Goldman Sachs pilots superhuman, generative AI banker project

Goldman Sachs is piloting a generative AI project that could boost the efficiency of bankers' meetings and their relationships with clients, chief information officer Marco Argenti tells Axios.

Why it matters: Wall Street giants are racing to figure out how generative AI could reshape the industry, while keeping regulation and privacy safeguards in mind.

How it works: The pilot, which is in proof-of-concept stage, suggests speaking points sourced from the minutes of previous meetings and additional data points that it believes will best resonate with a client. This aligns with the heavily relationship-based nature of banking.

• Other data sources for the project could include public filings, market research, and compliance documents.

This conversation has been edited for clarity.

You told the WSJ that gen AI could have an impact on the organization in months rather than years back in May. Give us an update.

• I've been excited since the beginning of the year, so we're about 8-10 months into the journey. We learned a lot, and we're seeing some applications that are starting to show promise, as well as some that are not a particularly good fit.
• We incentivized each team to come up with ideas within three layers of governance. We have an overall AI steering group, steering teams within each of the divisions, and a group that approves proofs of concept or pilots. We ended up with, say, just over 100 proposals, Now, we have about 15 ongoing proofs of concept.

What use cases do you find the most compelling?

• When you start seeing this many ideas, some primitives emerge. For example, there's a big divide between use cases that are 100% focused on productivity and use cases that are focused on, let's call it, super-humanizing. [That latter case] makes people smarter and capable of doing things they were not able to do before.
• So the difference is: Am I doing the same things with less resources, or am I actually doing things that I couldn't do before?
• The productivity case is where a lot of companies are prioritizing their investments, because it's easier to get the return on investment right away. You can say, "I used to need 80 people for this, and now I need 20."
• But I think empowering people to do things they were not even able to do before is worth the competitive advantage of the alpha in the future.

Can you tell me a bit more about what it means to super-humanize versus focusing on productivity use cases? There seem to be places where the two overlap.

• We have a pilot going on for our bankers. So when they have a meeting with their clients, the AI can suggest speaking points that resonate with this client based on the minutes and the notes of the previous meetings or current events.
• It's basically a chat interface similar to Chat-GPT, with a difference that is actually plugged into our platforms and decides what's the best model to use, and which sources of information to use. So it's all plugged into, for example, market research, investments, and compliance documents and has one place for reinforcement learning.
• Let's say that if out of 10 meetings with the client, normally three or four ended up really providing new information, and the rest is more like relationship. Now, maybe nine out of 10 are new information, and one is just a relationship.
• Is that more productivity? Well, I don't know. It's still the same number of meetings, but it is adding a lot of value to the conversation.

The banking industry has been careful around gen AI given concerns about compliance and data privacy. How are you approaching the issue?

• There is always a human in the loop in every single thing we do. We talk to regulators, and we talk to compliance — and this human-in-the-loop method is more, in a way, tolerated. We never go straight to the client.
• Given there is a single platform, we apply a ton of controls. Some of the models are internal and hosted by us. With the larger models like Microsoft or Google, we work with them very closely to make sure that we run secluded and private versions in a way that there is absolutely no data exfiltration and that they can't read our prompts. It's completely stateless.

There are also areas that you don't see gen AI being particularly useful.

• Yes, I think, in cases — which are a lot of them — where the visibility of the algorithm needs to be absolutely clear, explainable and auditable. In these cases, traditional AI or no AI might be better.
• There is a whole class of tasks that need to be highly explainable where you need to say: "I gave you this, and I didn't give this other person that." KYC, for example, when you onboard a new customer.
• The other one is cases in which you're trying to predict the next point in a sequence. Right now, I don't think there's evidence that generative AI could predict the behavior of a certain asset in the market because it's just not designed that way.
• We're of course testing it, but we're skeptical on the whole prediction of the market. I think that part is more science fiction right now than anything else.