Baby Don't Hurt Me: Bankers and AI

This is one of those learning and sharing moments that I just get so excited to do.

You pumped up? Are you pumped up too?

I am pumped up.

I am excited.

There’s just been so much talk over the last few months about artificial intelligence, generative AI, and the opportunity to really dig into this topic with somebody who’s written some exceptional thoughts.

It was something I just couldn’t pass up.

I’m Al Dominick. This is Steve Williams. We’re being joined in the proverbial studio by our good friend Sarah Hinkfuss from Bain Capital Ventures.

Sarah, thanks for taking some time with us today.

Hey, Sarah.

Thank you, guys. I’ve been looking forward to this. Can’t wait to get into the content.

We’re going to geek out a little bit on some of the things that we believe bank executives are trying to get their arms around right now.

Steve, I’m going to challenge you to challenge me, if you will.

It seems like the banking industry is really at this crossroads. It’s at a period of change that we probably haven’t seen since 1918, when the Federal Reserve was stood up.

The banking panics that we’ve seen over the decades really came home to roost this spring.

If I’m running a business today, I’m trying to figure out what my future looks like.

I can take my optimist hat, put it aside, and say, “Are better days ahead, or should I be feeling somewhat pessimistic?”

We talked about Matt Harris last week and his prediction that this may be the end of 8,000 banks in America.

I go back to the Fed in 1918, the Great Depression, and the creation of deposit insurance.

It was a business-model shift that was huge.

I think what we just saw with SVB and Signature is another one.

Because of the influence of fintech, someday decentralized finance, and now AI, you layer all that onto digitization, which has occurred for 20 years, and I think we’re at that breakpoint.

How does the business model evolve?

What does the final structure of the banking industry look like 10 or 15 years out?

Absolutely.

We’re in a really interesting spot.

I was afraid I got up on the wrong side of the bed today when I thought, “Is the business of banking a little bleak?”

I don’t know that it is.

Hopefully, Sarah, who works with Matt, will be able to shed a little bit of light on what she’s seeing now.

We did warn her that at Plugged In we have an affinity for pulling some classic tracks off the proverbial shelf and using music to inspire our conversation.

You got really poppy today.

I went super poppy.

You went super poppy.

In the past, we’ve talked Led Zeppelin and Frank Sinatra.

Today, we’re going totally different.

We’re inspired by somebody who’s having her own moment of fame, and that’s Taylor Swift.

We have the Barbie movie, which has done some pretty awesome things, so Dua Lipa is going to make an appearance at some point.

David Guetta has some “Baby Don’t Hurt Me” lyrics that I think we can use.

OneRepublic made an appearance during Top Gun, and who doesn’t like the Top Gun: Maverick reboot?

We’re not always going to go with these pop stars.

We’re going to have a little country fun with Morgan Wallen as well.

Those are going to be our five artists that kind of get things going.

But Steve and I are doing a lot of talking.

The person we really wanted to talk with wrote, with her partners at BCV, a really interesting Field Notes directory on how and why financial services need generative AI today.

We took the opportunity to reach out and get connected so that we could have this conversation around this new industry, if you will, that’s doing some pretty amazing things already.

I mentioned we’re going to start with Taylor Swift.

She’s got so many songs to choose from.

So many.

But we know them all.

It’s amazing.

Her repertoire is incredible.

We’re doing some pretty good things on the Swift side at Cornerstone, but we’re going to use “Cruel Summer” because “Devils roll the dice, angels roll their eyes” is a lyric that we can use to ask you to take us back a step and talk about the catalyst that started the pragmatic and widespread use of artificial intelligence.

Absolutely.

AI is nothing new.

Artificial intelligence has been leveraged within financial services for decades.

Financial services, like any other industry, is always thinking about how to do business better.

That’s a part of leveraging the latest, greatest technology.

In fact, AI has been pervasive in financial services.

One of the reports that I love to quote is a 2022 report published by NVIDIA on the state of AI and financial services.

This was published before the generative AI hype, to be really clear.

They actually identified that over 75% of financial services companies, across the different components of financial services, utilize at least one of the core computing cases of machine learning.

Now, with the advent of generative AI, we’re calling that “traditional AI,” which I think some people may scoff at.

But it’s not to say AI is anything new.

Rather, all of the hype we have around generative AI is a new form of AI that is complementing, for many intents and purposes, what has been a pre-existing part of the technology landscape that’s been critical to the way financial services are distributed, manufactured, and serviced.

It seems like machine learning was still back there with the experts, engineers, and programmers.

The thing about generative AI is it brought it right to the front, didn’t it?

It really brought it to the layperson.

I agree and disagree.

Again, back to this report, the second most common use case of traditional AI was actually conversational AI.

Twenty-eight percent of financial services institutions were using traditional AI in the context of providing service to customers.

But the problem was that it wasn’t fully complete.

I think we’ve all had the experience of a very frustrating bot that we’re trying to communicate with.

You’re trying to get an answer to something as simple as, “What’s my checking account balance?” and the bot doesn’t fully understand or doesn’t get the language you’re using, so you don’t get the answer you want.

I love that example because what we’re seeing now in the leap forward that generative AI is taking is that we’re able to have a much more empathetic and complete conversation with a bot because it is language-based rather than just rules-based, which was the domain of traditional AI.

Interesting.

Oh, the bots that we’ve had to deal with.

I think my big thing on this is it’s not an either-or moment.

I think that’s what’s sometimes missed.

When there’s a platform shift in the way people talk about it, it’s often a conversation around, “Out with the old, in with the new.”

That’s not at all what’s true in this case.

I think, especially in financial services, this is a both-and.

We’re building on this scaffolding and architecture of traditional AI and all of the ways that it’s really enabled business to be much more productive.

Then we’re adding on top of that and filling in the holes with generative AI in the places where it’s superior.

I really liked your buildup.

You’ve written about the different milestones that are now kind of accumulating.

It is an “and,” not an “or,” but there are some really big milestones in the last 10 years that you highlight.

Absolutely.

On this point around why is it now and why has it captivated our attention in such a way, I think something that’s really neat about this technology is that it is such a network-effect-based technology.

From the very beginning of ChatGPT launching publicly, it was something that anyone could get on and use.

There’s a very low barrier to entry.

There were these aha moments that made you want to share it with someone.

“Oh my God, look, I can look up a recipe for dinner tonight.”

“I can ask what type of wine I should have.”

“I can come up with a name for my baby.”

Not saying I’ve used that one.

There are all these cool things that you could do and share.

That matters because rather than it being, to your point, something hidden in the back with a bunch of PhD statisticians, it was something that actually started in the forefront, in the layperson conversation.

Because of the way that everyone can access it, those who are in positions of leadership have been forced to address it and figure out how they’re going to leverage it internally.

What’s funny is, as you’re saying this, Sarah, I had a chance this summer to speak to an audience.

This was when ChatGPT was wicked hot. Everyone was talking about using it.

I said, “Okay, let’s use some practical applications. How many of you have tried to have it write your own resume so you could find a job?”

I thought nobody was going to raise their hand.

There were like four people who were like, “Yes, I did.”

I thought, “Okay, you’re at risk for leaving the company.”

Interesting data points.

I mentioned we were going to talk Dua Lipa at some point.

I have not seen the Barbie movie, but I think some people probably have at this point.

Dua Lipa’s song “Dance the Night” kind of makes me think, when you go out to a club, at least when you were single, you probably weren’t just dancing with one person.

You danced with a few different folks.

I was looking again at this Field Notes guide that Sarah put together.

In it, she talks about three models for leveraging generative AI in the financial services space.

One is construct, one is configure, and one is adopt.

If you were to borrow from Dua Lipa and talk about each one in its own way, help us understand the dance partner that you’re looking at for those three models.

Absolutely.

I used a different analogy in the piece, which is baking a cake, but talking about dancing in a club is even better.

We can bake a cake if you’d like.

Three different dance partners.

Thank you for the introduction to this.

I would say this framework is probably the one that we most often talk about when we’re meeting with companies in financial services because it’s such a clear way to make sense of how to approach generative AI.

Exactly to your point, for a given use case, an organization could approach embedding generative AI in a completely different way.

We see that as a function of the resources they have available, as well as the goals or strategy of where generative AI fits into that use case.

The three different models, exactly as you said.

Construct is to create something from scratch.

I don’t know if we’ll be able to fully get there with your club analogy.

We take some liberties here.

We take some liberties.

Back to baking the cake, this is if you’re creating something from scratch.

You have the ingredients.

You have your flour, eggs, butter, vanilla.

You’re combining it, mixing it, then putting it in the oven and preparing it.

As it relates to generative AI, this is actually creating your own LLM trained on your specialized data.

This is, of course, the most work.

It requires the most specialized talent, and you have to have your data already arranged and structured in the right way.

But it’s also the most proprietary because no one else can replicate exactly what you’ve done.

It’s not just your data, but it’s also your workflow and your distribution as well.

The next is configure.

This is assembling from a kit.

This is as if you were buying a box of cake mix, then combining maybe vegetable oil and water and baking it, but most of it was already ready to go.

This is leveraging existing LLMs and then adding on top of it some level of customization.

That could be fine-tuning.

That could be embeddings.

There are many different iterations of this that we’re seeing.

This is one of the most active areas of innovation today in the space.

Then the third is adopt.

That’s basically just buying your cake.

You’re transacting with another vendor who has developed, usually, an application that leverages generative AI that can fit into your ecosystem.

Maybe I’ll just give you examples of each of these three, if that would be helpful, in financial services.

The piece that you’ve been referencing actually has 93 examples of real applications.

I’m just going to pick a few.

There actually are a lot of real ones.

On the construct side, one to mention would be what Bloomberg has built with BloombergGPT.

Bloomberg, obviously a data services company, extremely successful for the last 40 years, has an incredible amount of data that they’ve collected and constructed.

What they were able to do is basically work with the Amazon team, the AWS team, and Bedrock with all their tooling layer to construct their own LLMs with a set of researchers internally that leverage all their data.

Now they’re in a position where they’re trying different use cases within the company to see where that LLM would be most productive.

They’re looking at new ways to interact with the Bloomberg terminal, new products they’d be creating, and new ways of creating research reports.

It’s kind of like a font that’ll keep giving, based on this incredible data set of 40 years of longitudinal public-market data.

One example of configure, and this is actually where you’ll see in the report there are the most examples across fintech companies today, both fintechs and incumbents, which we can get more into as well.

A few of the applications that we’ve seen are in the underwriting process for loans or insurance.

Slope is one example of a company that’s doing that today.

Another is enabling your users to automatically tell the story of data.

Pigment is a company in the FP&A space, and they’ve released a new feature within their product that helps you go from, “What was the variance analysis within a company?” to, “How do I tell that story to the executive team?”

It makes that a lot easier and really makes heroes of their users.

Another example is payables categorization.

How do I match the expenses themselves to the expense categorization, so whether it’s approved or not approved, and then also receipts to the actual categories?

There are a lot of companies doing that.

That includes Brex, Vic.ai, and Puzzle as well, which is a really early-stage company in the accounting space.

Then finally, on the adopt piece, and this is where I’d mention as well the question you raised at the beginning for your banking partners.

One example is Kasisto, which is an out-of-the-box conversational AI company that supports customer-service applications.

They released at Westpac, the bank in Australia, as one of their early customers on that platform.

There are plenty of examples where, even if it doesn’t make sense for a company or bank to actually build this internally in the construct or configure category, adopting out of the box is a great way to get started with a use case that they think is highly strategic within their company.

We’re seeing a lot of Kasisto interest among our regional and community bank space.

Part of that, to your point, Sarah, is that they are building kind of a one-to-many library of use cases and interfaces that have to plug into the legacy bank environment to make it work, whether that’s the core system, digital banking system, or phone system.

I think a lot of our clients will probably start with adopt.

But over time, I love where you’re going, which is, “What’s proprietary about us that might someday become, as we mature our team, that construct layer or that configure layer?”

As you’re saying this, Steve, and I’m thinking about what Sarah is sharing, change is not cheap.

When you think about the different models that you’ve given, you talk about making cake.

Sometimes you just want to go in and buy the cake off the rack and not have to worry about it.

I think that’s where a lot of the banks we spend time with probably will start.

“That cake looks tasty. Let me see how it actually is.”

The maturity of some of these larger organizations allows them opportunities to do some pretty cool things.

But I just look at this whole space as shifting really dramatically, at least from public perception, even if it’s been going on behind the scenes much longer.

This is where I want to put myself again in the role of one of our troublemaking clients who’s saying, “I need to be able to separate the hope from the hype from where something is actually happening.”

Again, because I’m being a big music nerd right now, I’m going to use David Guetta’s “Baby Don’t Hurt Me,” because the lyric, “What is love? Baby, don’t hurt me,” I think that’s how some bankers feel when it comes to AI.

They are probably using it.

They probably know they’re using it.

But they still would be a bit trepidatious.

One CEO in my client base says, “I need an offensive strategy, but I need a really big defensive strategy on this as well.”

Yes, totally.

Sarah, any thoughts on this?

How can people start to separate that hope from the hype from the actual, “This is going on today,” and from the threat?

I hear you asking two questions.

One is, when you hear announcements or proclamations, what is actually real versus not?

Then secondarily, how do I make sure that, as I navigate as an organization, I’m staying on the right side of it and not letting it threaten my business as well?

Maybe taking the first question on what’s actually real.

We confront this all the time as investors.

I think the joke was, when generative AI hype started, all of these companies all of a sudden were “AI” at the end of their URLs.

Trying to really separate out where someone has an actual strategy, and furthermore where is it differentiated.

We see companies that are in the hype phase as those who have broad pronouncements about the intent to use it, but they really can’t answer first-order questions around the what.

What models are they using?

Where in the organization is it being used?

Who is the talent they have leading it?

How are they implementing it or creating value in production?

That last point really matters because it’s not just about introducing something.

It’s about actually capturing the value.

They’re not the same.

If it’s hope, where we see the further evolution of hype is that there’s a lot more specificity.

They can talk about the data sources internally that are relevant.

They can talk about the users.

They can talk about the partners they’re using.

The named vector-database company, the named security company, or the named model-testing company to figure out which of the various models is the best one or in what combination.

They also know the particular value that customers are getting out of including it.

Then finally, in the small number of use cases where it’s actually something happening and these use cases are fully in production, they’re able to capture the particular value of higher pricing, more customers, higher conversion, higher retention.

Whatever is that funnel accelerator that they’re seeing from generative AI.

They’re also able to talk about the early proof points of winning against competition due to these advantages.

Those are kind of the layers that we see to separate through it.

The one caution I would have, though, is not to underestimate the value of hype, especially in financial services, which is such a necessarily conservative industry.

There are really good reasons to be conservative because of the regulatory framework and the responsibility institutions have to their customers.

It is easy to continue doing business as usual and not see the opportunity for some of these more transformative technologies.

The great thing I’ve seen in conversations is that when there’s a lot of hype, like Morgan Stanley is a large bank that’s made a lot of noise around the great initiatives they’re doing here, I think that requires everyone else in their board meetings to really talk about, “Okay, so what is our strategy?”

Exactly to your point, there are some things that are scary.

But how do we understand what those are off the bat and then build around them rather than using those as excuses to not go down the path at all?

I’d love to amplify what Sarah’s saying because you and I have talked at length about how banks can be too risk-averse and could be using the moment that they’re in right now as the reason to pull back from some of their ambition.

Just because it’s not 100% ready for their use doesn’t mean they can’t start to think about, “How do I step into this? Where do I create some opportunities?”

I think this lends itself to a question around creating competitive advantage, higher performance, and performance layers of data and capability.

I’ll spare you my Top Gun: Maverick music reference, but just know that I was thinking OneRepublic on this one.

I do think it’s important to think about performance.

How do you create that little separation from others so it’s not just a commoditized opportunity?

I worry, Sarah, that the fintechs are going to be laser-focused on building those proprietary LLMs and the workflows.

They’re going to have an entrepreneurial approach to it.

To your point, I worry the banks might be too diluted about that and not having that conversation.

They can write a marketing letter using ChatGPT.

That, to me, is operational. It’s not strategic.

Are you seeing the fintechs get to very focused roadmaps to value capture faster than traditional institutions?

Or are traditional banks looking in that proprietary layer already?

I think there’s a huge opportunity for partnership here.

If you look at the last big wave of innovation in financial services, it brought us the BaaS model and all these other models that had partnership at their core.

I see that being true here.

The reason is a lot of the fintechs and these native LLM companies have incredible talent ecosystems that they’re a part of.

That is really hard for a lot of incumbent financial institutions to access.

At the same time, incumbent financial institutions, the banks, have this incredible set of existing customers and data that is not available.

If you see it, both sides have something that the other needs.

I think we’re going to see a lot of really productive partnerships.

They’ll come at those points of strategic advantage.

It will be in, how do customers get better served and understood?

How do wealth managers put themselves in that position?

How can I better underwrite bespoke customers who have until now been unprofitable for the bank because they’re so complex in terms of what their credit history looks like?

There are going to be these spaces that look disruptive, but the way that both sides will be able to create value is that they each need something from the other.

Do you think you’ll see kind of consortia between financial institutions?

Meaning, if we’ve got data, why not amass it with other institutions to create a proprietary pool, but maybe co-owned by multiple financial institutions?

There are a lot of great examples of that from financial services history.

Just thinking about payments, in many ways Venmo and PayPal are what launched Zelle.

That was obviously a consortium enabled by Early Warning Services.

A consortium among banks has become more popular than Venmo.

I do think we’re going to see very similar things like that as well.

The question mark I have around that is the regulatory context.

How will regulators think about the pooling of data?

What privacy will they want banks to ensure to their customers?

If you could tell us more about the regulatory crystal ball, we would appreciate it.

There are some things that we know Director Chopra at the CFPB said.

“Just let your data run wild. We’re not worried.”

I don’t think that’s what he said, Steve.

I don’t think that’s what he said.

All right.

We’re sitting here, and we’re going to wrap things up.

I’ve got to tell you, watching this evolution of technology, I kind of feel like we’re all maturing in different ways, but we’re all still kind of young in our understanding.

When I talk about this being an opportunity to learn and to share, I’m going to share a lyric by Morgan Wallen from his song “This Bar.”

It’s all around when he was starting at 21.

He says, “Making mistakes and making new friends. I was growing up and nothing made sense. Buzzing all night like neon in the dark. I found myself in this bar.”

I took that lyric as he’s kind of new to the world.

He’s trying.

If you listen to it, he’s drinking a little too much and getting into fights.

Made me feel like a cold beer just now, and it’s 11 in the morning.

I’ll take care of that later on.

But I think you walk into a bar and you’re 21 years old, you’re full of wonder and excitement, and things are cool.

I think that’s where we are, at least in the banking space, with generative AI.

What are you, Sarah, looking forward to and watching come over the horizon?

Absolutely.

To amplify your point, I think this is such an exciting and history-defining moment for all of us as well.

It’s changing so fast.

There’s such potential and possibility.

I personally have never had more fun intellectually than in the midst of all of this.

Some of the things that we’re watching and really excited about.

One is domain-specific models.

There’s been a lot of focus, obviously, on the really big LLMs that we all know of.

OpenAI GPT-4, Anthropic.

You have a lot of these really big models, and they’re widely generalized across a lot of different use cases.

But there is the opportunity, such as I talked about with BloombergGPT, to create a model.

There are a lot of very specific models within domains that are useful in cases when the data that’s relevant to that domain is proprietary and nonpublic, which is the case in financial services.

We’re really excited about seeing these domain-specific models be created.

As one aside from the news, I was just reading yesterday in The Wall Street Journal that we’ve actually seen that the generalized models, because of a concept called drift, there’s an AI lab at Stanford that identified this, generalized models are actually getting worse at math.

You actually need those other models to compensate and be able to work within a domain, even if the average question they can answer is not as broad, because it doesn’t matter as much if it’s not within financial services.

That’s one.

Another one I would say is talent flows.

I talked about this already, but I think we’re going to see that the limitation in the market and development will be talent.

Who attracts talent, retains talent, and is able to generate additional intelligence in this talent ecosystem?

Then regulation, which you touched on very briefly as well.

I think you’re right.

What these big models have done is create a golden ticket for a ton of niche, domain-specific things.

The real wealth creation and explosion is going to be this huge tapestry of domain-specific stuff.

For our clients, I think we’ve got to be real hunters of domain development and what capabilities and value capture are being built.

I love your point.

I don’t think we can win the talent war in banking.

I think we have to partner to access the talent.

New ways of partnering have got to be on the strategic agenda for a lot of banks.

I think I might go update my LinkedIn profile and just say I’m a prompt engineer, thanks to Plugged In.

I’m a senior prompt engineer, just so you know.

Our starting salary is going to be probably what, $375,000 plus some benefits?

Interesting times.

We can’t say thank you enough to Sarah.

I really encourage people, if you haven’t read what Bain Capital Ventures just put out, it’s super accessible.

It’s a great read.

It’s not going to take you a long time.

It is going to spur a lot of ideas and hopefully conversations because there are just some really amazing things going on.

Steve, I appreciate you taking the time.

Sarah, I appreciate you taking the time.

And everyone, thanks for getting Plugged In again with Cornerstone.

Thanks for you all taking the time.

Thank you, guys.

My pleasure.