
Transcription:
Penny Crosman (00:04):
Welcome to the American Banker Podcast. I'm Penny Crosman. In a recent survey, Boston Consulting Group did of more than 1,800 C-level executives around the world, 41% said they expect their companies to spend more than 25 million on AI in 2025. Yet across the board, only 25% of executives said they're seeing significant value from AI. We're here today with Vlad Lukic, managing director and senior partner at Boston Consulting Group and one of the authors of the company's latest AI report, to talk about this gap between AI investment and return. Welcome, Vlad.
Vlad Lukic (00:42):
Thank you. It's great to be with you.
Penny Crosman (00:44):
Thank you. So the report that you co-wrote said that BCG expects generative AI investments to increase by 60% in the next three years. Why do you think that is? Is this just the inevitable direction that technology is going? Are companies becoming dependent on generative AI for certain parts of their business, or is there some other reason why you see such a huge increase in the investment in gen AI?
Vlad Lukic (01:17):
Yeah, I would say it's all of the above, right? The technology at our fingertips. I was just with my 11-year-old showing her how to use ChatGPT, and she's like, oh, but I use it already. I'm like, wow, OK. Right. So it's there. We are using it in our daily lives, and it's fascinating. You talk to the, even with her, I was like, "Hey, but honey, do you know it can give you wrong information?" She's like, oh, yeah, I know ... I'm like, OK. So it's present in our personal lives. It gives us the art of the possible is at the fingertips and therefore by default we are expecting to see it in our workplace as well. So one, it's kind of grassroots, it's happening, it's available. Two, the costs and availability are just dramatically decreasing. We still have even just the new models that were released out now that are open source, et cetera.
(02:12):
So I anticipate the costs are continue to keep decreasing while the performance is going to keep improving, and it's moving at such a rapid pace that it's not leaving anyone oblivious to it. So everybody's like, we got to do something. So there is genuine desire given those things that I mentioned, genuine desire to lean in and make sure that organizations are ready to benefit from it. And it's fascinating. It's also not coming from fear, which is like a lot of the media narrative can be around, oh, is this going to take away the jobs and this is going to replace us, et cetera. It's coming from genuine excitement to, and this is me synthesizing general sentiment from folks that are leaning in. It's coming from a genuine sentiment of it can help eliminate toil from our day-to-day jobs and eliminate things we don't like doing. Two, it will kind of add a superpower into our toolkit to move faster, and three, it helps the companies grow faster than they could have been all essentially constrained with the labor and number of employees that we can bring in, and this allows us to keep scaling up with the teams we have.
(03:23):
So those are a few different factors and reflections.
Penny Crosman (03:27):
Your team identified three specific ways companies can get value out of AI. One was deploying AI in everyday tasks, and you felt that they could realize 10 to 20% productivity improvements that way. What are some examples of those everyday tasks that you're seeing being automated with AI?
Vlad Lukic (03:52):
And if I think about in financial services, it could be as simple as using it to do a first pass of analysis of a specific document. If there is a report of some source before you read a hundred pages, you can put it into one of these tools and get a summary in few minutes and you get a sense for the key points, and now you still will go and read the full report, but you're already sharper in how you're going to go about it. Or if there are 30 reports and you're expected to analyze them, you can get a summary of key themes literally in a few minutes across all of them versus spending a few days ... so you can get faster [in picking] up the information across different sources. That's one simple one. The second one is in generating first drafts of documents and doing a first pass.
(04:47):
You can now focus on editing and making sure the key points come out, versus starting from scratch to write a document. Another interesting one is ... I was with a number [of wealth] advisors where they would say, I need to have a bunch of conversations with clients first. Then based on those conversations, I create a first draft of a portfolio assessment. Then we have a conversation again, and there's a lot of steps, and then every single one of those steps you can shave off a lot of time by deploying some of these tools. So that piece of deploying is really just thinking about giving the tools into people's hands to use in their daily jobs. And the work ends up being not just giving the tool into the hands, but teaching them how to and having them creating a forum for them to share best practices and learnings around "When do you trust the tool?" "How do you validate that the quality of output coming from it is right?" "Which tools are better for different tasks and how do you compare them and how do we access different ones depending on the task," et cetera. So that's all within that bucket of deploy.
Penny Crosman (05:57):
So then the second area where you guys determined companies can get value out of AI is reshaping critical functions. I think companies can get 30 to 50% improvement in efficiency that way. What's an example of a critical function that these companies can reshape?
Vlad Lukic (06:17):
So that was actually a very interesting insight over the last year or so as we're seeing the tools being deployed, where we see the biggest value unlock is in reshaping the core functions of a given company. So if you're a software company, that's software engineering and writing code. If you are an insurance company that is rethinking the underwriting process or claims process. If you are in biopharma, that is speeding up the R&D process and increasing the quality of the first pass of analysis that comes. And the value comes not by just deploying the tool, which was in the first example that I shared, but in rethinking the end-to-end workflow. So for example, if in insurance in a claims process, you have, let's say, it's around car accident claims ... you have an accident currently you need to send an image and have maybe some phone calls before it triggers.
(07:19):
Then a set of workflows that are one agent is compiling all the information about the accident and photos, et cetera. The other agent is compiling all the information around your actual policy that you have. The third agent is compiling a document around current claims from your area. All of that then comes together to a fourth agent that then compiles all of that into an assessment of your individual situation. All of that goes to a committee that reviews all of that. And if there is stuff missing in the information, [it] goes all the way back to you or one of those agents to refine the inputs. And then when it's already it goes to the committee that decides what they're going to do, and then you get informed and everything I just described can take 10 to 20 days. And with infusing gen AI into these workflows, you can essentially bring the individual steps down to minutes of generating those documents, et cetera, and the process can be done in probably half a day or a day.
(08:19):
The issue is that the committee that meets still meets only once every 10 days, so by the time you hear back, it's still going to take 10 days. And so what we mean by reshaping is do we even need a committee? And you start asking, why did we have the committee in the first place? Well, when humans write those documents, the accuracy is maybe 50, 60% in acceptance rate, and the rest of it needs to go back at the beginning. When we use the gen AI tools here, the accuracy is 90% plus. So then maybe we don't need the committee that does these approvals. We can automate that step. If we eliminate that, we can get down to one-day turnaround on it, and maybe we don't need all the agents, the workflow, we can liberate them to actually do other tasks. But everything I just described needs to be a set of deliberate conversations and assessments and a set of deliberate choices to then, at the end of it, rethink the whole workflow and start using these tools as a natural part of it. Does this resonate? Does it make sense?
Penny Crosman (09:21):
Sure. And in financial services, you could probably apply the same ideas to a mortgage or some other kind of loan process.
Vlad Lukic (09:31):
Correct. Initial mortgage assessments or prequalification, right? Or initial investment advice and suggestions on how do you invest into, what should be the mix of your portfolio. Those are very similar analogs to the claims process I just described.
Penny Crosman (09:50):
Sure. As long as people don't fall between the cracks, because something that I sometimes worry about with things like insurance claims or loans. But that's a broader conversation about how to make sure that there's thoughtfulness and not just blanket giving "nos" to people because the AI came up with some reason to say no or something.
Vlad Lukic (10:17):
Absolutely. And this is where the question is in a steady state versus in an interim state, right? And in the interim state, humans in the loop need to be aggressively part of it, that until we get the confidence that these systems have been trained the right way, even then we need to design human intervention and choice in between the steps to make sure that the quality control is there. That's one piece. The other one is something that we call purposeful toil. Let's say that you have your best agents right now, even in the advisory are the folks that have been doing their job for a long time. They've also built intuition and because they've done some of the work themselves, and now if the junior ones never have to do some of those tasks, fast-forward 10 years from now, they will not have the intuition and will not be able to even quality-check some of these systems. So you need to design what we call purposeful toil into some of these steps where some of them need to be remain manual so that you're building the skills of your core workforce as part of it.
Penny Crosman (11:24):
So then the third area where you said companies can get value out of AI was in inventing new products and services that could give them a competitive advantage. Have you seen anyone do that yet — invent a new product, incorporating AI — or what might be a possibility there?
Vlad Lukic (11:43):
Yeah, so the example I use is coming from a non-intuitive place ... just to illustrate the power of this and how one needs to think end to end as they reshape the offer. So let's say you are in the construction business and you're selling a home and the current pitches, it's going to take 18 months between now and the completion of the task, it's going to take actually anywhere between 12 months and 18 months between completion of the task, and then you get the keys and the price is X. Now, if I use some of these tools in the design process upfront where I pre-align with the customer, if I use the tools to actually do the scheduling of the trades, if I use the tools to manage the delivery of the right components on the job site as well, et cetera, and the outcome of that is that I can consistently deliver these homes in eight months.
(12:46):
I can still pitch 12 to 18 months to the customers in the market, and I have the efficiency and I just delight them and I'm a little bit more efficient in how I do my work. But if I rethink the pitch to the market, I could also say, you know what? I'm not going to pitch 12 to 18 months and get that buffer. I'm actually going to pitch that we're going to deliver this house in eight months, and by the way, it's going to be 10% more expensive than if you were to do it with others. But we also guarantee we will deliver it in eight months, and for every day that we go over, we will pay you the penalties. We are convinced that we're going to deliver it, and I increase my conversion rates. So I'm now selling something for more money with higher profit.
(13:31):
I am delighting them, and I'm essentially now capitalizing on that totally different capability that I have. I have a step change in performance, but I've totally changed the offer in the market. It's not any more 12 to 18 months for delivering the house. It's actually in eight months and with penalties if I don't deliver. So you actually as a customer end up with a better deal and along the way, full transparency and visibility into it, and you're involved in the design process. So I'm just totally rethinking a category. So we are seeing more and more of those types of examples across the industries where sometimes it's hard to do it with your existing offers as companies would launch a parallel business under a different name to just provide that service at a totally different price point and a different level of experience. But I'm seeing that those who lean in are actually getting a lot of traction.
Penny Crosman (14:26):
Given all of these things, why do you think it is that when you guys surveyed all these C-level executives, only 25% said they're seeing significant value from AI. Do you think it's because it's too early and people just haven't built these things out yet? Or do you think people are maybe not choosing the right projects, or is there some other reason why they're not getting this value so far?
Vlad Lukic (14:55):
Yeah, so there are a few different things. One is in some instances, they just appoint the wrong team to drive it without accountability for results or without organizational position to actually drive impact. For example, a lot of clients have really well-established data science teams, and they are the ones who the CEO asks, Hey, you go deploy this gen AI thing. Well, deploying gen AI is different than building custom models or analysis. And the data scientists are not anthropologists or people that know how to rethink business processes. So they will usually take these models, they'll keep fine tuning them, increase the accuracy, but then they depend on the business process owners to deploy it. And if they're not given an explicit mandate to drive it, they usually just build a tool and then they're waiting for business owners to do something about it. And that's a very frequent failure point that we see.
(15:56):
And so the unlock of that is that you actually appoint business process owners to drive this, and they can be given access to the data science and other technical teams to deploy this, but they are on the hook, they need to drive it. That is a very typical part of the reason that we see. Second one is a lot of them are still seeing this as a silver bullet type of opportunity. We'll just deploy these tools, we'll give them to people and then magic will happen. Well, let's say I'm using the tool now in my daily job and I can do my task now in two hours versus eight hours, but if I don't have a conversation with someone, what do I do with the six hours that I've freed up? I'm not going to redeploy myself onto something else. So there needs to be an explicit conversation around, if you free up the time, what is the time going to be used for?
(16:49):
And then the third one is, many times companies will focus on the wrong metric or wrong KPI and will therefore not result in efficiencies. For example, the example that I gave earlier on the claims process, right? The bottleneck in that whole workflow was the frequency at which the committee meets, which is once every 10 days. And what the team was measured on was can they do some of those tasks faster? And the answer was absolutely. They brought some of those tasks from eight, nine days down to minutes. And so the KPI they measured was there and was a huge success, 90%-plus reduction in time on task. But KPI should have been the overall outcome. Will the customer get a response faster than the current 10 days? And because if they had measured that, then the question of the committee would've come up much earlier in the process and they would say, "Let's remove this constraint and the bottleneck, and then I can actually benefit from the efficiencies." Because they didn't do that, they just freed up some of the time of the agents, but there was zero impact on the customer and the speed of the whole process. So those are few of the factors that I see why the companies don't realize the benefits of this yet. The good news is they're building the muscle and this is going to be something they're all going to be leaning in on in the coming years, right? And they're just going to reenergize it and rebuild it.
Penny Crosman (18:24):
That makes sense. Well, Vlad Lukic, thank you for joining us today and for all of you, thank you for listening to the American Banker Podcast. I produced this episode with audio production by WenWyst Jeanmarie. Special thanks this week to Vlad Lukic at BCG. Rate us, review us, and subscribe to our content at www.americanbanker.com/subscribe. For American Banker, I'm Penny Crosman and thanks for listening.