00:00:00
Not only are we seeing AI being used to find out things about our work or our studies,
00:00:06
but it's increasingly being used in the world of marketing. The goal is to better connect with the
00:00:12
consumer, but with the growth of large language models, do companies need to also better connect
00:00:19
with the public, especially bots, because they in many cases are making the connections for humans.
00:00:26
Pleasure to be joined once again by Stefano Pantoni, marketing professor here at the Wharton
00:00:31
School. He recently delivered a presentation to Wharton's executive education course on this
00:00:36
topic. Stefano is also co-director of the Wharton Human AI Research Program. Stefano, great to catch
00:00:42
up with you again. How are you, sir? I'm great. Doing great. Thank you. Thank you, Dan. Nice to
00:00:47
be here. You as well. So how much are we seeing humans rely on bots when they're making these
00:00:54
decision processes, like when you're talking about in retail or some other area as well?
00:01:01
Yeah, there's a lot of exciting change happening in marketing. You mentioned retail, certainly that
00:01:07
is one very important domain. There are others. We were convening a group of executives as part of
00:01:13
a new program on AI marketing that Eric Bradlow, our vice dean for AI analytics and myself were
00:01:19
co-leading. And a lot of the conversations were in this phase. And I argue that we are basically
00:01:25
undergoing two revolutions in marketing. The first one is well underway. And that revolution started
00:01:32
in 2022 when ChatGPT was released. And that revolution is about search and discovery,
00:01:38
how consumers find information about the marketplace. People increasingly learn
00:01:44
about the market through interactions with chatbots like ChatGPT and others.
00:01:48
The second revolution is just starting out. This year is turning out to be the takeoff year for
00:01:56
personal agents. We'll talk about that later. But essentially AI agents are becoming more common.
00:02:03
They're still adopted by only a small minority, but we can see that trajectory moving very fast.
00:02:08
And that means that it's no longer about using LLM-powered AI agents for search and discovery,
00:02:15
but actually delegating decision authority and making decisions on behalf of consumers or informing
00:02:20
different steps on the customer journey. So in that sense, we are seeing increasingly
00:02:24
the buying decision being co-created by humans and algorithms.
00:02:30
So because you have so many people seemingly using ChatGPT right now, and to a degree,
00:02:36
I think we still have many more people that will use it in the years ahead. How then should that
00:02:42
impact the use of that technology really in the world of marketing?
00:02:48
Yeah, so if you look at a recent paper that the team of economists at OpenAI,
00:02:54
led by Ronit Chatterjee, published on what people actually do on ChatGPT,
00:03:01
one of the three largest users was search and discovery. And of course, you can search and
00:03:07
discover about all kinds of stuff. But one of the main elements of that process will have to do
00:03:13
with a marketplace in some shape or form, people trying to solve a problem, people looking for a
00:03:19
product to do a particular job, people learning about product category and a lot of other similar
00:03:24
things, where in the past, you might have gone on a search engine like Google and then mined
00:03:31
through a bunch of different links to try to explore what are the available offerings
00:03:36
in a category, which brands offer which attributes at which level, for which price points,
00:03:42
what is the availability of these products and everything else. Now they might do that
00:03:47
more efficiently using chatbots, where you can ask a chatbot to tell you about that ski jacket
00:03:54
versus that, you know, bicycle versus that whatever. And it will tell you which brand for
00:04:01
which attributes that you're interested in, and then even help you design like a comparison matrix
00:04:06
and point you to price points and trade-offs and help you make a decision. So take us into the
00:04:13
world of the AI agent and the impacts and the bigger role that it's playing right now.
00:04:18
Yeah, AI agents are basically systems. They're not just one thing.
00:04:23
They are systems powered by language models, but that also integrate a lot of other things,
00:04:28
including the use of APIs and script and other things that basically support decision making.
00:04:34
And they are big in professional context, obviously, but they're also becoming big in
00:04:40
consumer-facing functions. This year, we've had OpenClaw and Clawbots and
00:04:47
Multibook, that kind of discussion where these open source agents have found very wide adoption
00:04:52
in, let's say, tech-savvy markets. Then you've had perplexity, increasing revenues
00:05:01
very quickly on the back of the release of AI agents. You've seen, for example,
00:05:07
Anthropic releasing lots of plugins to the Clawbot to give it increasingly agentic features,
00:05:15
and a lot of other examples of that sort, which shows how AI agents are, you know, after about
00:05:21
a year of promise, what everybody was saying is they're going to come. They're actually here.
00:05:27
So how then does this impact firms, companies, and how they have to look at their approach
00:05:33
in that relationship they have with the consumer with the LLMs, kind of as the vehicle in between?
00:05:39
Certainly, there are many ways in which these revolutions are going to impact the way the
00:05:45
company is thinking about their marketing, or generally, they think about the way they
00:05:49
interface with external environment. Some of them are quite obvious, already apparent today.
00:05:55
Others are unclear. And we know things are going to change, but how exactly and for whom
00:06:01
there is some uncertainty there. But for the more certain part, certainly we know that companies
00:06:08
need to think very carefully about their visibility to language models and bots generally,
00:06:17
so that when you type a query on GCPT or when you ask your Clawbot to do a search for you,
00:06:24
it will find you as a brand and make it an available option to the consumer. Because obviously,
00:06:29
if LLMs ignore you, then you're not even there as part of the conversation. So for example,
00:06:36
many companies sell their products across many different platforms, many different, for example,
00:06:42
e-commerce websites. And for many organizations, the back end is a bit of a mess, meaning
00:06:49
that the same product might be described with different text tags or even different codes,
00:06:55
depending on which platforms it's being sold on. Now, when these bots send out search queries
00:07:02
across the internet to gather information to support a consumer who wants something,
00:07:07
they will use the authority of the brand or the product with a very strong cue for deciding whether
00:07:15
to recommend it. Now, authority can be gleaned by a lot of different sources, including customer
00:07:19
reviews, including brand reputation and awards of different shapes. But one important thing is
00:07:26
consistency across platforms when the bot is searching for the same products in a lot of
00:07:32
different places. So if your back end is a mess and the same product is appearing as it's potentially
00:07:38
to be different products on every platform, the bot will not find you. They will think that
00:07:43
you are selling a very diffused set of stock keeping units and that has little authority,
00:07:49
when in fact it's always the same product. It's just that the bot doesn't realize. This is one
00:07:53
concrete example, which is really about getting your IT infrastructure, product coding, and
00:07:59
interface with the vendor's systems in a way that supports more decision making. Now, another one
00:08:06
that is quite obvious is also that your website, your own, let's say, web-facing real estate,
00:08:12
has to be amenable to bot search. And the Internet that we have today is an Internet that was
00:08:18
designed for people, not for bots. So it is highly visual. It supports the way that we
00:08:22
like to do things. So you've got these drop-down menus that make search and navigation easier,
00:08:28
but those are hard for bots. And so we need to think about what can we do to make it easy for
00:08:35
bots to find information that they need on our websites. And so you started seeing now the
00:08:39
emergence of protocols and standards and, you know, companies starting to release pages that
00:08:45
are not designed for people, but are designed for machines. And so that is going to be a big area of
00:08:52
investment and improvement too. Right. And don't companies have to also figure out a way to
00:08:56
integrate the two of them together as well, right? Absolutely. We will need basically two Internets,
00:09:02
one for people and one for machines. And these two Internets need to somehow speak the same
00:09:06
language, meaning that there has to be consistency in the product information that, you know,
00:09:11
things like availability and prices and product attributes and all the rest of it. And so we don't
00:09:16
quite know exactly how to do that, but there are lots of startups and lots of efforts in this space.
00:09:20
So I expect a lot of, let's say, investments and interesting innovation happen at this
00:09:28
boundary between marketing, retail and supply chain.
00:09:32
So in the presentation you made to this exec ed class, you talked about a topic that is
00:09:39
very much in and around AI in general, and that's that of bias. Where does bias kind of factor into
00:09:46
this discussion around AI? Yeah, the moment that we delegate the decisions to agents, then it becomes
00:09:52
important to understand how do they make decisions. And we know that there is bias emerging
00:09:58
potentially in language model output. Bias can mean different things depending on the discipline.
00:10:03
If you are a statistician, bias is basically, you know, deviation from validity. So it's basically
00:10:08
a formula. If you are more into, let's say, culture and society, bias might be about discrimination.
00:10:14
And if you are a psychologist, bias might be about cognitive heuristics and biases that gets us to
00:10:20
make suboptimal decisions. All of those are relevant in the context of bots. I'm a behavioral
00:10:26
scientist, so basically I'm applied psychologist trying to understand how our thinking is shaping
00:10:30
consumer and behavior and behavior of other individuals in the market. And so I'm especially
00:10:36
interested in the cognitive bias aspect. And so it's a very weird idea where as a psychologist,
00:10:41
the question you might start to ask if you're interested in the marketplace is not only,
00:10:45
you know, how people think, but also how machines think. One of the things that you also
00:10:52
mentioned in here, and so much of this discussion is around the technology,
00:10:58
that there still is a human component that is going to end up being important in all of this.
00:11:04
Yeah, absolutely. So the human component is important in many ways. One of them is obviously
00:11:08
the consumer will want to retain control over that process to the extent that they feel that
00:11:13
control is needed. And so you're going to have a sort of delegation process where consumers are
00:11:19
going to learn to trust these agents, and we're going to delegate more and more authority
00:11:23
as the agent demonstrates that they're doing a good job. There is that process of human interaction.
00:11:30
But you also have potentially another process, which is that of thinking about then what is
00:11:35
the role of human labor. Now, if you are interacting with companies today, oftentimes you interact
00:11:40
through a chatbot, and that can be convenient and effective. But you can start thinking about
00:11:44
the company when the default shifts away from interacting with a human service provider
00:11:49
to interacting with a system like an AI system. Then the question then becomes,
00:11:55
how would you deploy human labor to have maximum impact in the market? Sometimes you might find
00:12:02
that companies can deploy humans to really signify care for the relationship, the importance of the
00:12:07
customer, that we care. And so companies should think strategically not only about the deployment
00:12:13
of AI labor, but also deployment of human labor, because I think doing so could really create more
00:12:20
valuable relationships for companies. Do you hope that there's an overall message that is gleamed
00:12:26
from this type of a presentation with the students that you had, but also companies in general?
00:12:32
Yeah, my big picture takeaway of all of this is that it's not going to be human alone,
00:12:38
it's not going to be AI alone, it's going to be human and AI. And companies that are smart
00:12:44
about capitalizing on the collective intelligence of the organizations, both its AI systems and its
00:12:51
human experts are going to be the ones that are going to thrive. Stefano, always great to talk
00:12:57
with you. Thanks very much for your time today. All the best. Thank you, Dan. Have a great day.
00:13:01
Thank you. Wharton Marketing Professor Stefano Pantoni.
