One year in: Lessons learned in scaling up generative AI for financial services
In the year or so since generative AI burst on the scene, it has galvanized the financial services sector and pushed it into action in profound ways. The conversations we have been having with banking clients about gen AI have shifted from early exploration of use cases and experimentation to a focus on scaling up usage. The technology is now widely viewed as a game-changer and adoption is a given; what remains challenging is getting adoption right. So far, nobody in the sector has a long-enough track record of scaling with reliable-enough indicators about impact. Indeed, some financial institutions have gotten off to false starts. Yet that is not holding anyone back—quite the contrary, it’s now open season for gen AI implementation and the learnings that go with it.
Much has been written (including by us) about gen AI in financial services and other sectors, so it is useful to step back for a moment to identify six main takeaways from a hectic year. With gen AI shifting so fast from novelty to mainstream preoccupation, it’s critical to avoid the missteps that can slow you down or potentially derail your efforts altogether.
- Make AI a strategy centerpiece
Gen AI amounts to a big bet, and as such, implementation and scaling need to be front and center. This is a core strategy topic and needs to be addressed at the CEO level. Financial institutions that have made the most headway have driven their efforts from the top. Treating AI strategy as a CEO-level topic both energizes the organization and eliminates potential bottlenecks. Moreover, the top-down prioritization that flows from elevating AI to top-level strategy means that efforts are sponsored and funded in ways that help gain momentum. We are already seeing a big change in thinking along these lines: at a recent roundtable with 26 data-and-analytics banking executives, some 60 percent said gen AI is now a top-of-the-house strategic priority that is spurring more widespread adoption of the technology.
- A centrally led organization is the key to scaling
Financial institutions using a centrally led gen AI organization are reaping the biggest rewards. A review we conducted of gen AI use by 16 of the largest financial institutions in Europe and the United States showed that more than 50 percent of the businesses studied have adopted a more centrally led organization for gen AI, even when their usual setup for data and analytics is relatively decentralized. Conversely, efforts have sometimes stalled, including at some leading banks, when organizations have launched more peripheral experiments that failed to gain traction. No single business unit has the scale needed to build the AI “scaffolding” for the whole organization, and thus harness the full power of gen AI. The case for centralization is strong when important decisions need to be taken on matters such as funding, tech architecture, cloud providers, large language model providers, and partnerships. Risk management and keeping up with regulatory developments is also easier with a centrally led approach. Centralization also allows a company to allocate talent in a way that is more likely to benefit the entire organization. For all the advantages that central leadership brings, there is also a risk of straying too far from core business. Striking the right balance is essential, and may involve centralizing only for a limited time.
- Sequence the gen AI roll-out across domains
To date, very few institutions have seen material benefits from Gen AI at scale. While we suspect that this will change over the next 3-6 months, the approach that some have taken to launch dozens of pilots at once is not leading to sustained impact. Instead, those institutions that are narrowing their scope and focusing on scaling 1-2 lighthouse domains appear to be getting more impact momentum.
To realize gen AI impact, financial services institutions will need to continue to reimagine the domains of application and deliver against the full range of opportunities within each domain—both simple and complex ones. Gen AI is a lever, not a silver bullet, and we can already see some sequencing at work. Early momentum has largely focused on customer servicing—mostly agent co-pilots with some banks offering a chatbot or thinking about it, alongside software development, mostly in the form of coding assistance tools for new code development, documentation, and testing. We’ve also seen some front-line support for wealth and commercial banking. More recently, there has been movement to domains that are in the thick of operations and highly regulated, including back office, credit risk, and know-your-customer. There are some moves to more complex, end-to-end automation for customer servicing, for example linking multiple agent co-pilot use cases including call summarization, knowledge access, post-call fulfilment, and seamless experience, as well as more complex coding transformations, such as legacy code base migration. Next up will be bringing together multiple domains seamlessly, including front and back office, outbound communications, and inbound chatbots. All of these applications will need to be integrated with complementary advanced analytics capabilities; hyper-personalization is one example of where traditional recommendation engines work hand in hand with gen AI.
- Robust (and reusable) scaffolding is crucial
Building (and trying to scale) pilots without a robust enterprise AI infrastructure can result in increased risks and Model Risk bottlenecks. The “knowledge scaffolding” that a financial institution builds for gen AI needs to be extensible and reusable. There is both a “what” and a “how” element here: the “what” is the product innovation you provide customers, the productivity gains you can extract, and the competitive shifts in market dynamics that need to be considered; the “how” is the operating model, the prioritization of use cases, the data, tech, and tooling aspects including data architecture and governance, and the risk and regulatory frameworks that are essential components. Scaling gen AI requires a multi-layered tech stack comprising applications that allow for end-user workflow integration, machine-learning operations that can tailor, deploy, and maintain models; access to different large language models, and reconfigured infrastructure that can support gen AI. Once the scaffolding is built, you can hang use cases from it that span from the back office all the way to the front lines, as well as legal and compliance. In other words, the same infrastructure is used to analyze different types of documents. The question here for financial services firms is whether they are building locally optimized scaffolding where it is needed that can also be used horizontally across the organization where it makes sense. Getting value to the customer does not always require the latest state-of-the-art model. What’s important is staying up to date on the developments in the field to keep selecting the most appropriate tech solutions.
- Treat data as a corporate asset
Data is the new gold dust and needs to be elevated as a corporate asset and governed in ways that extract maximum value. This is not specific to gen AI but is relevant to any digital transformation. It may sound obvious, but we continue to find institutions that have not yet done this and are thus struggling to gain momentum. Data is either a great enabler or a great blocker, and those that have elevated it to a corporate-level asset tend to do better. Yet the hurdles to doing so remain considerable: in our recent roundtable with data-and-analytics banking executives, one in four said that the quality of unstructured data was one of their biggest data-related challenges to scaling gen AI, followed by security classification of new data sources, and data permissioning.
- AI is a people play
A key differentiator that separates the institutions that are doing best with gen AI and the rest is how well they address end-user adoption and change management. Indeed, it’s no exaggeration to say that gen AI is not just a tech play but a people play. Firms need to ask themselves if they are spending enough resources to drive adoption, including through robust change management, reskilling, and measuring impact. The issue cuts to the core of cultural changes that most banks are grappling with. If you are a 30-year relationship-manager veteran, you have long been heralded as a rainmaker; now, along comes gen AI which aims to make you more effective. Not surprisingly, the pushback against adoption is real. A survey we conducted suggests that fewer than one-third of organizations use AI in more than one function, a share largely unchanged since 2021. The risk here is that companies will fall into “pilot purgatory” and capture only a fraction of the technology’s true value. The same rigor is needed in change management as it is in implementation.
These are certainly exciting times, and we have been impressed by the rapid embrace of gen AI by some of the biggest players. AI capabilities continue to evolve at lightning speed: latest developments include agent-based workflows that automate more complex processes, including, for example, writing credit risk memos. Assumptions made today on use cases thus may not hold true tomorrow. Ultimately, time is the best test of any technology adoption and, one year in, there is still a lot about gen AI and its current usage that we have yet to discover, let alone imagine future uses. What’s important is to keep the momentum going, to broaden the use cases, and above all to see the technology for what it is: a tool of incredible power, but one that needs to be managed with care.
Carlo Giovine is a partner in McKinsey’s London office, and Larry Lerner is a partner in McKinsey’s Washington, DC, office.
Copyright © 2024 McKinsey & Company. All rights reserved.