Goldman tech chief says AI agents will be your next coworkers

Thanks for having me.

I was running a number of areas like serverless compute. We started that with Lambda. I was running Internet of Things. We actually started that. I was running all the mobile services. So at the end, I was running maybe 20, 25 services overall.

It was pretty amazing because I came in in 2013, which was almost at the beginning of a mega cycle of cloud adoption, and then I left in 2019 and the company had grown like 50x or something like that. When I joined, we had something like 13 services; when I left, we had like 230. So it was one of those times where I saw a lot of transformation for sure.

It was exciting because we were really learning at the same time how to overcome certain technology barriers, but also most importantly, really trying to understand what clients really wanted. And the one thing that we used to say was always to work backwards from the client. And so a lot of the time was really spent to try to understand the why and then kind of adapt the how to that. For example, is it about reducing cost, or is it about increasing speed, or is it about increasing agility? And at the end, I think what we kind of concluded was that yes, cost is a very important thing, but acquiring agility and speed and velocity was really what companies were going for.

Well, I mean, AWS obviously powers a lot of the Amazon. So maybe a little bit.

Towards the end, the last couple of years at AWS, one of the things that I was running was the so-called Internet of Things, which was about wiring up industries and companies with sensors and then having machine learning in order to optimize processes. So I started to engage a lot with companies more at the strategic level doing digital transformation. And the real question there was, how do you really transform how a company is run and operates thanks to technology as an enabler? And so I started to develop an appetite of actually being on the other side of the company being disrupted rather than someone helping them by selling tools and services to disrupt. And so when a call came, and also our CEO David Solomon was like, 'Hey, we think that technology in finance needs to really step up to a strategic level because we will not be able to be Goldman unless we are leaders in technology five years down the road or keep our leadership in technology.' So that to me was almost irresistible because it was a different field, I really love to learn, big challenge, and also I was feeling that there was a big transformation ahead. And so that's kind of what drove me there.

Right now, we've been at this journey for about two years, I would say. We've been in the journey of machine learning for over a decade, but specifically generative AI. And we have people from all over the firm actually using it. I would estimate that at this moment, at least close to two-thirds of the organization in one way or another is exposed to an AI tool. And I think a good way to think about it is that you have three waves of adoption. One is more when you are using AI mostly for making an existing process more efficient. An example being developing software or running call center, etc. Those are the low-hanging fruit use cases, and of course we have them at scale. I can tell you a little bit more, but we're already seeing a pretty significant impact on that, say for example a 20% increase in productivity for software developers, etc. Then you have wave two, when you're starting to integrate AI more deeply into the fabric of the organization, which for us means that AI not only optimizes an existing process but gives you an opportunity to actually change the way you do things. So an analyst, for example, can fundamentally change the way they analyze earnings reports. Instead of spending maybe 80-90% of their time going through pages and trying to extract signals, they outsource the extraction of signals to an AI.

Well, I think cloning is probably overrated. We tried it with sheep, right? I don't see too many clones out there. But I think it's more like having this concept of an AI co-worker, which is slightly different than the concept of co-pilot. A co-pilot is someone that essentially does the same thing that you do and helps you with a specific task. A co-worker or agent is someone that you can actually delegate to, and that's slightly different. And then all of a sudden, everybody kind of becomes a manager. Even if you never managed anybody, even if you never managed a person, almost inevitably you might end up actually managing an AI agent.

I think it will be an interesting transformation. At the minimum, you can start from the tasks that you're already doing and then just ask questions. For example, 'Hey, we have earnings season. Please analyze the reports of the companies that have announced in the last week and extract anything that has to do with usage of GPUs.' Then you start basically extracting that, assigning it. And think about it, the way you scale is that those agents in turn can outsource to other agents. So you have the same way as you have managers. When you ask me how do I manage 200 agents, how do you manage 200 people? Well, you manage 200 people by actually having maybe 10 people that you directly manage, and then those 10 people manage 10 others.

Well, I think that would be an interesting one. Actually, one of my predictions was that in 2025, we might see the first layoffs of agents. So layoff of agents, but essentially you just shut them down. You shut them down, and you can retrain them. You can put them back to the drawing board and retrain them. What's interesting is that at a minimum, everyone that has never even managed a person is going to have to develop three basic management skills, which are going to become essential for almost any job where you want to employ agents. One is the ability to properly describe what you want, which is not obvious. Sometimes you know how to do things, but if I'm asking you to describe in detail to someone else, you have to have a very good mental model. So the ability to actually describe is an important one. The second one is the ability to delegate. You have to be able to break up your work into chunks that you can give to different people. So you take what you do and say, 'You know what, I can parallelize these actions,' and then start giving them to someone else. And the third one is the ability to supervise. You now need to understand that the same way as a person might not be perfect in what they do, obviously an AI in some cases is even less perfect. You have hallucinations, you have process errors, and so you need to find a way for you to verify your work. At Goldman, one of the things that we really cherish is the fact that we always want to have human supervision of AIs. And so those three skills are something that I think will become a basic requirement in the tool set of any employee, the same way as today you need to know how to use a computer. 50 years ago it was not a requirement. I think in a few years, it's going to be, 'Hey, you need to be able to manage an agent to actually be able to work.'

I think actually, what's interesting is that what does a motivational town hall do? It gets people in a state of mind where they tend to be more excited and more productive. But you can do that to an AI as well. The way you write the prompt, the AIs respond really, really well to the way you ask the question. So in that case, it might not be a motivational town hall, but it might be you realizing, for example, and believe it or not, it is true that if you're kind to an AI, the AI will actually give you better answers. I've verified that myself. For some unknown reason, being kind to an AI is a good way to actually get them to give you a good answer. Maybe because in the training set, there was a bias towards a certain type of expression or kindness that was correlated to good material being trained upon. I don't know, but there is a psychological element. And so it would be wrong to think about an AI as something that is just as mechanistic as a computer. A computer program is 'syntax error, just tell me the same thing the same way all the time.' It's not the same for AI.

Listen, this is just another transition, like before Excel, after Excel, before you could use Google, after you could use a search engine. So there is a certain pattern.

Absolutely, absolutely. Why did I say Google? It's okay. Of course you use Yahoo. And so I think that will happen and is happening naturally. It's more about what part of your job you really feel that you add unique value. It kind of elevates your work rather than doing repetitive tasks. So I think the bar is just going to increase. We have this platform called the GSAI, which is a platform that we built to give access to the latest and greatest models like OpenAI, GPT, Gemini, Anthropic, etc., in a way that is safe and compliant for a regulated industry. And then the tool called GSAI Assistant, which is essentially a chat that gives you access to all those models but at the same time links to all the data that is relevant for you that lives inside the bank. And so what we see is that people exposed to that, first of all, you can literally see this learning curve that goes super fast, to the point that we get people using it with 40-50% a month growth in terms of how much they use it. But also you see what kind of questions they're starting to ask, and that's how you know their work and their attitude is starting to shift. At the beginning, you ask more basic questions. For example, you want to know an acronym, or you want an explanation to try to explain to a client. Or for example, you have a portfolio of stocks inside a basket, and the client is asking, 'Can you tell me some information about those stocks?' and you can ask an AI. And then as you go to more deep usage, you're starting to see what I said before: 'Hey, why don't you do this for me? Why don't you retrieve a chart? Why don't you create a presentation that will include some macroeconomic indicators?' So the sophistication of usage becomes a little bit like the interaction with a colleague. At the beginning, you exchange ideas, and then at the end, it becomes entrenched in the way you work. I think seeing that revolution is really fascinating.

In terms of tasks, yes, we're seeing shifts. For example, for developers, we're starting to see that a lot of the mechanical jobs, like migrating from an old version of Java to a new version, migrating to a certain web framework, writing documentation, writing unit tests, etc., is starting to be taken by the AI. But I wouldn't go as far as saying taking an actual job.

The answer so far is no, because there is so much more that we have below the line every year when we do our budgeting, and there is so much more that we want to do because it's a competitive environment. So at the end of the day, what we've done so far is really to reinvest the extra capacity, which will give us additional competitiveness. I think for the foreseeable future in a competitive market, most companies will actually decide to reinvest.

A reflection that I've been having is that there have been a lot of transitions over the years: before computers, after computers; before mobile, after mobile; before the cloud, after the cloud; before outsourcing to other countries, etc. And inevitably, the pattern has been you might have some job displacement at the beginning, but then those industries tend to create a lot of new jobs. Just look at how we have 12,000 developers within Goldman Sachs out of 45,000 people. If we didn't have computers, we wouldn't have those people. So generally it rebalances itself, and I think that will be the same for AI, except there is a certain velocity. The transition to computers took decades, the transition to the cloud took several years, the transition to outsourcing took several years, so there was more time to reabsorb and reskill. This transition is happening very fast. It feels like there are months instead of years. So I think one of the things I've been reflecting on is, can the job market shift as fast, or could you have a deeper displacement while those new skills or new jobs get created? That's something I obviously don't know the answer to, but I think it's one of the most poignant questions that we should ask ourselves.

Yeah, it's going to be an evolution. I don't know how slow it's going to be. In fact, I think it's happening very fast, to the point that if you think about it, six months ago we almost didn't think about agents. Now it's all the rage. And one specific moment in time was in Q1 when a lot of the AI model providers started to introduce reasoning models, like o1, o3, or DeepSeek, or all the others. That means they're not just answering your questions; they're actually starting to think about how to answer. They can do plans, they can create processes, and that is what really created agents. Because an agent at the end of the day is you give them a question or a task, they create a plan, and they know how to execute that plan. That's reasoning in a way. So the biggest shift has been models that are capable of actually reasoning and doing tasks on your behalf. And if you look at that in a timeline, you see that the acceleration has been pretty remarkable, to the point that today pretty much all major models have reasoning capabilities and can be used to create agents.

First of all, I don't know how much clients will enjoy that experience. 'Hello, sir.' But I've seen a lot of robots. My prediction is that humanoid robots first are going to be used in business in industrial settings, like in a factory to assemble things, to move things around. I think human interaction is something different, and so personally I don't look forward to that day for sure, where I'm going to be greeted by a robot.

Well, we've seen many buildings in many countries have robots that are delivering packages. That's absolutely great because they're extremely efficient, they don't misplace packages, and the interaction is very simple. So I think I can see a bright future for robots, but I'm thinking that a certain type of warmth and human interaction, especially in businesses like banking or investment banking, are going to be very important.

Thank you so much.