How banks can extract value from generative AI

Dan Latimore recently joined The FR as chief research officer. He will produce ongoing research and provide insights into the difference financial services companies are making via innovative technologies, including through a recognition program called Most Impactful. The FR’s editors recently sat down with him to get his insights on the state of generative AI and banking. The following Q&A has been edited for clarity.

It’s been almost two years since ChatGPT rolled out. With consumer expectations rapidly changing, how should banks harness generative AI’s benefits?

Gen AI is a generationally important development that will democratize many aspects of technology, bringing cutting-edge tools to non-technologists. All banks should be looking at how they can incorporate GenAI into their workflows, but they shouldn’t pursue it just for the sake of using Gen AI. Do it to create specific value, where the three classic levers are increasing revenue, decreasing cost, or mitigating risk. 

Since you need to focus, you can deprioritize increasing revenue, at least for today. As you look at becoming more efficient and decreasing costs, the low hanging fruit lies in chatbots, process improvements, and coding.

Banks have long been boasting about how many tasks their chatbots can do. How does generative AI fit into that picture?

A constant theme with generative AI today, particularly when it's customer facing, is that you've got to have a human involved at some point. There's still some hallucination that goes on, so the output that comes out directly from the LLM [large language model] may need some help. Also, if you're using a chatbot up front, have a clear off-ramp so that when a customer needs to go from the AI to a human, they can do so seamlessly. You also need to be transparent by telling the customer what just happened.

Gen AI is also great to generate initial responses in a host of different situations. A chatbot is a nice way to start a conversation with the customer, with the proviso that they can switch to a human when needed. For example, if a customer starts to discuss bankruptcy, the chatbot should hand off the conversation to a human quickly.

Recommendations are another use case. As an adviser discussing financial planning, you can use gen AI to generate a first draft of a financial plan. Gen AI can save an underwriter a lot of time in doing the initial preparation and input and even offer some initial suggestions about what the outcome should be. There’s also process automation: you can synthesize and summarize long documents. If you get a 100 page PDF, you can ask what the main points are. AI can summarize Zoom calls, and then you can do a sanity check on it, because you do need that human interaction. You can also generate training material efficiently by taking advantage of the gen AI’s ingestion capability to produce a first draft of a manual.

Circling back to chatbots, have we really moved the needle past these sometimes annoying rules-based chatbots?

Absolutely! One of the biggest advances, obviously, is that customers now can ask questions in natural language and they won't get bumped out of the process if they don't call something by the name the rules-based system recognizes. That’s a big deal. 

Two years in, consumers have gotten much more used to interacting with bots. If you do the models right, you can reduce your wait time quite a bit, which is a great customer experience. And the AI can make better recommendations via the agent, who can direct the customer to the right product out of a couple of dozen potential choices.

How can generative AI be used to draw insights from customer interactions with their banks?

If you've got a large body of interactions, you try to look for recurring themes. So maybe you find out that a lot of customers are asking why the interest rate on their checking accounts are so low. If that leads to many closed accounts, maybe you want to think about reexamining your rates.

On the predictive analytics side, you can use this to augment and test over and over again the churn and customer attrition algorithms that you've already got. You can also identify customers to go after proactively, and cross sell and upsell.

We’ve heard a lot about generative AI helping people code and build new apps. What are your thoughts on where that is at right now?

Goldman Sachs has been doing this since the dawn of ChatGPT. They recently said that they’ve become 20% more efficient in their coding across the stack: creating, testing, further debugging and rewriting. And then there are new apps like Cursor allow someone like me the capability to code and write my own apps. As firms become more experienced, progress will only accelerate.

How about risk mitigation? How are banks using generative AI here?

Mitigating risks is huge: fraud detection and prevention, predictive analytics, and AML/KYC, along with the rest of governance, risk and compliance (GRC) are potentially huge beneficiaries. But if you don't have a good data infrastructure in place right now, you're going to be hard pressed to get the full value out of generative AI. For fraud in particular, the large language models can ingest a huge amount of internal data and look for patterns and anomalies much faster than an individual person. They can flag those to a human analyst, but the analyst still needs to go in and investigate authentication in traditional and nontraditional ways. For example, they may match duplicate IP addresses across different applications initially flagged by the AI. 

How should institutions handle bias in AI algorithms?

You've got to be transparent about it, let your employees and your customers know when you're using it and be sure that you can explain to regulators. You want to have a robust governance structure in place, and make sure that your entire GRC mechanism takes account of your gen AI-oriented processes. 

We are excited to have you lead research efforts at The FR. Can you tell us a little about the initial initiative you’ll be leading? 

I’m spearheading Most Impactful, a recognition program that identifies firms in financial services that are making a difference. We’re starting with those institutions — banks, insurers, and capital markets firms — that are creating impact with AI, and we’ll expand from there. It’s exciting to be leading this groundbreaking research effort at a time when consumers and businesses need help from their financial partners more than ever. We’re flipping the process by selecting firms based on objective, publicly available criteria rather than asking them to submit nominations. The firms that we recognize will be able to let their stakeholders know that they’re making a difference in the world of financial services.