative data.
7:42
But what we miss, and a lot of times
7:44
marketers miss this, is the qualitative
7:46
data that comes with it.
7:47
So talking to the marketing manager,
7:49
talking to the founder, like what they're
7:50
actually trying to do.
7:52
And everyday conversations, it's not just
7:53
conversations of like, we're
7:55
sitting down to talk about this.
7:57
It brings in the conversations from the
7:59
account manager with them, or it might be
8:00
talking to the
8:01
operations manager in the company.
8:02
All of that's recorded and brought into
8:04
Kaizen, which is
8:05
where we record our calls.
8:07
But it actually allows us to understand
8:08
sentiment analysis, understand what's
8:10
really good and what's
8:11
really bad, what they really like.
8:13
And brings that
8:13
information, too, into the strategy.
8:16
So it's like the founder might have
8:17
talked about it this way, or whatever the
8:20
conversations might happen.
8:21
That connects in them with motion and
8:23
meta data to be able to give you
8:25
information in a much richer format
8:27
rather than just looking at numbers.
8:29
Because we're not.
8:29
Okay.
8:30
So Kaizen is like connecting all the
8:32
transcripts and all that data.
8:36
So that is kind of the intelligence, the
8:40
client intelligence layer of it.
8:43
Yes.
8:43
Got it.
8:44
Yeah.
8:44
Okay.
8:44
And again, your team is using cloud,
8:47
co-worker, cloud code with these MCPs to
8:50
kind of connect everything.
8:52
Every single person in our
8:53
company has a Cloud Pro account.
8:55
Straight up.
8:56
Like that's just the standard.
8:57
Because so many people in the industry,
8:59
in the agency world have gone, I'm not
9:01
too sure, should I be using AI or not?
9:03
We just went headfirst into it.
9:04
I said, right, everyone has a Cloud
9:06
account, a paid Cloud account that they
9:08
can use that comes with
9:09
Cloud Code, Cloud Design.
9:12
And we keep those in Cloud
9:13
projects as well per client.
9:15
That's cool.
9:16
Just to jump in, because again, I'm super
9:18
nerdy about this and
9:19
this is exciting for me.
9:20
So I've looked a little bit at Arc ads
9:24
and those kind of things.
9:26
I think that's really, really good.
9:28
And I see a lot of
9:29
other agencies are using it.
9:31
Cloud Design is a bit new, I think.
9:33
And so how do you work
9:36
together with those things?
9:37
Like super practically, how would your
9:39
team combine those
9:41
two to create an output?
9:43
Yeah.
9:43
So the way we would do it is we create
9:45
the initial designs and the
9:47
initial briefs with Claude.
9:49
And Claude would then be connected.
9:50
And we'll go into the
9:51
slow thinking now shortly.
9:53
But they would design briefs and
9:54
everything else with that.
9:55
That would input into Arc ads.
9:57
It would actually bring in different
9:58
aspects of winning creative and winning
10:01
performing creative out of
10:02
that into a brief for Arc ads.
10:04
Arc ads then goes and creates the basic
10:07
structure of the ad with
10:08
all of the creative in it.
10:09
It might be a UGC person talking about a
10:11
certain product with
10:12
the product in their hand.
10:13
We then bring that into Claude design to
10:15
actually be able to build out statics,
10:17
build out kind of iterations on that.
10:20
So that we're not just launching with one
10:21
version of it, but
10:22
several versions of the one.
10:24
So it might, and we use something like
10:26
HeyGen that might take that and actually
10:28
translate that into
10:29
six different languages.
10:30
So we work with people all across Europe
10:31
that want to translate into those
10:33
different languages.
10:34
So it'll take the initial ad and
10:35
translate those really, really quickly.
10:37
So it's all about one of the things when
10:38
you're using across like Claude design
10:40
and Arc and everything else, it's kind of
10:43
token maintenance, I call it.
10:44
So if you just start throwing it all into
10:46
one AI, it uses all your
10:48
tokens really, really quickly.
10:49
But if you're clever about where you
10:51
brief it, so for Claude design, if you
10:53
get your briefing and everything else
10:54
done before you get to the design stage
10:56
and use it in actual Claude code,
10:58
it doesn't burn as many
10:59
tokens in Claude design.
11:02
And that will actually save
11:03
you a lot in the long run.
11:04
It'll be a lot cheaper.
11:05
Tris, I'm wondering of the stuff that
11:07
you've done under fast AI, do you see
11:10
like have client results
11:12
increased as a function of this?
11:15
Because you're able to spool stuff up
11:17
faster and have client relationships have
11:20
gone better as well.
11:21
Like, can you give us some tangible
11:23
examples of like where
11:24
it's impacted both of those?
11:26
Absolutely.
11:26
Yeah.
11:26
I mean, that's, that's exactly it, right?
11:28
We always talk about two parts of the
11:29
job, doing the job and telling someone
11:31
you've done the job.
11:32
The first part is actually really
11:33
increased because what we're able to do
11:35
is make decisions faster.
11:36
Is that going to lead to a better result?
11:38
Is this, is one connected to the other?
11:40
Is a sale on a meta connecting to a sale
11:42
on the actual, on the Shopify?
11:45
So there is a very tangible seat, like
11:48
speed in terms of the speed increase
11:50
there in terms of making decisions.
11:52
But then the second part as well is
11:54
actually allowing the client to
11:55
understand what we've
11:56
actually done and how we've done it.
11:58
So when we're making changes, when we're
12:00
doing certain things, we'll allow our
12:02
agents that we'll talk about shortly to
12:04
be able to actually write summaries for
12:05
that in a very simple format, not just
12:07
spit out slaw, but actually explain,
12:10
explain to clients, this is what we've
12:11
done and why we've done it.
12:13
We'll read it and go,
12:13
yeah, happy with that.
12:14
Send it on.
12:15
So with that constant communication, that
12:17
constant contact allows that step up in
12:20
trust, but also allows us to be able to
12:22
analyze and go, that's what I've done.
12:24
Yes, I'm happy with that.
12:25
Or no, undo that and
12:26
change it and move it back.
12:27
So in terms of tangible results, that's
12:30
actually being, if that being creating
12:32
ads for a specific product in a campaign,
12:36
it will actually create ads a lot faster
12:38
and actually start running ads a lot
12:39
faster as a result of this.
12:41
I think that another tangible result is
12:43
actually part of our kind of, one of our
12:45
daily agents is sending a daily update
12:47
and actually getting that daily update
12:49
written and transcribed from six or seven
12:51
different data sources into one and
12:53
actually written and then analyzed by a
12:56
human and sent across.
12:57
That would typically have been a lot more
12:59
of the person's time.
13:00
Now they can actually spend the time
13:02
going and analyzing that that worked and
13:04
actually going and making decisions on
13:05
where they need to go next.
13:07
So that's some two
13:08
tangible results on those.
13:09
The processes that you've
13:10
already done here with Fast AI.
13:13
What are the other things that you're
13:14
like, look, like what's the six month
13:16
vision of doing some
13:18
of this stuff, right?
13:19
It's, it's, I mean, it's an analysis.
13:22
It's, it's building more, you know,
13:25
building better ads.
13:26
It's discovering new personas.
13:28
Like what are, what's
13:29
the roadmap for you?
13:30
Cause I assume what happened is you like
13:31
had them sit down and say, look, this is
13:33
where you're spending time.
13:34
This is what we're, you know, and like,
13:36
we're going to optimize that.
13:37
But like, what's the looking forward for
13:38
you in terms of
13:39
continuing to make this better?
13:41
Yeah, I think the biggest thing for us,
13:43
the biggest kind of growth lever for us
13:44
is actually not just building it for
13:45
ourselves, but allowing others to be able
13:47
to use these products themselves.
13:49
Cause when I'm starting to realize a lot
13:50
more quickly is that, you know, we'll
13:52
build it for
13:53
ourselves and it'll be great.
13:54
But the democratization of all of this is
13:56
going to be huge where people will go,
13:58
well, I could just make that myself.
14:00
So we want to put those tools into the
14:02
hands of owners and founders so they can
14:04
go and use those information, have an app
14:06
on their phone, or even be able to
14:08
connect to when Siri
14:09
eventually starts to connect.
14:11
To Claude, we can actually start to make
14:12
artifacts that people
14:13
can see on their phone.
14:14
So it's actually starting to roadmap over
14:16
the next six months
14:17
with WWDC coming up soon.
14:19
They'll release obviously anything that,
14:21
like I said, connecting Claude to their
14:23
iPhone, where they'll be actually able to
14:25
make ad changes on their iPhone or
14:27
actually be able to get that
14:28
information directly to them.
14:30
Because if you think what a founder wants
14:32
is the clarity and the knowledge that
14:33
their agency is making the right
14:35
decisions, but also doing that and
14:37
leading them in the right direction.
14:38
So our roadmap is to be even clearer and
14:41
give the founder even more knowledge and
14:43
power to be able to go and make things
14:44
like ordering decisions or marketing
14:47
decisions they want to talk about.
14:48
So it's empowering
14:49
founders a lot more quickly.
14:52
So creating agents that we like and then
14:54
making them publicly available.
14:55
Interesting.
14:56
That's super interesting.
14:58
And then as a business model, you're
14:59
thinking about like a monthly retainer
15:02
per agent or something like that?
15:03
Or what's your thinking around that?
15:05
There's two parts of it.
15:06
There's either a retainer model on our
15:08
side where we would say, right, we'll
15:09
continue to make you those agents.
15:11
I will make you newer and more improved.
15:13
Or it's going to be a
15:14
simple case of this.
15:15
You know, just because you have a
15:17
screwdriver doesn't
15:17
mean you can build a house.
15:18
You still need somebody to run it,
15:20
operate it, make those decisions.
15:21
We always make these, we design these
15:23
agents with you and in the loop.
15:25
Because we don't like, there's no point
15:26
in having a fully autonomous platform.
15:29
Because what happens is it'll just go and
15:31
start, you know, making all sorts of
15:33
decisions that may not actually line up
15:34
with your overall goal.
15:35
I don't know if you remember Silicon
15:36
Valley with Son of Anton, where he
15:38
started making AI decisions like that.
15:40
That's the thing we're
15:41
trying to avoid here, right?
15:43
So I think, you know,
15:44
yes, that's a silly example.
15:46
But I think that's something that, you
15:47
know, we want to make sure that we still
15:48
have control and direction
15:49
and be able to make those calls.
15:51
But it's something we're seeing for like
15:53
time and time again, that every time you
15:55
kind of think you've made something
15:56
innovative, Claude or something else will
15:58
release something that will just
15:59
completely squash it.
15:59
So if you're making a SaaS
16:01
product, you know, that's great.
16:03
But then you can go and spin that up in
16:04
five minutes, which one of the things I
16:06
would say is that power is the data.
16:08
So that you can make a SaaS model, but if
16:10
it's not based on anything
16:11
concrete, there's no point.
16:13
So our business model then around that,
16:15
Thomas, to a question is like,
16:16
what do we actually monetize?
16:18
How do we work it out?
16:19
It's looking at the
16:19
data and how we use it.
16:21
That's the direction.
16:22
We've always looked at ourselves as a bit
16:23
of a fiduciary and actually managing
16:24
people's money correctly.
16:26
That's one of the things that we will
16:27
continue to do, but at a higher level now
16:29
that we've got AI plugged in.
16:31
Can you walk us through the agent team
16:33
that you have and like the different
16:35
things that they're doing?
16:36
Just run us through the whole list of
16:37
every day, every week and every month.
16:41
It's a perfect transition, I think, then,
16:44
Andrew, from what Chris was saying, is
16:46
like the data part of this, right?
16:48
Because your agents are able to do what
16:52
you're going to walk through now because
16:53
you have this data warehouse where you're
16:55
connecting everything, right?
16:56
Absolutely.
16:57
Absolutely.
16:57
Yeah.
16:57
So this is exactly it.
16:58
So we keep our data in this very secure
17:00
data warehouse on Google BigQuery.
17:02
And what we do is we import the data on a
17:04
daily and sometimes hourly basis,
17:06
depending on the connector.
17:07
But what we then bring that information
17:09
in is that we have a central store of
17:11
everything that you need to
17:12
know about a certain brand.
17:14
So that'll be whether it be quantitative
17:17
data, where it's literally all the
17:19
numbers and everything else, or the
17:20
qualitative data on like, is this
17:23
happening or when's that happening?
17:24
What's the plan and everything else?
17:26
All of that comes into one place.
17:27
But then we start to identify, for
17:29
example, what the goals are,
17:30
what the ROAS is, et cetera.
17:32
And every morning we get a
17:34
performance anomaly detector.
17:36
That's agent one.
17:37
And what it does is it catches any ROAS
17:39
or CPA problems that are
17:40
happening across all the agents.
17:42
And it'll DM either me or it'll DM the
17:44
account manager that's working on it
17:46
saying, there's a problem here.
17:47
Take a look.
17:48
So when they open their phone or open
17:49
their laptop in the morning, that's the
17:50
first thing they see.
17:51
Have a look at this.
17:52
And we've refined that over time because
17:54
sometimes the CPA goal might
17:56
not be clear over what period.
17:58
So we started to refine it where it gives
17:59
it a 3, 7, and 14-day view to see if
18:03
those anomalies are working.
18:04
And these agents, are these cloud managed
18:07
agents or are these open claws or what
18:10
kind of tech is in the background here?
18:12
So these are cloud routines that are
18:14
running in the background.
18:15
Got it.
18:17
So you have a separate machine that's
18:19
like, these are all the routines and it's
18:21
just running through.
18:22
Yeah, a good friend of mine, Thomas Moen,
18:24
actually showed me that.
18:27
Yeah, it was one of the things that, I
18:29
mean, it's great to share this, but it's
18:30
like where we actually were able to say,
18:32
right, we have a machine
18:33
there that runs all this.
18:34
And we've got a bunch of other things
18:36
around, uploads and stuff like that, that
18:37
runs through the machine.
18:39
And the idea then is that this is a agent
18:41
to run off the back of this, but we have
18:43
a self-improving agent that actually goes
18:44
back through any
18:45
feedback that our team gives.
18:47
So if we're missing, for example, missing
18:49
ROAS or the daily update we'll talk about
18:51
in a second is slightly incorrect.
18:54
That will be corrected in Slack.
18:55
And then the agent will go back through
18:57
Slack and read all of
18:58
that and self-correct.
18:59
Cool.
18:59
That's very cool.
19:00
Okay.
19:00
So keep going through
19:01
your different agents.
19:02
It's just cool to hear how
19:03
it's in the background as well.
19:05
So sorry.
19:05
Of course, yeah.
19:06
So this is part of the slow
19:08
AI, as I was talking about.
19:08
So you've got your budget pacing to make
19:10
sure you're on budget.
19:11
That will, again, be
19:11
part of your overall.
19:13
Then you've got your morning briefing
19:14
agent, which is what we call agent 00,
19:16
because the idea here is that this starts
19:18
to gather all of the information in one.
19:20
So how the performance is going, where
19:22
you are in terms of your budget spend and
19:25
how your CPA is going, it
19:26
starts to come together.
19:27
Think Jarvis out of Iron Man.
19:29
That's what you wake up in the morning.
19:30
It's like, this is what you've got.
19:32
So then we obviously have a daily report.
19:34
We've got a detailed performance report.
19:35
Then in terms of underneath it, it goes,
19:37
okay, so this is how your CTR is going.
19:39
This is how many add to carts you've got.
19:41
This is what your conversion
19:42
rate is and so on per product.
19:45
If you want to call out something in the
19:46
previous day that you want to keep an eye
19:47
on, it will then call those
19:49
out in those daily reports.
19:51
So this is almost like having an exec
19:52
directly with you on each of those.
19:54
We then obviously have our meta ads
19:55
operator, which is connected
19:57
directly to Facebook's MTP.
19:58
So it verifies everything that comes into
20:00
in through our connector into the back
20:02
end of BigQuery, but
20:02
verifies it through meta.
20:04
Because sometimes you want to make sure
20:05
that the data is correct.
20:07
But then what it does is suggest any ad
20:09
approval changes we've got in there.
20:10
So we've trained this agent on our SOPs
20:12
in terms of ad buying.
20:14
And then it goes and makes suggestions
20:16
and gives you the option to thumbs up or
20:18
thumbs down in terms
20:18
of making those changes.
20:20
So those changes, again, it's required as
20:22
a human in the loop.
20:23
Because we, again, we're responsible.
20:25
Regardless if we put an agent running it
20:26
or not, we're still responsible.
20:28
So that's really, really important.
20:30
So, again, I'm just trying to dig down to
20:32
the specifics of things
20:33
so I can understand it.
20:34
So it's cloud and schedules and kind of
20:38
the outputs of thumbs up, thumbs down.
20:40
Is it generating artifacts?
20:42
Is this like built-in notion?
20:44
What's the interface here?
20:46
All of this is created into Slack.
20:48
So we created a Slack app for this.
20:51
And so what happens then?
20:52
It drops it into Slack.
20:53
And the communication is back and forth
20:55
with a human in Slack or a user in Slack.
20:59
So when you wake up in the
21:00
morning, you have an exec.
21:01
If you're an account manager, you have an
21:03
exec that's connected to Cloud, powered
21:05
by Cloud in the background on a separate
21:06
machine, but it's connected in Slack.
21:08
And you communicate it like that.
21:10
So it can be done through your phone.
21:10
That's super cool.
21:11
That's super cool.
21:13
So, yeah, at the end of the day, then you
21:14
get sent in kind of a flag or any actions
21:17
that are needed because we usually have
21:18
external and internal
21:19
Slack channels with clients.
21:20
So we have our own Slack channel where we
21:22
talk about stuff with the design team or
21:23
with the data team or
21:25
anything else like that.
21:26
And then the
21:26
communication with the clients.
21:27
And it scans all of that throughout the
21:29
day and gives a daily brief of
21:30
like, this is what's happened.
21:31
This is what you've done.
21:32
This is what you may have missed that you
21:33
need to update before you leave.
21:35
And so that's a very clear kind of close
21:37
of day, making sure
21:38
that everything is done.
21:39
And we've seen client satisfaction
21:41
results really increase.
21:42
People are super happy.
21:43
They feel like we're on the ball.
21:45
And we are.
21:46
But this is a real helping hand in that.
21:48
So that's kind of what about your team?
21:50
How did that they kind of adapt this new
21:52
world of agents, giving
21:55
them a hand, basically?
21:56
How has that been?
21:57
It's good.
21:59
You know, the way I look at it is we've
22:00
got the right people on the bus.
22:01
The people who are working at this
22:03
company are very forward thinking, very
22:05
innovative and very focused on how they
22:07
can improve themselves and the work that
22:09
they're doing, make a
22:10
more meaningful change.
22:11
So it works if you
22:13
have the right mindset.
22:14
If they come in and have fear that
22:15
they're going to be replaced by this,
22:17
they're going to be
22:18
rejecting of it straight away.
22:19
The ones that stood up and went, I can
22:21
make my work 10 times more effective by
22:23
using these agents and this product that
22:25
we're building as well.
22:27
I think it's generally the type of people
22:29
we want to work with.
22:30
So the vast, vast majority of people in
22:32
our company took this and ran with it.
22:34
A lot of this stuff I
22:35
haven't built myself.
22:36
The team, we're building it as we go.
22:38
So it's really, really exciting.
22:40
That's really cool.
22:41
One thing I'm curious about is like, we
22:43
know that Meta is coming out with more
22:45
tools as time goes on.
22:47
We know that AI tools, right?
22:49
We know that the way that Meta serves ads
22:53
with, you know, Gem and Lattice and all
22:56
that shit continues to evolve.
22:59
You know, like what, and you've been in
23:01
the agency space a really long time.
23:04
Like, what do you think is the future of
23:06
what your ad agency agency is?
23:08
This is it.
23:09
I mean, what's the race to the bottom
23:11
there is the way I look at it, right?
23:12
What's the lowest common
23:14
denominator you can get to?
23:15
And that's the individual business, a
23:16
single person running an entire business.
23:19
That is, I'm on podcast here talking
23:21
about it, but that's not
23:22
going to happen, right?
23:23
So where's the low bar in that?
23:25
What's the race?
23:25
What bottom are you getting to there?
23:27
And to be honest, I think it's not as low
23:30
as you might think, because if you think
23:32
back to, you know, just five, 10 years
23:34
ago, Meta have always been kind of trying
23:37
to make it easier for
23:37
advertisers to run ads.
23:39
It's easy to run ads on Meta.
23:41
You can just turn them on.
23:42
You get in, you throw
23:43
your card in, the way you go.
23:45
That's simple.
23:46
Running ads is fine.
23:47
Running good ads is impossible.
23:49
It's really, really difficult to always
23:51
just look at DTC Twitter.
23:52
Look how difficult it is to run ads and
23:54
actually make a huge amount of money.
23:55
We spend literally millions a month and
23:58
actually do it really, really profitably
23:59
for every client that we've been working
24:01
for the last six months and
24:02
being really, really profitable.
24:04
What I would say to you is
24:05
that didn't come by accident.
24:07
The decisions that were made may not have
24:09
had all of the data that the AI may have.
24:12
It may be some software
24:13
metrics and everything else.
24:15
There's a human element to making
24:16
decisions that what AI cannot capture.
24:19
It's just it will never be able to
24:20
because you have to then program it to
24:22
feed in this
24:23
information and that information.
24:24
Look, context windows can get better.
24:26
You can feed more information in, but
24:28
there will always be something missing,
24:29
something there as intangible, that
24:32
limbic system that
24:32
we're missing in an AI.
24:34
But what I will say is when you start to
24:36
think about building an AI for somebody,
24:40
again, the way that we think about it
24:41
isn't necessarily as building a
24:43
replacement for a
24:44
human, as enhancing a human.
24:46
So the people who think of this as
24:47
reductive are the ones
24:48
that are going to lose out.
24:49
The people that think of it as additive
24:51
and kind of increasing where we're going,
24:53
but the people that are going to go.
24:54
Look at the, I think it's NVIDIA, the
24:56
likes of those kind of guys.
24:57
When AI started to come in, they go,
24:59
right, well now we can go and do this and
25:00
this and this, rather than looking at the
25:02
other kind of large Fortune 500 companies
25:05
that just cut half their people because
25:07
they're like, we can just AI this.
25:09
Those are the people that are going to
25:10
lose out versus the people
25:11
who are trying to scale with AI.
25:13
So to answer your question in more of a
25:15
short form, I think it's going to be in
25:17
two years' time, I say that we're going
25:19
to have the same amount of people.
25:20
We're going to be helping a lot more
25:22
clients and it's going to cost them a lot
25:24
less to work with us.
25:25
I think it's interesting.
25:26
Like, I think it's already, it helps you
25:29
unlock so much more in terms of finding
25:33
new places to advertise, spooling up
25:35
landing pages that are better because
25:38
they take every learning you've ever had.
25:40
Right, like Cody Plofker talked about in
25:43
a previous AI D2C WTF episode about how
25:46
he basically, when he builds landing
25:47
pages now, he takes every call he's ever
25:50
had with any CRO firm.
25:52
He took like all these PDFs he's
25:54
downloaded from CRO experts.
25:56
He took the literal entire text of a book
26:02
about a CRO, like from a CRO expert.
26:05
He did, and like, it's like seven things
26:07
and he's 10 for 10 on the landing page
26:12
that he's rolling out, the
26:13
PDP, improving CVR by 5 to 10%.
26:17
So he's increased it by
26:18
almost 100% in two months.
26:20
And like, that's what I'm saying.
26:23
It's like, it's crazy
26:24
how big that can be.
26:26
And I think it's going to, you know, I
26:29
think that we also, a lot of times people
26:31
get worried like, oh, wouldn't the client
26:33
do this themselves as
26:34
an agency or whatever.
26:36
But I think you have to remember, like, I
26:38
could paint my house.
26:39
Like, I, you know, I could plumb, I could
26:41
go on YouTube and plumb my shit, but
26:44
like, I'm not going to.
26:45
Like, you know, because it's not.
26:47
And so people are always
26:48
willing to pay for an expert.
26:50
And so I think that staying on top of
26:52
this stuff, and I'm not saying you have
26:53
to adopt it all or any of it.
26:55
I mean, some of it probably, but like
26:58
being on top of it's massive.
26:59
And being able to say
27:00
like, yes, let's try that.
27:02
Because that's a huge value add.
27:05
If you're able to say, look, we have
27:06
three people working on your account, but
27:08
it's the equivalent of like 15 people.
27:10
Yeah, exactly.
27:11
That's a crazy like value prop, you know?
27:14
The thing I would say to your point there
27:15
is like, yeah, Cody was able to do that
27:17
10 for 10 the whole way up.
27:19
Cody was.
27:20
But John down the street, he's not going
27:22
to be able to do that.
27:23
So it took Cody 10 years of actual
27:26
working on stuff, probably longer.
27:27
Sorry, Cody.
27:28
But longer for actually go through all of
27:30
this and say, right, you know, you've
27:32
learned the basics of it.
27:34
It'll spit it out and it'll be like Will
27:35
Smith of, you know, 2023 AI.
27:38
It won't be the Will
27:38
Smith of today AI, you know?
27:40
So it's like it will spit it out, but you
27:42
still need someone to look at it and go,
27:43
ah, that's crap because I
27:45
need to move that there.
27:46
If Cody spat that in and put all that
27:48
information in and iterated again, 10 for
27:50
10 and didn't touch it, that's when it
27:53
becomes, okay, this is
27:54
autonomous, but it's not.
27:56
Because Cody went in and made those edits
27:57
and did all the changes,
27:58
which is perfectly fine.
28:00
But to your point, Andrew, you could
28:01
paint a house, but you're not going to,
28:03
you're going to pay the guy or the girl
28:05
to go and paint the house
28:06
and go and get it done right.
28:07
Yeah.
28:07
I think it's, I think
28:08
it's really interesting.
28:09
I think, you know, the way that you've
28:11
gone, gone and sort of looked at like
28:13
enhancing the employees, do, do employees
28:16
get bonuses for innovating on AI?
28:21
Like, or how are you incentivizing them?
28:23
Because a lot of times I feel like the
28:24
top down edict is really hard because
28:26
it's like, do more with AI.
28:28
And then everyone's like, when?
28:30
Yeah. Yeah.
28:31
And so you're like, so how do you
28:32
incentivize employees to be like, Hey,
28:34
like if you save three hours of your day,
28:37
like go be with your family.
28:38
I don't care.
28:38
I'm still going to pay you the same.
28:40
Or here's a financial incentive.
28:42
If it's something the
28:42
entire team can use.
28:43
The reason I ask is because talking about
28:45
Kaizen, like when Kaizen came out, the
28:47
process of Kaizen in Toyota in like the
28:50
sixties, they paid proportional to the
28:53
amount of time that it
28:54
saved every employee a bonus.
28:57
They saved like three minutes of time.
28:59
They paid like whatever it was.
29:00
And if they saved 30
29:01
minutes, it was like much bigger.
29:03
And every employee was incentivized, even
29:05
on a micro adjustment to, to do that.
29:08
So I'm curious if you're doing that.
29:09
Yeah.
29:09
I guess the, the, I mean, that worked
29:12
super well because the, the, the, from a
29:15
Toyota standpoint, because the goals of
29:16
the company was to get cars out fast.
29:19
Right.
29:19
And so that was what
29:20
they, they wanted to do.
29:21
Good quality and get them out fast.
29:22
Our goals of our company is one to
29:24
improve the performance of
29:25
companies, not just get shit done.
29:27
Right.
29:27
Get shit done is a function
29:29
of what we're trying to get to.
29:30
So while we try and incentivize, we try,
29:32
we focus on performance of clients.
29:33
We focus on, on happiness of clients and
29:36
we bonus them based on that.
29:37
If they can achieve that in a shorter
29:39
amount of time, great.
29:40
But our focus is on keeping that high
29:42
level, high standard.
29:43
And we pay on a quarterly basis where we
29:45
say, if a client is higher on a certain
29:48
CSAT score, which we would then record
29:50
through Kaizen, which also, by the way,
29:52
takes in Slack communication as well.
29:54
But then we have that and then
29:55
performance of clients as well.
29:56
So as we start to scale the client's
29:58
results, that will then be fed back into
30:00
our, our people that are doing it.
30:02
So they're incentivized
30:04
to improve performance.
30:05
And if they can use AI to do it quicker,
30:07
go be with your family.
30:08
So it's kind of a mixture of both.
30:10
But we want to make sure that our
30:12
incentivization is lined up with what we
30:14
are trying to do as a company.
30:15
And that's, you know,
30:15
create commercial masterpieces.
30:17
That's what I talk about all the time.
30:18
Yeah.
30:19
Yeah.
30:20
Great.
30:22
Thomas, any other questions?
30:23
Oh, I'm just excited.
30:24
And I think this just reminds me, and I
30:27
think in every interview you've done so
30:28
far, AI is great, but you need experience
30:33
and you need some, you need to know what
30:37
you're doing to actually get this to work
30:39
in a high quality way.
30:41
And I, I'm not afraid at all that AI is
30:44
going to replace my team or myself.
30:46
It's just going to
30:47
enable us to help more people.
30:49
So I think it's just really cool.
30:51
One more thing on that.
30:52
And I think this is super cool.
30:54
We built a very specialist agent that is,
30:56
is a kind of a different tact to what
30:58
we're talking about here entirely.
30:59
But we, we call her Mia.
31:00
She's a very special agent.
31:02
And she's separate to everything we talk
31:03
about here in some parts,
31:05
but kind of still connected.
31:06
So Mia is a Google ad specialist.
31:09
What she allows us to do is a lot of the
31:11
kind of the grunt work, the basis work.
31:13
So she runs a service, so she's obviously
31:15
there access to everyone, but she does
31:17
research across all Google accounts to
31:19
understand what a search intent is
31:21
happening across their accounts with a
31:22
very similar level of accounts.
31:23
She'll then bring up kind of search terms
31:25
that are relevant for like when she looks
31:27
at the search volume.
31:28
She'll bring up search terms that we can
31:29
bring into keywords.
31:30
She'll bring those into shopping titles
31:33
and actually be able to work out what we
31:35
need to turn negative in terms of
31:37
non-branded keywords.
31:38
So a lot of the basic works that are in
31:40
there, she's just started two weeks ago.
31:42
We've started running that, but actually
31:44
started to see a measurable increase in
31:46
terms of performance.
31:47
Where that's starting to go is actually
31:49
that's starting to build out that
31:50
autonomous, autonomous buying.
31:52
What's really interesting though, is
31:54
where we're starting to link
31:55
that into our communication.
31:57
So when we're talking about what's
32:00
performing and what's not, that then
32:01
feeds that into the
32:02
direct client communication.
32:04
She's building out Google performance
32:06
through AI specifically.
32:07
So we're just, we're
32:08
hedging a little bit.
32:10
So we're still focusing on ourselves, but
32:11
we're building out platforms that are
32:13
allowing us to actually
32:14
buy autonomously as well.
32:16
Because agentic buying is going to be,
32:18
it's not even six months
32:18
away, I'd say it's three months.
32:20
It'll be big.
32:20
I mean, I just was talking to a colleague
32:22
yesterday, his team, they spend about,
32:25
you know, six million a month, probably
32:27
across the brands they work on.
32:28
And they've been doing agentic buying and
32:31
it's like beating any
32:33
baseline of a human.
32:35
And so he's like, I don't know how
32:36
quickly we're going to roll this out, but
32:37
it'll be faster than I thought.
32:40
So it's interesting, like where we're
32:42
headed with that for sure.
32:43
That's a whole other episode.
32:45
Tris, thanks for joining us, man.
32:46
Pleasure.
32:47
I really enjoyed it, guys.
32:47
Bye.
32:48
High-end.
32:49
Bye.
32:50
We'll see you next time.
32:54
You