Retailgentic, presented by the ReFiBuy Media Network and hosted by Scot Wingo, is where Agentic Commerce meets retail innovation. We help retailers and brands prepare for the future of agent-driven shopping with actionable news, expert insights, dxis, forward-looking predictions, and practical guidance on Agentic Commerce Optimization (ACO).
If the, the retailers choose the
right strategy, they can increase,
like, the omni-channel as well,
not only the e-commerce part.
So I think the contribution of agents
will not be only, uh, like, to the
e-commerce part, which is, will
be key part, but also I think it
will increase a lot the awareness.
Because really the agent, the main, main
thing to your point is they reduce the
search friction, and I think it will
increase more foot to the stores as well.
Welcome to the Retailgentic Podcast,
where we explore the intersection
of retail, e-commerce, and AI
agents, which we call retailgentic.
Our mission is simple: we are shining
a bright light on the new agentic AI
shopping megatrend, which will disrupt
the trillion-dollar digital retail market.
The stakes are high.
Google, OpenAI, Microsoft,
Perplexity, Amazon, Apple, Meta, and
Anthropic are all fighting to be the
consumer AI shopping agent winner.
There are gonna be some tough decisions
for online brands and retailers to make.
What happens to checkout?
How should payments work?
Should you continue to invest
in retail media networks?
What does the world look like when
over half of your traffic is not
from humans, but from AI agents?
These existential questions
are the tip of the iceberg.
On this podcast, we're talking to industry
experts, analysts, and influencers
to get answers to these questions,
identify trends and best practices.
This is Retailgentic, your
podcast partner in navigating
our AI shopping agent future.
And now here's your host, Scot Wingo.
Welcome to the Retailgentic Podcast.
If you are as big an agentic commerce
nerd as I am, and let's face it, that's my
audience, today we have a special treat.
On August 4th, 2025, researchers from
Columbia and Yale dropped this research
report, What Is Your AI Agent Buying?
This is a great report.
It's linked in the show notes,
and I strongly, uh, recommend
everyone listening download it.
The best parts are the conclusions,
which is where you start to
see these screenshots of how
agents are buying things.
So it's very interesting
It's, uh, an amazing report,
uh, but it's more than that.
So it's more than just a report.
It has an entire open source agentic
buying arena or sandbox, uh, an
evaluation system, and then they've
also open sourced all the code and
data that, that they built for this.
So you could literally, uh, if you're so
inclined, um, download this, uh, create
your own sandbox site, point it at your
own retail site, and have a bunch of
agents, uh, work on your site to see
how they do So today on the pod, we
have one of the authors, Amine Alaoui.
He's, uh, on the Columbia side
of things, where he went to
school, and it was his idea.
He has, uh, before, uh, even, uh,
writing this paper, he had left Columbia.
He graduated with a PhD.
Um, went and worked on AI at Meta, also
known as Facebook/Instagram, and then,
uh, actually more-- most recently started
his own company called MyCustomAI.
Uh, the name implies exactly what they do.
They build and train very specific
bespoke AI models for their customers.
Uh, and I found that
conversation interesting as well.
Uh, I'll note that this research uses the
class of agents that are called operators.
Uh, since they wrote the report,
ChatGPT renamed it to agent.
Um, but, uh, the way I think about
these agents are, is their tool set
is, uh, that they use, uh, a computer
in the cloud to do their work.
Um, so that's interesting.
That being said, I think most agentic
commerce that I see happening is gonna
be done by what I would call, uh, this
is what Perplexity, ChatGPT, Copilot,
uh, and Gemini have announced at I/O.
Uh, it's a combination of a
product-focused web search, a product
database, uh, a merchant center usually
that feeds that product database.
Uh, I think eventually we're gonna have
hosted checkouts, like has been leaked
for ChatGPT, um, or agents that are
really good at completing checkouts.
That's like more the Perplexity
way of doing things, and
Google I/O, Google, uh, Gemini.
Uh, that being said, even
with this approach, there's a
lot to be learned from Amine.
Um, the paper, this framework
I think is really fascinating.
Um, and most importantly, the learnings.
I think the learnings cut
across both types of agents.
So for example, um, one of their
learnings is how, uh, the agents
view advertisements, and I think
this holds across both of the
types of agents I'm talking about.
So I think you're really
gonna enjoy this one.
I could've gone literally three
hours on this, so I had to cut it
off at an hour, so it's a little bit
longer, but I think it is worth it.
Now without further delay
is my discussion with Amine.
Enjoy.
Amine, welcome to the
Retail Agentic Podcast.
I was, uh, telling you in the green
room that, uh, I read your paper.
Uh, it's in my…
It's always next to me right now.
I, I, uh, refer to it often.
And, uh, I reached out to you and some of
your colleagues, and I was really excited
to be able to get you on the podcast.
Welcome.
Welcome to the pod.
Uh, thank you very much, Scott.
I'm also very excited about, uh, this
discussion, and I'm v- big fan of your,
uh, podcast, as I mentioned earlier.
Thanks.
It's like the two biggest agentic
commerce nerds got to hang out,
so it's gonna be a lot of fun.
So, uh, tell me, let's
start with your background.
Um, so let's start, uh, you're, uh,
an academic, so let's walk through a
bunch of your academic background first.
Uh, absolutely.
So, uh, so I did my, um, undergrad and
master's at, uh, École Polytechnique.
Uh, maybe to set the stage, École
Polytechnique, it's a French school, one
of the top in the France, and, uh, is
leading in mathematics and also in AI.
Uh, as a matter of fact, uh, the
alumni from the school, for instance,
are the founders of Hugging Face and
the, and the founders of Mistral.
All of them are from École Polytechnique.
So yeah, we, we, we get a lot of, uh,
like strong foundational mathematics,
uh, training that allows us to work on
the field of machine learning and AI.
Cool.
Do you know those guys?
Are they your buddies, your AI buddies?
So, uh, Arthur Mensch, I know him,
uh, like we were in the same batch.
Uh, but we didn't catch up
since then, I would be honest.
Uh, for the other ones,
they are elder than me.
Hugging Face, they were in like, uh,
in a, in, in a, in a different batch.
That's about my ma- masters.
Uh, I'm happy to cover, uh,
my, uh, my, my PhD and yeah.
Yeah, then what brought you to the States?
So you're, you're in
California now, I imagine.
So how did you make the trek
from, uh, Paris to California?
So I stopped by New York for five years.
I did a big stopover.
Uh, when, when I was, uh, doing my,
uh, PhD, I was in the department
called Decision, Risk, and Operations,
part of the Columbia Business School.
Uh, the reason I went there, because
that department faculty is one of
the leaders in taking, uh, real world
applications, model them mathematically,
solve them to have real world impact.
As a matter of fact, um, uh, our faculty,
for instance, uh, helped designing the
Google, uh, like, uh, ads algorithms,
especially the, the budget part.
They have also designing, and they are
very active in terms of, uh, Amazon as
well, especially, uh, the logistics part.
Columbia has a research center that
Amazon Columbia jointly in Columbia,
given the deep, uh, like, uh, relationship
and expertise that Columbia has.
Um- So little bit maybe about,
uh, my own expertise and my
PhD li- little bit thesis.
So I, I, I, um, so I am expert in
what we call algorithmic game theory,
which is, uh, developing algorithms
for when we have lot of actors that
are using different machine learning
models, how can one optimize the, its
own algorithm given that others have,
uh, like their own algorithm that where
the outcome will depend on the others?
So o- one concrete example is the ad
system, where your own bid does not
determine the outcome of what will happen,
but will depend on what competitors are
doing, as well as the platform, what
Google or Facebook also decides to do.
So that's at high level a little
bit, uh, about my background.
Very cool.
Did you, uh, complete your
doctorate and do a, a, a thesis?
Uh, yes, I completed my doctorate
at, uh, uh, at, uh, uh, Columbia.
It took me almost, uh, five years.
Uh, and then I, I defended
my thesis there before, like,
uh, moving to the West Coast.
Got it.
Ca- uh, it's always fun.
What…
Tell us about your thesis.
I know I won't understand it, but I just
like to hear all the fancy terminology.
So in this field where y- your algo-
like the, the outcome of the, uh, like,
uh, uh, your decision depends not only
on you, but depends on others, uh, lot
of algorithms are developed when we have
large data sets, we have infinite amount.
But in practice, that's not the case.
We have less data sets, we have less data.
So my thesis is ab- was
about from the other extreme.
When we have very limited information
about the, our, uh, our objective-
Mm-hmm
… then how to develop those algorithms.
And- That's in nutshell
what my thesis was about.
Ah, got it.
Cool.
And then, um, there's a, a…
For those that are interested
in game theory, there's a really
good movie, Beautiful Mind.
Uh- Absolutely.
Do you know that one?
Yes.
Okay.
Yes, yes, I have.
Yeah, I had a feeling you'd,
you had probably seen that.
Uh, that's j- about John Nash,
um, and the Nash equilibrium.
So, so you're kind of taking that same
kind of setup where I think he, he won
the prize for game theory looking at
multiplayer auctions and, and, you know,
collaborating and competing and things.
So you're, you're taking that,
but then arming them with LLMs.
Is that a fair-
Uh, so I
would- It may not be the exact Nash setup,
but, but similar kind of game theory.
So yeah, indeed you touch
up on different notions.
So equilibrium is, uh, this
concept, uh, that is, uh,
conceptual that Nash has developed.
Uh, in practice, is that, uh, w-
what my focus on and my colleagues,
lots of my colleagues, there are
those that study those concepts.
My focus was, like, how can we
reach that equilibrium in practice?
And typically, those AI models,
ML models will interact together,
uh, uh, before LLM era even people
studied those in, in, in some sense.
Okay.
Got it.
Okay.
So you graduated, uh, we'll,
we'll ca- call you Dr.
Amine now 'cause, uh, I
now know you're a doctor.
So, uh, so you graduated and then, um,
what, uh, what gave you the bug to go to
Silicon Valley and be an entrepreneur?
Uh, so, uh, indeed, so I, when I
graduated, I had like different
offers from Amazon, from Uber, from
different two-sided marketplaces,
and one of them was Meta.
And the reason I picked the team
at Meta there is the reason is the,
the, it was called Central Applied
Science team, and it, it has like
this central place that allowed us
to work with different verticals.
So Instagram ads team, notification
team, uh, marketplace team, which is the
e-commerce part, and other, uh, verticals.
So it allowed me to get lots of
exposure about lot of, uh, uh, systems.
But also, uh, at Meta scale,
there is a lot of impact.
A small change on algorithm can
change the life of lots of people.
That's what really has
driven me to go there.
Yeah.
And it can, uh, also make a
lot of money for, for Facebook.
Yeah, absolutely.
Yeah.
Their last quarter, I think,
was, uh, an illustration of that.
They're really starting to use AI across
all the, the ad, you know, ad creation,
ad optimization, the bidding systems, and
it's really starting to, to bear fruit.
Absolutely.
Yes.
Yeah.
Was there a particular, um,
how long were you at Meta?
So I was there from 2019, and I decided
to leave in 2022, like September 2022.
So, uh, that is a big timeframe in our
world because that's when ChatGPT…
You were probably tracking
GPTs early, I imagine.
But that's when most people became aware
of, you know, the three-letter acronym
GPT, and, and that's when ChatGPT came
out of, it was more like October, but
you were right in the sweet spot there.
Uh, so indeed, uh, so, uh, to give
you the full story, I was, uh,
like, uh, in parental leave at that
moment, uh, uh, actually, so before
I leave, and we were struck in
different trends, uh, at that moment.
And ChatGPT came just across the corner
in November, at the end of November.
And this is where, uh, like, uh,
I saw big opportunity indeed.
Got it.
And then you started MyCustomAI.
Uh, that's correct.
Now I'm the co-founder of MyCustomAI.
Cool.
And then, um, tell us the-- I'm, I'm
a fellow entrepreneur, so I always
love to hear the founding story.
Was this a set of, of friends
you had, and you decided this is
time for us to start a company?
Or, or how did this all come to be?
Uh, so the, the reason is that when
ChatGPT came, um, so I-- to be a, to
step back a little bit, I always had this
desire and passion to go build something
from zero to one, such that it has huge
impact on what we're doing day to day.
My dream is like when I'm 70 years
old, I look back and I see if I left
some legacy that people are, like, uh,
have high impact on them in some…
That is what's driving me personally.
Um, and to do that, like, I was wondering
what do I have as, like, skills in
some, what is, uh, my unique edge?
And my unique edge is developing
these custom algorithms.
And when I saw ChatGPT coming, there
was this, uh, like people were believing
AGI is coming in some sense, and I
always believed, you know, AGI is
coming, and there will be always a
gap between what model provider will
provide and really the end use case.
So this is when I decided that
this niche, Custom AI, is our
sweet spot where we can shine.
Got it.
Cool.
So tell us, so you founded Custom
AI kind of in late '22 or early '23.
Um, what is it-- So you guys have
been at it for two-ish years.
What, what is it that you do
that you can share with us?
Yeah.
So we are, think about us like
an AI agency, so we're a service
company, and we help businesses, uh,
customize AI to their own use cases.
A lot of us today use either ChatGPT,
and a lot of our customers will
believe that it can do anything.
However, ChatGPT is very good,
or LLMs in general are really
good at specific use cases.
So what we offer them, we
offer three services typically.
So the first one is like what we call
AI feasibility study, where we sit
with our customers, tell them where
they can apply AI, which specific
tasks, which one has high ROI.
So this is the first part.
The second part, which is more our
core service, is that we train models
for our customers, then we deploy
them in the customer's, uh, cloud.
And the last part, which is very
critical even in, especially in, uh,
company's cultures, is we, uh, like we,
we are-- we always try to forward look.
So we train our custom-- we do workshop
training where we share with customers
the latest findings that we know of.
We cannot follow everything, of
course, but can be relevant to them
Okay.
So there's a thought
leadership element of it.
Um, there's, uh, you know, helping
them doing an analysis, so kind
of understanding where they're
gonna have impact, and then you go
and build models for, for folks.
That's pretty well.
Um, what are, um…
And I know, um, you know, you probably
don't wanna release specific client
stuff, and I'm not asking that.
What's-- Help our audience understand,
like, what's, what's the type
of use case that, you know…
So we have some things on the
inference side, so you can use
retrieval augmented generation.
You know, my company is doing agents,
which are largely on the inference
side and, you know, maybe eventually
we'll, we'll have some little
pieces where we need to build our
own models, but so far we're just
using the more, more horizontal.
Um, but tell us about, like, what
are some common, you know, things,
flags that, that, you know, uh…
So some of our listeners, for
example, are large retailers.
Is that something where there's a
day where maybe not external for
customers, but internal on their
data, there's some training or…
Like, help us understand, like, when
do you need your own LLM to be trained?
Uh, yeah, so good question.
So in general, like, where we--
like the type, maybe I will
describe type of customers that
we work with as an illustration.
So we work with two type of customers,
either large, uh, like companies
with large datasets, or, or those
that are startup, they are trying to
disrupt like specific industries to…
with AI.
In both cases, uh, what is happening,
there are three pain points that we solve.
The first one is accuracy.
Off-the-shelf models are not necessarily
accurate on the specific internal
datasets that a big retailer will have.
The second critical point
is about, um, uh, privacy.
There are some datasets that you, you
don't want to share with, uh, outside
third party, and you would like the models
to run on your own, uh, like a cloud.
Uh, and the, the, the third reason,
which is also very critical, is costs.
These model providers, they have
their own margins, uh, for a reason.
So what is…
By deploying specific models on your
own cloud, one can save a huge cost.
As a matter of fact, I
can give you an example.
With one of our customer that has very
large use case, we were able to reduce
the cost by ninety-nine percent for,
for that, for their specific use case.
Uh, so it's big ROI for companies to,
if they have many use case that is tied
to their core business, it's critical to
think about using local models internally.
Got it.
The…
Okay.
And, uh, I could ask a
million questions on this.
This will be my last one, I promise.
Uh, I imagine being a Meta person, um,
Llama is, like, one of the core models
you guys like to work with 'cause,
uh, y- most likely you're doing a
lot with open source models, right?
And there's, there's, there's not infinite
of those, but there's, there's a lot.
Um, so is that…
So, so you would take like, let's say
just Llama or maybe DeepSeek or one
of the open weight models, um, and
you, you can then, uh, train parts of
it, uh, you know, based on adding in
the proprietary pieces that, that you
want, and then you help your customer.
Um, they're up-- It's up to them to deploy
it, but I'm sure you help them figure out,
all right, this is a great way to host it.
Let's run it through some tests and tune
it and, and get it, get it operational.
Is that, is that kinda how it works?
Uh, yes, that's high level.
Yes, your understanding is, uh, correct.
So, uh, so- Uh, so depending on
the use case indeed, and for the
chatbot ba- parts, yes, we use open
source models if we need to, and we
can use Llama, we can use DeepSeek.
But think about recommendation systems,
which are critical in marketplace.
LLMs are not good at those use cases.
There are some…
lots of use cases where LLMs are not
the best technology for those things.
And as a matter of fact, uh, LLMs
in terms of research have already
reached a plateau, and people are
looking for new architectures that
will overcome their limitations.
So as a matter of fact, what we start
from the business use case, we try
to find indeed open source models.
Some, some of them that we use is used
called vision transformer, VIT typically,
which is the, the vision part of ChatGPT.
There are some audio models
we use, depend on use case.
So we really look at what's
the business use case, try to
find the open source model.
Then we have a tr- uh, annotation
team internally that will annotate
the data for that specific use case.
We train the model, we do an AB testing
to see the impact, we measure the
impact, but we help them deploy it.
We have an infrastructure engineer
within the team that will integrate with
their own, uh, like, uh, internal team.
So at high level, yes, and uh, but
we don't only rely on LLMs, we rely
on other open source models as well.
Got it.
Okay.
So there's multimodality, and so
depending on the mode, which could
be visual, audio, some combination
of those things, you're gonna use
different models and- Absolutely.
Okay.
Okay.
I got it.
That's correct.
Very cool.
Well, that, I think that, uh, uh, I
think you've proven to our listeners
and, and you know, I already knew
this, but, uh, that you, you know a bit
about this stuff, so, so that's great.
Um, and then what we wanna spend the bulk
of our time talking about is the paper you
guys published about a month ago, and it's
called ACES, the ACES framework, A-C-E-S.
Uh, this will be available
in the show notes if anyone
hasn't had a chance to read it.
It is, it is very meaty, um, and I've
really enjoyed it and, and gone back
to several different aspects of it.
Um, and the full title is, so,
uh, it was published by you and
several of your colleagues at
the Columbia Graduate School.
Um, and then one was from Yale.
I guess he's like- The, the, the
one crazy friend that went to Yale.
Uh Yes.
"What is your AI agent buying?
Evaluating-- Evaluation
implications and emerging
questions for agentic e-commerce."
So, so let's start at like, what, what was
the inspiration for doing this project?
Um, how, how did you guys
come up with this idea?
Does this go back to your time, um, in,
you know, working on your PhD or, or-
The-- So the story is as follows.
So, uh, in November, um, twenty
twenty-four, Anthropic released the,
the, the competitor use agent model.
Uh, and we used that for one specific
vertical, and we developed a very novel
use case that, uh, that is not related
to e-commerce, but it's attracted Disney
and PayPal with an-one of our customers.
So we were really understanding deeply
the how, what we call GUI agent works or
agent that use the browser work very well.
And then in January twenty-third,
twenty twenty-five, ChatGPT released
their own version called Operator,
where you can chat with it and to
book a hotel or plan a trip, and it
will go, like, do that on your behalf.
And when I was reading that article,
something caught my eye, which is
the logo of Instacart being a partner
of OpenAI And that sparks an idea.
Given my expertise that I developed
at Meta, and the fact that we know
Instacart were, were not profit- was
not profitable till they introduced
the ads, sponsored product.
I was wondering, is it rational or
is it the best decision for Instacart
to partner with OpenAI and open
their website to them in some sense?
Maybe that is backfiring on them.
Maybe now agents might
not click on ads anymore.
And that's really what has
driven this initial idea.
So what happened after?
I called my alma mater, like my previous
department that are expert in this,
uh, designing the algorithm, and they
said, "Can we partner on this, guys?
We are like, our focus
is not purely research.
We are very applied."
So…
And that's how this idea came,
and then we formed team together,
and this is how the, the, the, the
full project starts since then.
Interesting.
Yeah.
Did it surprise you that, uh, OpenAI
ended up hiring the CEO of Instacart?
Uh, that's
Yeah.
Pretty wild,
huh?
Indeed.
Indeed.
Yeah.
Yeah.
You called it.
That's good.
Yeah.
Um, okay.
Let's, let's kind of, um, instead of
building up to the conclusion, let's
start with like some of the things you
learned, um, with this, and then we'll,
we'll kind of dig into how you built
this framework and, uh, and whatnot.
So what, what were some of the s- you
know, maybe, maybe for, for the audience,
set up the experiment, um, and then walk
through some of the big things that you
guys learned when you, when you ran this.
So the, the, the experiment is as follows,
is that we created like a website that
looks like an Amazon website, where
we, we give, we give the agents access
to this website, and the agent will
intru-- like search for specific product.
Let's use the example of fitness watch The
agent will search for fitness watch, and
then we display eight products in a grid.
So think of two rows by four
columns, uh, in this case.
And then we ask the agent to
pick one of those products.
Here, like, pick will
be more nuanced later.
And then what we do, we observe its
choices based on the product attributes.
Talking about product attributes,
one of them is the price, number
of reviews, number-- uh, like the
average rating, tags, et cetera.
And then what we did is that we
varied those product attributes as
well as product position in the page,
and then we collected the different
choices of these, uh, uh, agents.
So that's at high level how
we designed this, and then
we move to an analysis part.
So the main implication at
high level are as follows.
Model, uh, like we…
First thing implication is that the
model, uh, like, uh, uh, um, uh,
quality is not hundred percent there.
What I mean, model are not fully rational,
and I will go into details later.
Second point, which is, I think,
very critical, MCP is not enough.
The protocols that people are advocating
for are not enough in e-commerce,
meaning that we don't believe headless
websites is the future The third
thing that implication at high level
is that there is a need for different
agents in the ecosystem, and we
believe that will be the future.
And the fourth and last implication
is that there is still, there is
a need for alignment between the
agent and the buyer preferences.
So these are at a high level.
I'm happy to go into de-
to the details of each.
Yeah, let me restate this to make
sure that, uh, everyone understands.
So, um, so you used off-the-shelf,
um, you know, agents.
So ChatGPT, it was called
Operator at the time.
They've since rebranded it Agent, right?
Um, uh, and so, so walk us…
So you did that one.
Did you do Claude?
Uh, so we did indeed.
We did the leading provider.
We did ChatGPT, we did
Claude, and we did Gemini.
So, uh, in-
Gemini.
Yeah, yeah, yeah.
Yes.
Uh, okay.
So you, you used the three, three agents
like any human would use, but instead
of using Amazon, you built kind of a
fake Amazon, and that's the experiment.
So that enabled you to-- And these
agents would go shopping in that,
that kind of controlled environment,
and then you could tweak things.
So, so some of them would be sponsored,
some of them wouldn't be sponsored.
And then you could see how the agent
behaved, um, to get an idea for…
And then once we understand
that, then we could back into,
well, if you're a merchant, you,
you may wanna do A, B, or C.
Or, you know, if, if you're really
focused on the buyer experience, here's
where these are deficient in some ways
that, that maybe if you wanted to compete
at, you know, A, they should fix that.
Or B, if you wanted to compete
with them and have your own
buying agent, this gives you an
opportunity to, to be differentiated.
Okay.
Got it.
Uh, uh, let's start with, uh, you know…
Well, uh, in AI circles,
you said MCP is not enough.
I've, I've believed this for a while,
and I haven't, like, said it as, as
kind of brutally honestly as you have.
I just, I just feel like it's too
early, like having been around
technologies for a while, the…
One of the first things out
is usually you realize is not,
like, the best thing, right?
Like .NET
programming, that was, like, the
big thing, and now no one does it.
Um, uh, or, you know, I'm sure the
Assembly people thought, "This is
the best thing since sliced bread.
There'll never be anything
better than Assembly."
And now hardly anyone does Assembly.
So it feels like MCP
is Assembly for agents.
It just feels, like, a
little too low level.
It lacks a lot of discoverability,
auth- you know, authorization and,
and a bunch of things like that.
Um, what are…
Say more about why MCP is not enough
and, and what you, what you believe.
Uh, so the context of MCP, uh,
like we start with the console.
MCP is a model context protocol,
and it was meant to allow, uh, in
one agent to call different tools.
And then later people discovered,
hey, we-- can we use tool, this tool
communicate between different agents?
And then now they allow the, like, using
the same framework to call other agents.
So l- what I'm trying to say, the
first, uh, limit, like context is
that it was meant for tool calling.
The second thing is that it
was meant for, uh, mainly text.
Mainly text meaning, uh, like MCP, what
you send to the server is a blob of JSON,
and you receive blob of JSON And LLMs,
uh, which maybe help in the third point
of context, which might help is that like
agents, when you open them, typically
open source models, they have two
brains, one for text and one for vision.
There, there are some approaches where
try, try to fuse them, like Chameleon
by, by, by, by Meta, but typically there
are two different independent brains.
And then each input depend on modality
will go through different almost
process, then at the end they get
merged or in the middle, depending
how the model is architected inside.
So MCP really focus on the textual part.
Does it make sense?
So that, that's the context.
Now, what about the facts in factual?
What we see in our paper, the vision part
is a critical piece of decision-making
of this buying or agents in general.
Relying purely on text is
almost missing an opportunity.
We don't know yet the lift how much
it is, but the retail industry,
the margins are very slim.
Any lever that can help push or
monetize for the platforms and the
sellers, I think should be emphasized.
And that's why I was very bold
and said MCP are not enough.
Yeah.
So some people would say MCP
UI solves that, but do you…
Does that, does that solve
it or does it still, still…
It, it only solves one
piece of the many, the…
It's, yeah.
So I think MCP, uh, like, uh,
like, uh, like th- there are some
variants of MCP that can solve it.
For instance, what if we pass
in the MCP the image, for
instance, as a part of the MCP?
That can be solve it, but
it will be not efficient.
It was not in the first…
When it was designed the first day, it
was not designed with this perspective.
What I'm trying to say is that we need
to go to first principle and rethink.
I think the closest protocol to, uh,
that can be used is not MCP, it's
agent to agent by Google, the A2A.
That is closer to what, uh, e-commerce and
retailers need rather than, uh, relying
on, uh, an MCP which is like a black box.
Yeah.
I gotcha.
Let's go back to your original premise.
So your original premise was, is
Instacart host, uh, or Instacart ads host?
Uh, um, is it, is it host?
So, uh, you know, is…
Do agents look at advertising or
do they just totally ignore it?
Or, or do they, you know, when I'm
a human buyer, I've learned most ads
are spam, therefore I don't click
on a lot of sponsored listings.
And do LLMs have that same behavior?
Uh, good question.
So what we find indeed, uh, like, uh,
agents also penalize, uh, ads, meaning
that they tend, if you have a sponsor
tag They tend to, like, ignore you.
However, what we find is that we
find the new levers to do sponsoring.
Instead of showing them a tag,
it's better to show them in
specific position in the page.
That is-- actually boosts your, uh, like
your click-through rate in some sense.
So in other words, we discovered indeed
they behave like human in some sense.
They ignore sponsor tag, but
they have different levers that
a retailer can pull in order to
boost specific products or segments
that they would like to advertise.
Yeah.
I gotcha.
Okay.
The…
Okay.
Uh, you also said there's a need for
different agents in the ecosystem.
What-- Say more about that.
So I feel like the main focus on the
industry, I might be wrong, is that
it's more focusing on what we ca- I'm
calling the buying agent, meaning agent
that will go search for your product,
help you maybe like buy it typically.
But I think on the other side, there
is a need of an agent that up the, the
extreme on the, the other side of the
marketplace or the seller, not Amazon,
I'm thinking about the, the company
that's selling mouse pad, that they need
their own agent to update in real time,
think about PDP in real time based on
the buying agent that within the page.
So in terms of buying agent,
there will not be hundreds.
We know the main providers.
And what we show in our paper is that
there is a potential for real-time
updates of the PDP based on the buying
agent when we upda- So that's new op-
monetization opportunity that I feel
there is less discussion about it.
So in my corner of e-commerce
I come from, which is the
marketplace world, uh, you end up…
A lot of them started as what we'd
call comparison shopping engines.
So, so they realized this problem
early on and, uh, Google Shopping's
kind of the last one around, but there
was like, there was a day where we,
at my previous company, we supported
like 25 comparison shopping engine.
Yahoo Shopping, My Simon, CNET
Shopping, all these things.
Um, and they were all, they
all had their own engines.
Um, and they had a merchant center.
So, so because of the PDP being hard
for any crawler to really get all the
information on, um, and PDPs notoriously,
they'll have a lot, lot of JSON on them,
and, and they're hard, they're hard to
kind of pull off any graphic elements.
Um, you would basically
convert it into data and, and
send it to a merchant center.
So, so it's public information that
Perplexity has a merchant c- center,
and then so does ChatGPT, and then it's
rumored ChatGPT is working on a checkout.
Um, is that one way to kind of solve
this problem, uh, is go to the retailer
and say, "Hey, we could always kind
of crawl your site or, or use an
operator-like thing, but if you just
gave us perfect PDP data, then that
would be much better and cleaner."
Oh, I see.
Oh, I see.
Uh, so that's, uh, will solve, uh,
uh, like partially the problem.
Uh, but, uh, how to, to
phr- phrase my answer.
Um, I think what…
Let me explain experiment then.
Uh, so the experiment that we run in our
paper- Mm-hmm … if it's fine with you.
Yeah.
Yeah.
Yeah.
So the experiment that…
Uh, I will give you two observations.
So we observed that, for instance,
I think it was Claude model, if you
had the word ergonomic in the product
title, you will be considered in
the consideration set, meaning that
you will be picked at least once.
If you didn't have the word
ergonomic, the mo- the…
Your product is completely
ignored, for instance.
Uh, uh, that's one.
Two, to emphasize this, what we did,
we took product attributes, and as the
agent is loading the page, like the
buying agent, not the selling agent.
The buy…
We know its identity.
Let's say we know it's ChatGPT.
We know its sensitivity
to different details.
We then adjust in real time
the title of the products.
So we, uh, so in some sense, we're
like, what you said could be a solution.
Yeah.
You just send a feed to Claude to-- that
has, uh, ergonomic in it, and then your
one to Ch-ChatGPT says, uh, you know,
uh, uh, amazingly wonderful or, you
know, whatever, whatever the, the right
phraseology is that each engine prefers.
Yeah, you can, you can kind of
achieve the same thing, but-
That, that's correct.
Yes, yes, that's correct.
Yes, uh, that's another
solution, assuming that, uh…
Yes, that's a good point to, to, to-
Yeah
… put away from
that.
Someone should, uh, start a company to
do that, but that's just a crazy idea.
So, yeah.
Yes.
I'll take that one.
You, you do what you do.
I'll, I'll…
That's my role.
Yes.
So, so I have this five levels of,
uh, a-autonomous, uh, you know, of,
of agentic commerce, and in there
I start to kind of pick away at…
I don't say it the same way that
what you're-- what I think you're
saying w-between alignment of the
agent and the buyer preferences.
Because at some point the, you know,
for the, for the retailer to really
provide the buyer the best thing, they're
gonna want more information, or there
may be a little negotiation, right?
So, so, uh, let's say you're gonna go
buy, um, like one of the most popular
loyalty programs is the Starbucks Stars.
Everyone knows that one, right?
So you may say, "Hey,
welcome anonymous buyer.
If you'll tell me Amines…"
You know, if you'll tell me your buyer
who happens to be Amines, um, you know,
a little bit of more about them, like for
example, are they a Starbucks member, then
I could probably offer you a better price.
And then they say, "Oh yeah, he is.
Here's his Starbucks loyalty program."
And then the s-system says, "Wow,
he's like right at 100 stars.
I'll go ahead and offer this one
for free," or something like that.
You couldn't have that experience unless
there was some negotiated exchange of
information about the buyer, right?
Uh, and that one-- in this
example, it's a loyalty program.
But I, but I do think we're gonna
get into these advanced phases that's
way more personalized and way more,
um, a lot of really interesting
optimizations around pricing and
loyalty and, and even like group buying.
We've had this promise of group
buying in e-commerce forever.
There's been a big startup every
like 10 years that goes and raise
$20, 40, $50 million and then kind of
doesn't make it because it's really
hard to get consumers to spend time
on, on all this group buying stuff.
Um, is that what you mean
by more alignment between
the buyer and seller agents?
Is like this, you know, learning
more about each other and being
able to kind of trust each other and
exchange a little more information
to make a better transaction.
Uh, so what you described is l- a
little bit down the funnel in some
sense to what I was describing.
Yeah, higher up is like just between
the buying agent and the customer.
Does the buying agent understand
what the intent of the customer is?
You know, sees…
Can I, for instance, the question or
what we see is that what you, what's
best products for you is not what's
best pro- products for me in some sense.
And c- how can we tune these models, uh,
like automatically to understand your
preferences on the buying side before
they go negotiate with the selling side?
There, there was a paper by Amazon,
I think last year, they, where
they developed like a, a prompt
eliciting, like where they will
ask you questions till you answer.
But I'm wondering if there is something
more systematic, like think of like
the most extreme is that each customer
in the world has their own agent
in some sense, uh, that understand,
a twin in terms of preferences.
That's the, uh, the extreme.
But what you mentioned is also
very relevant, which is very also
interesting point, which is how to
incentivize sharing information.
And I think like there are
two high-level mechanism.
One is through incentives, but the
other one is by blocking access, and
that's what Amazon is trying to do now.
If you don't reveal your pro…
Like your, your, your, like who you
are through our specific access, you
will never get access in some sense.
So I think there are two ways to do
that, and either by incentivizing
if you like, uh, for a medium, like
if you have less negotiation power.
But for large traffic like Amazon
platform, it might be forcing
the others to go through specific
protocol to get that information.
Hmm.
Okay.
Uh, so one, one of my theories on, you
know, back to your, your side of it, which
is how does my agent learn more about me?
Uh, I've been using Comet for a
while, and it knows a lot about me.
So the, the agents have, you know, the
models have memory now, which is one way.
So, so I've told ChatGPT, "Here's my
shoe size, my, my pants size," you
know, uh, you know, most of my kinda
like things that, that you a- get
asked for when you buy something.
And it, it automatically is injecting that
when I shop, which is pretty, pretty cool.
Um, so that, that's one way is you, you
basically have some memory and you trust
it and, and you tell it more about you.
But the ultimate, the ultimate form of
trust is giving it access to the, the
kings-- the keys to the kingdom are your,
your email and your calendar, right?
Because it can learn a lot about you.
Email gives you all past
purchases 'cause there's always
an email kind of chain there.
So, um, or even in the browser, like,
like I can log into to Amazon in Comet
and then say, "Go reorder something,"
and, and it also solves the blocking
problem because Amazon can't block
the browser 'cause it blocks me.
'Cause at that point, you're,
you're kind of, you're on the edge,
you're not out on a server, and
it's very hard for it to detect.
There's some little fingerprinting
you can do, but, but it's very,
very hard to con- detect versus
like Operator or one of those.
So, so maybe that's one of the ways
we solve that is just like getting
the agent into your browser in, in
a trusted context is a way to do it.
Uh, absolutely.
That's one way, and that will
solve when you know exactly
what you are looking to buy.
There are many times when you go,
let's say, to shop something you are…
You don't have upfront what-- You don't
know upfront what you are searching for.
And will this-- The one fundamental
question, what if the buying agent will
go start browsing the website and they
implicitly understand your latent need,
what I call latent need, meaning those
that you didn't express explicitly, and
how to come back and convince you as
it was convinced through the process.
I got it.
Okay.
Yeah.
This is where beauty products is a
really good one because, um, you and I,
uh, I won't speak for you, but I don't
spend a lot of time on beauty products.
But, but a lot of women have a color
palette, and then they have kind of like
a season of their color palette, like
a spring, winter, fall kind of thing.
And, you know, you may not tell
the model that, but if it can kinda
like start to infer it from the way
your either past purchase history
or the way you're shopping, it can,
it can start to recommend products
that are way more in your sweet spot
than others that, that may not be.
Um, cool.
So, uh, what are some of the
other things, uh, you know…
So, so we spent a lot of time
in e-commerce with these…
Amazon's been so successful with
advertising, largely sponsored
listings, um, uh, in, in…
It's called retail media networks.
I'm sure you've, you've run across this.
Uh, does this mean as more and
more of these, these types of
systems get used and the ads aren't
really effective, do you think…
You know, some people have said
that this is gonna be one of the
things that could kill that model.
Um, that's, that's-- Let me
just ask that question, and
I've got a bunch of follow-ups.
Uh, so I don't think it will
kill that because, uh, the big…
I think people that are postulating that
they believe that it will be mainly,
uh, MCP, and you can access the full
catalog through the MCP, so meaning
that that's the biggest assumption.
But I feel like through the paper
that we find is that, uh, browser
will still be a thing in some
sense, and, uh, so that's, uh, one.
Two is one of the thing that
we, we, we show in the, in the
paper, uh, and implications that
we believe, like in, like in ads,
there is real-time bidding, correct?
Or real-time ads.
I strongly believe there
will be r-real-time agentic,
uh, ads in some sense.
Uh, that's my belief, and
that's my vision of the future.
Yeah.
This ties back into your, your, your
thesis and your, your, you know,
everything you learned at Columbia, right?
Um, yeah.
The…
Do you think there'll be two ad networks,
a human one and a machine one, or agent?
Uh, good question.
Uh- 'Cause you would
bid theor- you would probably bid
different amounts for each one, right?
So I worked very closely on the
ads system at Meta, very closely,
uh, especially on the auction side.
Uh, there will be some questions
there is like, uh, indeed, like
should we treat them differently?
That's one question.
The second question is how, uh, like
one fundamental question is like since
it's two-sided marketplace, what--
how to make the marketplace grow.
Here it will be three-sided marketplace,
almost like ad-- like advertisers,
users, and the agents in some sense,
and how to incentivize three parties.
I think it's, uh, uh, like, uh, there
will be a lot of novel questions, but
I believe that personally, initially,
they will, uh, treat them the same.
If there is practical need to
differentiate them, then maybe
there will be the like-- then
one will think about separation.
Is this the end of websites?
Uh, is, uh, uh, uh, agentic commerce…
And, and, and so the logic path that,
that I get a lot of times is, um,
and a lot of it stems from this idea
of ChatGPT having a checkout, right?
So, so you, you used operator,
which is now called agent.
Another path is you can now ask ChatGPT
web search, um, and it's getting smarter
at products, and it'll actually insert
a UI element we call product cards.
So, so there's kind of a question there
of, um, you know, if-- So then you
get a product card, and then you, you
pick-- you, you, you see the merchant's
offering for that product, and then
you, you click one, and you go off.
Uh, so Perplexity has an agent
go complete the checkout.
So right now, ChatGPT is what I
call level two, and Perplexity is
level three doing the checkout.
So, so I think ChatGPT may skip that
part and go to just hosting a checkout
and then running it, uh, right there.
And in that world, I can s-see
the argument that maybe you
don't even need a website.
Um, I do think Perplexity will
eventually get to that same phase.
So that, that's the argument is,
um, kind of a different system than
you used, uh, in, in-- when you
use the, the operator-like ones.
But, but, but do you feel like there's
some risk there ar-around website use?
Uh, so there are like, I think
we, uh, like I think website
will not be dead to answer your
question, and the following reasons.
So indeed that the user interacting
with the-- to buy, make a decision
of buy might be through a chat and
not anymore through the website.
But I still believe that once the agent
instruct is instructed by, to search for,
uh, by the user to search for a product,
they still might need to use the websites
instead of using, like in, to your point,
when you click on search, uh, like search
now in ChatGPT, maybe instead of using
purely the API, which is equivalent of
MCP, it can go and browse as what we
do as a human on our behalf, then come
back and show these UI elements to us,
uh, in some sense, like what, uh, uh,
ChatGPT is doing on Shopify, for instance.
So indeed, maybe between the user
and agent, the browser will not
be there, but I still believe that
agents still need the browser.
Got it.
One another-- There's a bunch of startups
that are trying to solve some of the
problems you identified by saying the
reason it's hard to do this is 'cause
you have Claude, OpenAI, uh, and
Gemini, they have to be so horizontal.
So they have to be good at
shopping for roller skates, auto
parts, um, you know, B2B cogs and
widgets, and, uh, fitness trackers.
So, so that's really hard.
But what if you had vertical ones that,
you know, let's say there was an auto
parts, um, focused agent and a fashion…
The most popular is fashion.
There's probably like six
fashion agents that I've seen.
Um, we had someone on that was just
looking at gifting as a category.
Like, you know, what if you had this,
what if you had this agent that learned
everything about you, your family and
friends and all their, their holidays
and upcoming events and, and their
preferences, and it became kinda like
your agent for helping with gifting?
Um, do you think that solves some of the
problems you guys identified as if it,
if it goes kind of vertically focused?
Uh, good question.
I think, uh, the, like, the,
the, the future will be almost
in between, between vertical and
horizontal for the, for the following
reasons based on the, our findings.
So for maybe small ticket items
that are routines in something
that we buy every day, maybe
like horizontal agent is enough.
But think about on the other extreme,
big ticket item, let's say car or
house or some, some, like, luxury
products, maybe vertical agent will be
the, the, the like the leader there.
Now, what will be the
equilibrium in the future?
Who will be leading?
I think they will coexist, uh, I believe.
That's my, uh, belief.
Not only did you guys write this
paper, but you've, you've basically…
On your website, there's a whole Git where
you can go see you've made the source
code generally available and the data.
Um, so people can…
And you-- It's actually set up
almost like hugging, hugging,
little baby hugging face.
I get that it's not hugging face.
But it's set up that now, you know,
because you've open sourced it,
people can go and say, "I wanna run,"
you know, let's say some of the,
some of the vertical fashion things.
They could set up a little fashion
site and see how they do against the…
They could com- they could pit in an
arena-like setup their, their verticalized
agent versus the horizontals, for example.
Um, have you seen people start to use it?
The fun thing about open source things
is people do the craziest stuff with it.
Have you anything you can sh-- I know
it's only been 30 days, but any feedback
or anything interesting people have
started doing with what you put out there?
Yeah.
So the first company that started using
our framework is Amazon, by the way.
Okay.
Okay,
yeah.
So for, for- They're using it to
block stuff.
Yeah, they're like, "We're
gonna use this to block stuff."
Uh, actually I t- I talked to one
of the, the principal people there.
They re- like we connected and
the, our paper shifted the way
they are looking at things.
It's, it's like, uh, it The other
company where we heard the positive
feedback, uh, is Walmart Labs as well.
They thanked us for open sourcing,
uh, like our, uh, like our code.
There are other small
companies reaching out to…
Like, like there was a startup
that reached out to us saying,
"Hey, what you did is shifting
our vision about the future."
So we are glad.
So in some sense, like, uh, we
heard very positive feedback so far.
And the reason we did this because,
uh, we consider ourself the closest
competitor to our own, the Mercatus
MEI is, uh, we think about OpenAI,
like, uh, and other companies.
And the way they got traction and scale,
which is the phase we are at, we are
scaling, is by, uh, open sourcing things
so that we find these forward-looking
partners that we can work with.
In some sense, it's true that
that is bringing a lot of value.
You can, as you mentioned, retailer,
they can plug their website to
the framework and see how agents
are shopping on their website.
They can use it to see how they can
shift the PDP, for instance, how
to increase the click-through rate
for specific products, et cetera.
So I think there are lots of
opportunities, and I'm very excited to
see what people like, uh, will use it for
Got it.
Uh, you know, uh, a reta-- a use
case could be if you're a retailer,
you could download this, um, and
you could create your own faux
website that looks like yours.
So, so you can take what you guys
built, which has got kind of an Amazon
flavor to it, um, and inject your own,
you know, um, you know, put, put your
retail name here, whatever it is, and
then see how the agents react to it.
You could even point it to
your own website, I imagine.
Can you…
Uh, there's no-- As long as you're not
blocking the bots, you probably could.
Yeah, I imagine there's a way to do that.
Uh, indeed, so the little bit
backstory started with the Amazon
website initially before created our
sandbox, but we didn't have control
over, over the product attributes.
So to reverse engineer how things,
uh, are, like agent things.
But indeed, you can use
this on any websites.
Yeah, you had to use a sandbox.
Yep.
What are…
You know, so, so we have a lot of big
brands that sell direct and retailers
that, that listen and watch this show.
W- what do you think the paper reveals?
Like, what are some of the actionable
things they should be doing?
You know, one of the things I always tell
them is I, I get called into a lot of
these meetings with some of the execs,
and they don't re- realize they're using,
um, Cloudflare and they're just blocking,
indiscriminately blocking all bots today.
And they say they do that
to protect their data.
I'm like, "Your product datas should
be, like, freely available because
you're not gonna sell anything
until you give that away free."
And, and, you know, over time,
they, they warm up to that.
I think there's this fear that the bots
are coming in and getting, you know,
some personal identifiable information,
but they're, they're just trying to, you
know, look for the product on the site.
They're not doing anything
a human wouldn't do.
Um, and certainly, there's some
nefarious ones, but you can
whitelist the, the, the, the…
There's not a giant list
of, of bots to let in.
So that, that's, that's one of my
takeaways, but, but I'd, I'd love
to hear, uh, you know, what are
the things you learned from this?
Uh, I, I completely agree.
I think, uh, retailers should embrace
this wave, uh, in some sense, uh, because,
uh, like they, they need to act now,
that's my feeling, in a regulated way.
What does it mean?
So typically, like, uh, when the
agent does have well intention,
they will self-declare themselves.
They will say, "Hey, I am ChatGPT.
Hey."
Of course, there is a- Small fraction,
some agents might misrepresent themselves,
but the opportunity outweigh the c-
like, the, the risk in this case.
That's my own belief.
So that's the first recommendation,
is like to, to, to allow those traffic
and understand them and see what are
quick opportunities you can monetize.
That's the first recommendation.
The second thing, I see some, especially
like, uh, for instance, Shopify, they
publish their MCP, that's great, but I
feel like l- lots of companies are in
this mindset of, "Oh, MCP, it's just MCP."
It's not only about MCP.
Headless is not the future.
That's what I…
It's not the only future.
There's another future of
browser that already exists, and
that's the quickest, uh, ROI.
And, uh, the last thing that I recommend
is that, to your point, these models are
not like humans or where you need to do
user survey to know their preferences.
You can already know how
they will shop the…
on your website and understand
their behavior so that you can
monetize and be ready for the future.
One way is through, as you
mentioned, through our sandbox.
I, I think I highly recommend that
developing deep understanding of
these agents will help decision-making
makers go the right direction.
Yeah.
Throw, throw your site in- into your
sandbox and see if the agents puke on
something and can't find, you know,
a, you know, a product or, you know.
So that's like the top level, and
then you get more and more refined.
So you can kind of start high level
on the funnel, like, you know, help me
find a, a fitness tracker or whatever.
Make sure it's at least finding
your search box and doing the
right things, and then you can get
way more refined over time too.
That's where you'll start to learn where
the importance of the word ergonomic
comes in and these types of things, right?
So you can, you can really use this
system as a way to do some optimization.
I think.
And, uh, the best part is,
like, it's basically free.
So, so, you know, thanks for putting that
out for people to, to experiment with.
Um, I know we're getting tight on time.
A couple just quick ones that I
was gonna sneak in here before,
before I have to let you go.
Um, the, the CEO of PayPal has
said that by, uh, 2030, 25%
of e-commerce will be agents.
Do you think that's spot on, low, high?
Where-- Do you have a sense for that?
You're, and you're a technologist, so
usually I wouldn't ask this, but you have
a really good feeling for, for the bigger,
the bigger aspect of things as well.
Uh, so I, I think he's spot on in
the sense that it will become a main
channel, one of the channels that
e-commerce, uh, will get their traffic.
Will it be 25%, 30%, the exact number?
Uh, I, I would…
We have to say, but I think it will
be not trivial amount of, of traffic.
Yeah.
There's another kind of interesting,
y- you know, data point out there.
So there's a Wall Street firm,
Bernstein, and they say, if you
look, um, e-commerce is growing 6.8%
year over year.
Uh, and I've been in the industry
since the early days, and it used
to grow 20, 30%, and it just feels
like it's really slowed down.
Uh, I think it's due to, like, lack
of exciting new buying and ads and
that kind of stuff and, and just a
overwhelming amount of, of inventory.
Um, so their theory is that, uh,
agentic commerce will increase that
by one and a half to two and a half
percent, which isn't a lot as a
percentage, but it's against that six…
You know, it's on, it's mult-
It's, uh, additive to the 6.8.
So when you add two to 6.8,
it's actually a big percentage
increase, uh, of e-commerce, and
then we're multiplying all these
things times trillions of dollars.
So, so it actually ends
up being big dollars.
Um, it's like at Meta, if you change,
you know, if, if the auction gets 0.1%,
0.01%
more, like, you know,
money falls out of the sky.
Uh, do, do you, do you think this
will incrementally, um, accelerate
e-commerce, the, uh, agentic commerce?
I th- I think it will in,
uh, uh, increase two things.
It will increase the e-commerce,
but also more foot to the, the,
the stores, like the omni-channel.
I think it will bring…
If the, the retailers choose the
right strategy, they can increase,
like, the omni-channel as well,
not only the e-commerce part.
So I think the contribution of agents
will not be only, uh, like, to the
e-commerce part, which is, will
be key part, but also I think it
will increase a lot the awareness.
Because really the agent, the
main, main thing, to your point,
is they reduce the search friction.
And I think it will increase
more foot to the stores as well.
Yeah.
Say a little bit more about, like,
give us a use case just to, uh…
I, I understand, I wanna make
sure, um, the audience understands.
Yeah.
So let's say I'm, I'm looking to
buy, uh, sunglasses or maybe, um, I'm
just trying to, to find an example.
Sunglasses that have specific
like, uh, uh, like a glass because
my eyes are like, uh, have some,
uh, sort of, uh, uh, sensitivity.
Maybe I don't…
This is very niche, uh, like glasses,
and I don't know anything about it.
So I will go, go to ChatGPT,
ask about those, uh…
Like, I will describe the, uh, my
eyes, like condition, et cetera.
And then it might recommend me to ask
the optician, for instance, or the,
the, when I go to the store, ask about
specific brand that I strongly believe
that will be good fit for, uh, my,
uh, my, my, my, my, uh, my use case.
This is just…
I'm came up with this example.
There might-- There are many other use
cases where it will help the people
before they go to the shop understand
better what they are looking for,
given their own personalized needs.
Awesome.
Cool.
Well, uh, I could talk another hour,
Amine, but I wanna be, uh, uh…
If I'm, if I'm good with your
time now, you'll come back on.
So, uh, I wanna be, uh, very, uh…
I know you're a very busy person, really
appreciate you taking time to come on.
If, um, if people wanna learn more about
what you and MyCustomAI are up to and,
and these types of things, is there a,
a, are you big on, uh, X or LinkedIn?
Or should they just go to MyCustomAI?
Where can people find you to learn more?
So the best way is actually the
way we get, uh, our even customers.
We receive lots of inbounds
through our website.
And also, like, um, my LinkedIn, uh,
people feel free to reach out to me.
I would be happy to, uh,
to, to contact them back.
Awesome.
And, uh, we'll put links to all that
in the show notes and also to the
paper, which last time I'll say it,
ACES: What Is Your AI Agent Buying?
Evaluation Implications and Emerging
Questions for Agentic E-commerce.
Thanks to you and your colleagues
for writing this paper.
It, it's, uh…
And, and making everything open source.
I think it's a huge, um…
You know, you, you have-- I don't know
if you'll remember when you're 70, but
you've made our world a little bit better
by doing it, and we really appreciate it.
Uh, thank you very much,
Scott, for having me.
I really enjoyed our conversation.
Thank you very much.
Follow, rate, and share to stay up
to date with Retailgenic, your guide
to the future of AI shopping agents.