GenAI Questions, Answered: A Primer for Institutional Investors

Recently, Verity’s SVP of Product Katya Taycher sat down to discuss the current state of generative AI (GenAI) with Snowflake’s Principal of Financial Services Thomas Gray. The conversation ranged from the promise of AI-enhanced productivity to the practical realities of compliance to audience-directed questions asking for predictions on the future of investment analysis work in a post-genAI world. 

You can watch the full discussion here >> 

Below are excerpts from the conversation. 

How would you recommend an investment team approach using GenAI?  

Taycher: [There are three options currently: using ChatGPT directly, relying on existing providers to incorporate GenAI features, or a in-house build out]

If you’re going to use ChatGPT or other off-the-shelf generative AI solutions, check with compliance. What are  the guidelines for your firm? If your firm doesn’t have them, they will likely come up sooner or later. 

The next option is a vended solution. That was the most common option selected in our survey. [Ed: a survey within the webinar asked how attendees were using/considering using generative AI.]

Existing vendors will be incorporating generative AI. (Verity is one such vendor. We already have some of the generative AI features in one of our products.) Or, looking for new GenAI-enabled providers. Certainly, there are a lot of vendors right now looking at incorporating GenAI.  

The final option is to build your own solution. That’s harder in the sense that it’s a bigger undertaking, but it does address some of the privacy and security concerns that some of the larger firms might have. The biggest one is that your data is never commingled with other clients’ data and that it’s not available for the models to learn or, otherwise, leaves the firm. So those are the options. 

What kinds of tasks should funds consider enhancing with GenAI?  

Taycher: Generative AI has specific strengths. When you give it tasks where it cannot hallucinate, then it does quite well. Examples of that are summarized content. It summarizes quickly, efficiently, and fairly accurately. Also: entity recognition. So, when you have text identifying companies, people, locations, currencies. Now, this is something that could be done before generative AI, but it would take a lot of resources, was expensive to do, and was not quite accurate. Generative AI really made a leap forward in that regard.  

Sentiment analysis is another feature that was able to be done previously with other tools, but generative AI makes it easier to do and more accurate across a broader set of data.  

Another is translating text into code, which is something we’re seeing across our client base. Data analysts can become much more efficient if they can, let’s say, write SQL queries with the help of generative AI. Again, those do need to be checked and validated because the answers tend to not be accurate from the get-go, but they do provide a big help getting started and enabling productivity gains.  

Finally, answering some types of questions that have been validated and that have gone through prompt engineering can be good. Prompt engineering is when someone creates prompts and makes them very specific to make the models respond accurately. Usually, this takes several tries and a number of tweaks to the prompts themselves. But then you can get consistent answers to some general questions.  

What about tasks that funds should be careful about enhancing with GenAI?  

Taycher: Now, some of the weaknesses of generative AI. It generates text that sounds thoughtful and human-like, but might not be valid. So, generating thoughtful text on a new topic is not something that generative AI can do well — at least not today. This might change. The models are evolving, very quickly, much faster than any human could learn. So, we’re expecting much better quality of results and accuracy of answers down the road. But at this point, I would say you have to be very wary of any answers from these models. They always must be validated and fact-checked.  

Another area where GenAI generally is quite weak — although, again, getting better — is quantitative analysis. It can pull out numbers from documents, but if you wanted to do aggregations and comparisons and more advanced kinds of analysis, it typically has not done well. So, lack of reliability and having to fact-check answers are the major weakness of GenAI.  

Relying on use cases where it is strong and where it does not make up answers is the key to the successful use of generative AI technology.  

How are firms navigating the compliance concerns or potential risks of leveraging these LLMs in their workflow? What measures are they using to mitigate this? 

Taycher: I can say from my experience working with our clients, that’s question number two that I always get (question number one is the cost of this.) So, first of all, if there are any compliance concerns or any security concerns [at your firm], I would advise against using ChatGPT directly. They do say anything you put in there can and will be used for training models. Same is true of Google’s Bard and other similar publicly available services.  

OpenAI — if you use it via the API (which is how firms like us use it) — states in the contract they cannot use your data to train the models. So that addresses that question assuming you trust OpenAI as a company.  

OpenAI offers other options by default. Regarding the data that you input, they have a thirty-day retention policy, which is, in their words, to protect against weird things or safety concerns. But you can ask them to reduce that down to zero days. Basically to say, “Do not store my data at all.” That’s something we recommend to our clients.  

Finally, if you’re using a vended solution (let’s say if it’s not Verity), ask them how they’re making sure your data is not commingled with other clients’ data. What we do at Verity is we suggest our clients leverage their own OpenAI accounts so that any data sent to OpenAI is sent for each fund individually. Then there’s no chance that it’s commingled with other clients’ data.  

But that’s, again, up to each client’s infosec departments. My prediction though is that in a number of months, and I do say months and not years, GenAI platforms will become as commonly used as, let’s say, cloud computing. These concerns will be less prevalent because it is just something that everybody does.  

Can you provide a few examples of how investors can lean on AI to provide tangible results?  

Taycher: Absolutely. Investment firms have a lot of research that is, essentially, their crown jewels. As long as you have that research and data in one central place — be it Verity, Snowflake, or somewhere else — you can then leverage generative AI to share and provide answers from key research and have it be used much more effectively across the organization.  

On the other hand, investment managers and analysts still have to sift through tons and tons of unstructured data on a daily basis. Be that SEC filings, sell-side reports, MD&A, or anything else that they have to read through. Generative AI can be very helpful in reducing the amount of data that they need to look through by generating summaries, finding key points, and being able to answer What is the most important thing that I need to know about today? 

Again, answers are subject to [human] validation. They might be incorrect or incomplete. It’s also dependent on having all of that data in a centralized place. But as long as you have both of those covered, it should help with productivity gains across the investment industry.  

In your conversations with large asset managers, can you speak to the next phase of AI and what customers are hoping is on the horizon?  

Taycher: We’ve been working with a number of large asset managers. The biggest request I can think of is unlocking the power of their internal research. So, in other words, creating the ability to have their own internal chatbots that can answer questions about their proprietary research and, for example, compare research on multiple companies. Without the power of generative AI, you have to read all of the research written on two or three companies and then draw your own conclusions. Generative AI helps you do that more effectively. Again, generative AI at this point requires you to validate all those answers. But to the extent that these tools can provide sources — where it gets the answers from — that is helpful. So that’s something that we’re building into Verity products as the next phase.  

We’ve also seen some of the largest asset managers doing it on their own as well. So, to summarize, it’s basically the ability to access firm-specific data using GenAI and ask questions directly. This can be extremely useful in the private equity world, where you have proprietary research and other proprietary documents that are not available elsewhere.  

Watch the Webinar, Co-presented With Snowflake

The popularity & promise of generative AI is causing technology leaders to reassess their LLM strategies. How can investors effectively implement this groundbreaking technology in a rapidly evolving landscape?

Join Katya Taycher, Senior VP of Product & Operations at Verity, and Tom Gray, Principal of Financial Services Data Collaboration at Snowflake. Drawing on their experience working closely with the world’s leading asset managers, Taycher and Gray will discuss how funds are navigating the disruption and capitalizing on generative AI. 

Watch the Complete Webinar Now