Welcome to Goldman Sachs Exchanges. I'm Alison Nathan and I'm here with George Lee, who is the co-head of the Goldman Sachs Global Institute. Together we're co-hosting a series of episodes exploring the rise of AI and everything it could mean for companies, investors and economies. George, good to see you again. Great to see you. I really enjoyed our first discussion and super excited about this one today. So today we are going to discuss how AI can fit into enterprises. We're always talking about the use case.
And perhaps we should say it's a big question about how enterprises can fit themselves around this evolving and exciting technology. So set the stage for us. When you think about what companies are doing today with AI, how would you explain the usage and how has that evolved over the last year or two?
Yeah, great question. Look, I think we're in the early days and that's consistent with the state of development of the technology itself, which is quite novel as you and I have discussed before. That having been said, there is definitely a value gap opening up between the trajectory of the technology and the amount of enterprise adoption of that technology. And I think there are really three reasons for that.
First is, always was it thus? And by that I mean that the diffusion of new technology in businesses is always a little bit of a slow, deliberate process.
Second, the pace of progress of the technology itself is an obstacle to adoption. Any technology that is moving as fast as this one makes it hard for CIOs, CDOs, et cetera, to figure out what technologies to deploy, when, where, and how. And so almost it's funny, as much as I admire their trajectory of improvement, for the sake of adoption, I almost wish it would slow a little bit.
Third, it's a general purpose technology. That's part of the genius, but that's part of the challenge for an enterprise. So its use cases are bounded only by human creativity and imagination, but it takes human creativity and imagination to determine how you want to use this very broad technology. Now, I have two caveats to that. One is there's a generational component to this. In the department of it's hard to teach old dogs new tricks,
As new people join our organization and organizations like ours, they inherently bring with them an appetite to use the technology, a facility with the technology, and they will, in a way, drive adoption naturally. So that will happen. The other thing I would say is there's a big distinction today between big incumbent companies who are using this technology or attempting to use this technology and de novo enterprises.
If you're starting a business from scratch today and you can write code unfettered by legacy, et cetera, you can instantiate processes that are built from first principles around AI, it's an amazing piece of leverage. We have encoded workflows, habits,
regulatory obligations, new technology, old technology, and fitting this new tool into that mosaic is just inherently harder. But man, if you look at the data from, I saw something from Y Combinator recently, new businesses being formed today that are using this technology to build the software platforms to allow them to operate are just zooming. They're moving really fast. So look,
early days, I'm very optimistic and it'd be fascinating to talk to our guest about the actual practical day-to-day challenges of deploying this new technology.
So let's bring Marco Argente into this conversation. Marco is our chief information officer here at Goldman Sachs. Thanks so much, Marco, for joining us. I'm happy to be here. And this is definitely an exciting topic for us to discuss. Great. Welcome. Thank you. Marco, given the context that George just gave and the many developments in Gen AI over the last even weeks, certainly months, has your thinking about AI changed at all in this time period?
I think the world of AI has changed like three or four times over even in the last year or so. And I think what's interesting is that every time we seem to have reached some form of a plateau, here comes sort of a bolt from the skies that, you know, just kind of makes us rethink any illusion of the technology having reached any sort of scaling limit. So if you look at even like the last two months,
There has been some of the most fascinating evolution with regards to specifically thinking models and models that have the ability to reason. And if you look at those are not incremental changes. Also for the user standpoint, you now have models that, for example, can do
real research for you in a depth that was unprecedented before. I personally tried it. I compared all these models on a topic that I know really well, which is kind of the last few years of the history of Nokia smartphones because I used to work there and it's an interesting, like, obviously case study of, you know, strategic choices and I had six different AIs doing the research and some of that came out as
the best book I've ever read on Nokia. It felt almost like it was written by an insider. And, you know, then we tried other subjects. So I think what I'm trying to say is, you know, I've been in technology, like, sadly for about, I would say, since I wrote my first line of code. Now,
45 years, almost exactly. And I've been through revolutions such, you know, the beginning of mobile or the internet or the beginning of the cloud and Amazon. This is compressed in a way that I've absolutely never seen before. So we all agree that the advances, the technology itself...
is making is absolutely enormous and unprecedented in terms of the speed. But when you think about this topic of enterprise adoption, I mean, George already said it's not necessarily keeping pace, but what are you observing from your seat?
In general, I agree with the fact that enterprises in general are lagging the state of the art with regards to adoption. Also because, I mean, let's face it, it took even like decades to be perfectly proficient with computers and then it took years to be implementing the internet in a meaningful way. It took a couple of decades even for adopting the cloud.
I would say we're like in year one and a half of actual useful products and not toys. So let's give ourselves a break a little bit because guess what? On one side, you have a brain that is evolving on a daily basis. On the other side, you have a human brain and habits that generally tends to evolve in hundreds of years or millennia biologically and at least years from a habit perspective.
And so I think right now the biggest friction point to adoption of AI in the enterprise is people, is behaviors, is the fact that you have muscle groups that you need to retrain. Although I have to say that we're starting to see areas where there is actually fairly consistent and meaningful progress. And those are in areas that maybe are a little bit easier to change,
Because the people that work in there are more, in a way, open to change. And one of them, for example, is the developer community. The developer community, in many cases, tends to be eager to experiment new things. Historically, they know how to deal with imperfect products.
A new operating system is released three, four, five versions to developers before it actually hits the market. And so you kind of know how to use stuff that is not there yet. And you kind of use the good stuff. And then you know how to deal with the bad stuff. And so I think that community has been very receptive. And that's one of the first area that in Goldman has gained real traction of really improving our developer experience. And then as a result of that, gaining more efficient. There are many other cases. But I will say that today,
The real answer to your question is this is probably one of the biggest change management challenges and exercises that any corporation have ever seen in their history. Totally agree. And Marco, one of the things I talk about with this technology is, especially as it relates to adoption, there is this magic button hypothesis.
The idea that our users tend to expect it to work perfectly and solve all their problems and unless and until it does, they tend not to want to adopt. Again, caveat as you write certain populations more inclined to do it.
How are we getting people on the journey with us? And how do we identify people who are willing to experiment, to change habits, to reconsider? Talk a little bit about what we're doing to foster adoption by finding those people inside the firm. You know, if you want to have a disruption or if you want to kind of be ready for a disruption, you need to find people that have already crossed the bridge of accepting that a disruption is going to take place and be very open to it.
And so we have really started to seek for disruptors
that have taken a very kind of purposeful approach on how to navigate through that disruption and ending up on the winner's side, even at the sacrifice of questioning their own beliefs or even sometimes really profoundly rethinking their own businesses. And so we identified quite a few of those. And the trick when you're finding those disruptors is to find people that are actually in a position where they can drive change.
And I think we have a few, you know, I don't want to, you know, call out names, but it's starting to become a movement. And those people in turn will influence others, will become role models, and then they will actually apply their judgment. So because we don't want reckless disruptors, we want mindful disruptors.
And I feel pretty good about how we've been doing so far. Exactly. Now, I think the dynamic inside businesses, once you identify those disruptors, they're willing to take some risk. Think from first principles. They start to generate positive results. There's a follower effect in enterprises. And so that's what I think we're hoping to catalyze.
Marko, you talked about one of the obvious use cases, which is instrumenting our developers with this technology. I can think of a few other sort of more pedestrian use cases. Let's dream big for a second. What are some of the most ambitious use cases of this technology for a knowledge-intensive business like Goldman? Let's fast forward a couple of years, and what are some of the inspiring use cases you can imagine in a business like ours?
I think my belief is we're heading towards a world where we're going to live in a hybrid workforce, where you're going to have to have the ability to manage interactions with human colleagues and AI colleagues with the same ease. And so in this hybrid workforce, one of the great advantages is the fact that you can achieve elasticity the same way as you do with the cloud.
If 10 years ago someone would have asked the CIO of any large company how many servers you have, they would have said the number with very great precision. Today, if you ask me how many servers I have, I tell you I have no idea because it might change on a minute-by-minute basis because with the cloud, if you have a high workload, especially in the world of serverless, you can go up and down in minutes.
And I think that is actually going to be the same answer if you ask me 10 years from now because you could actually surge agents for a closing of a quarter or for an earning season or for particular times in the market and then shrink and then go up and down on the agent side where your human side kind of remains more stable.
And talking about the way this is going to impact our daily life, I think one of the first objectives that we kind of already, I think, kind of realized is that pretty much every developer right now has AI in their what's called IDE, which is their interface to do coding. I want that to happen to every single employee, but not for coding, but it should be basically as natural as Google. It should be as natural as doing a web search.
It should be actually most importantly as natural as sending an email to someone asking, can you do this for me or can you find someone that could give me this information, like client report or research, et cetera. So going into really a world where you can ask natural language questions and you get very complex, very GIS-specific answers. And for that, we built a
and we're deploying at scale our GS AI assistant, which is really like your desktop for AI. That's the way to think about that. And from there is where you go to find answers. And I think to me, the evolution and what I see as a really fascinating future is where
this assistant is actually becoming and sounding more and more like a very experienced GS employee or a very deep expert on a certain subject rather than giving you generic answers. Because at the end of the day, the companies that are going to win are the companies that are going to have a workforce that is not only embracing AI, but also in a way knowing how to make AI part of their daily life and their daily work.
Marco, there are some guardrails that an institution like Goldman Sachs has to be aware of and put in place. So, I mean, talk to us about what we might be doing to try to overcome the limitations that we face in the type of information we deal with.
We're using the latest models, but we're very careful at controlling any potential infiltration of external data or data externally. And most importantly, we are doing a lot of work to "ground" those AIs so that they don't make stuff up. That is the most important thing. Right. So rather than removing stuff, we tend to actually make sure that stuff is actually accurate.
by cross-checking it with sources. So I think that's one of the things that unfortunately we live with when we talk about AI, that when you use AI as the wild type of AI that is out there, it's kind of a medical term, which means what's normally out there.
You, in many cases, tend to get information that is not necessarily accurate, but it sounds really plausible. And then if you kind of try in a sort of a heavily grounded AI, it might sound a little bit less creative or imaginative, or maybe it's more correct. And so I think we've been putting a lot of our biggest effort there has been to increase accuracy, reduce hallucinations.
and also to make sure that we minimize the risk of prompt injections or data somehow exfiltrating out or being inserted in the mix. - This leads me to something you've written about, Marco, that I'm fascinated in, how the managerial challenge will evolve over time where we're not only managing humans, but we're managing AI agents. How will that shift the job of being a manager and what skills do we need to build to live in that world?
Yeah, that's an incredibly fascinating question. So first of all, one of the challenges is to somehow inject, call them the cultural traits, the leadership principles, the tenets, like I like to call them, of the organization into the AI agents the same way as you do with humans. Because
Cultures are very peculiar. Goldman is very peculiar. Amazon, where I was before, was very peculiar. And I tell you that people tend to kind of unify their decisions, their way of thinking around certain tenets, sometimes written, sometimes not written, and cultural traits, which you will get when you're here and they're hard to describe, but you see them when you see them. Whereas agents that you implement from the outside,
are going to be like employees that we hire the first day. And so the challenge is agents will get smarter and smarter, but not culturally smarter if you don't do something on that front. And I think this is not something that has been talked about too much. We talked about mostly like a specialization of expertise. We're not talking too much about, in a way, enhancement of the cultural aspect of agents because those agents eventually will
Even today, we have, let's say, AIs or kind of agentic AI that does things like providing or creating drafts of research. Or for example, in private wealth management, when it comes to portfolio analysis, to use agents to summarize the nature of business of dozens of positions and in a way that puts
a Goldman lens into the answer. Even like something as simple as translation, which obviously should come out of the box, we have adapted the tool of Translate AI to translate in a way that is actually very specific to our financial content, but also that somehow is specific of the Goldman language in a way, the way we call things. Imagine just acronyms, how complex that is.
Those are just early examples, but in general, I think as we use agents,
to more and more interact with people, but also produce content for people. I think there is a little bit of a sort of an extended example of the three laws of robotics for Misa Kasimov. There are, you know, what are the actual leadership principles or tenets that an effective corporate agent should adhere to in order to fit into the culture of an organization? And I think that's a very complex problem that I don't think has been solved yet, to be honest.
I mean, this hybrid workplace is just such a sea change from where we are today. I mean, do you see it really just driving corporate strategy overall? Yeah, there are companies like I was on a panel actually yesterday with a chief people officer of a major software company and her new title is chief people and AI officer.
Because, you know, they're already thinking that even today, they want to look at how do we make sure that employees that communicate with our AI tools are not exposed to biases, for example.
example, and also the cultural aspect of the answers is very important. And then agents are definitely something that is already knocking at the door. And so when you do like resources planning or when you actually like, you know, the training right now is such a big deal because we are all pretty much AI illiterate. This is comparable to a change on pre and post computers. And by the way, funny enough,
If you talk to people like in the, I don't know, 40s and 50s or 60s about whether they could imagine a workplace without computers, a lot of people will tell you, absolutely. You know, the computer is always going to be this big thing over there that a few people with white lab coats are going to use. But on a day by day, there's nothing like a physical signature or a paper. You're going to sit down and read it and whatever. And
Think about like today, imagine any company that doesn't have a computer, what would they do? They couldn't scale even like minimally.
And I think in this new world, there is going to be actually an interesting, like you mentioned that, but a bit of a generation shift where you now have this new generation and my daughter is 20 and she knows how to use AI very naturally because she's been using it for at least two or three years. And a lot of the kids that are even coming into Goldman, they're extremely versed on AI. They're great prompt engineers and it took them years to perfect that.
The same way as computers. And, you know, like I've seen that with, you know, with my parents or etc. that
There was a point in time in history where if you actually never used the computer until you were like 40, you would be like never having ridden a bicycle until you were like 40 or 50 or never done certain things early enough that you will never be natural around it. You could use it, but even today when I look around, I swear there's people that are still typing with one finger or two.
Okay? And those are fine. They can be very effective. But there has definitely been a generational gap before on the use of computers and then on the use of the internet. And now with AI, I think it's even more profound.
And so the change happening in a generation that is not AI first is the same as a generation that was not digital first. And that's why I think that's one of the biggest challenges for human resources and human capital professionals to try to understand how to navigate into this. And what makes it, I think, different is that all this is happening in weeks versus years.
And that is what I think are raising some of the most profound questions on how to actually drive that change or how to actually somehow survive that change. So how long will it take until we get to a hybrid workforce?
I think for developers it's already kind of happening because there are today virtual developers that you can buy for $20, $30 a month, literally. And you can chat with and they do things for you. You're always going to need a developer because I don't believe in a future that is dominated by AIs and humans of a marginal degree.
I mean, I tell you that is not even in sight, not even for, you know, the most futuristic predictors. In fact, AI, I think, makes the role of humans even more important.
Because you now have an added responsibility. Everybody becomes a manager. Everybody becomes a manager of agents. They can have hundreds if they want. And everybody needs to be somehow taking responsibility for the work that is done. And the amplification of your good and your bad can be incredible. Your mistakes can be amplified and your successes can be amplified. And so you even have more responsibility. Therefore, this is a wake-up call to step up for everyone. And I think the centrality of...
professionalism and high judgment has never been more important than now. Speaking of being an AI native, you and I both push ourselves to be ardent adopters of the technology, but you have a very specific personal use case that I admire and I've seen a lot of cool evidence of. Why don't you talk about some of the ways that you use AI and the one in particular you know I'm thinking about?
So I do, in my spare time, which is not much, but I'm a little bit of an amateur musician, and so I composed a few songs with AI that actually some of them my band has actually performed live. And I have to say that, okay, AIs can do terrible compositions, but actually if used properly...
They really accelerate or make it, you know, make it much easier to kind of go through the creative process. If you try to kind of help it for some of the mechanical part of composing a song or like, you know, creating rhythms or loops or so there is a lot of musicians that are actually using it creatively. And I'm not obviously one of them, but I aspire to be at least okay on that. And so I've been doing that with pretty good results. And all it takes is just knowing how to ask.
First of all, Marcos being humble, he's a terrific musician. And I must say the music that he co-composes is remarkable. The lyrics are poignant. The rhythms are interesting. And so it's one of those things that you see that really makes you sit up and take notice of the progress of the technology and the ways that human creativity can be used, can be elevated with it.
So, Marco, thank you so much for being with us. Fascinating discussion. We could go for hours. Thank you, Allison. Yes, absolutely. I mean, we started this conversation saying adoption. We got to give ourselves a break. We're only a year and a half into this. So I'm really interested to see how this all evolves. Thank you so much, Marco. Thank you so much. Thank you. And thanks to you, George. Great. That was fun. This episode of Goldman Sachs Exchanges was recorded on Thursday, March 6th. I'm Allison Nathan. Thanks for listening.
Thank you.
is the property of the company to which it relates, is used here strictly for informational and identification purposes only, and is not used to imply any ownership or license rights between any such company and Goldman Sachs. The content of this program does not constitute a recommendation from any Goldman Sachs entity to the recipient,
Thank you.
which may vary. Neither Goldman Sachs nor any of its affiliates makes any representation or warranty, express or implied, as to the accuracy or completeness of the statements or any information contained in this program and any liability, therefore, including in respect of direct, indirect, or consequential loss or damage, is expressly disclaimed.