Key takeaways
- Despite significant investment in AI, only 8 percent of banks are able to apply predictive insights from their machine-learning (ML) models to inform campaigns.
- Although banks know that time to insight matters, just 16 percent have standard protocols for algorithm development.
- By codifying, unifying, and centralizing key analytics and supporting processes, these organizations generate 5 to 15 percent higher revenue from their campaigns and launch them two-to-four times faster.
A large retail bank invested in ML technology to take its personalization initiatives to the next level. The goal was greater predictive power and efficiency in designing and automating customer campaigns. But two years later, the bank was still managing its personalization program much as it always had: manually and in silos. Although it had acquired a sophisticated analytics engine, the bank had overlooked the elements needed to turn that engine into a smoothly functioning “brain.” The result was a perpetual cycle of subscale efforts.
This bank is not alone. We see many well-intentioned personalization programs stumble from a mix of inconsistent data gathering, sluggish production models, and bespoke algorithms that are hard to update and share.
There are organizational issues as well. Too many businesses continue to view personalization as either a marketing initiative or an analytics initiative, when it needs to be managed as a joint initiative across the business. In the bank’s case, the marketing team was tasked with managing the program, but marketers had only sporadic access to analytics resources, forcing them to fall back on basic heuristics that were easier to manage but less effective at personalization. The bank’s aspiration also proved narrow, with a heavy focus on boosting click rates and conversions rather than on long-term drivers of customer value. As a result, the institution struggled to meet its retention and satisfaction targets for key segments.
There is a way to crack these challenges, and it starts by reframing them. We’ve seen organizations across the banking sector test a model that puts customer value at the center of personalization efforts. The outputs are individualized, but the inputs and algorithms that produce them are codified, unified, and centralized. Companies that have employed this model have generated 5 to 15 percent higher revenues from their enhanced campaigns and have halved or quartered their time to market. The bank referenced above, for example, learned to shift from equating value with mortgages sold and accounts opened to measuring it based on customer outcomes, such as retention and willingness to recommend. Instead of sales messaging such as “Try our zero percent introductory rate,” they developed journey-based communications, such as “Here’s how to make your holiday stress free.” Instead of monthly or quarterly campaign releases, they shifted to a daily or weekly tempo. Other companies can do the same.
Would you like to learn about our Growth, Marketing & Sales Insights Practice?
Rather than creating an engine, most banks focus on parts
Often banks claim to already work in an agile way or to have the analytical tools in place to drive personalization. But without effective mechanisms to coordinate and amplify customer initiatives across the organization, many end up with one-off use cases, hard-to-replicate models, and limited knowledge sharing—all of which run counter to scale. We’ve observed the following six common challenges that banks across segments face/have faced (Exhibit 1):
- Sporadic and inconsistent customer data: Our research suggests that only about 28 percent of banks today have the ability to rapidly integrate internal structured customer data into their AI models.1 For many, a key struggle is that customer information is housed in different ways across the business and stove piped within functions, preventing the bank from gaining a complete view of the customer. To get around these constraints, some banks overindex on third-party analytics and unstructured data. But these can add cost and complexity while ignoring valuable customer data within a bank’s own systems.
- Narrow scope of machine-learning models: Only about 9 percent of banks have a full suite of ML models capable of driving personalized engagement at every touchpoint. Most are trained on isolated moments with short-term, product-driven aims, such as boosting mortgage applications or account openings, rather than on identifying drivers of customer lifetime value (CLV) and shaping their customer interactions based on those insights. As a result, they spend time, attention, and budget on interactions with only marginal returns and underspend in areas with higher potential.
- Subscale analytics development: Just 16 percent of data-science teams follow a standard protocol to develop AI tools. In the majority of cases, analytics are developed on a campaign basis, which slows build times and limits the number of personalization initiatives that can be launched.
- Poor campaign integration and tracking: ML models and campaign-management systems often lack feedback loops to connect them, with the result that a mere 8 percent of banks are able to apply predictive insights from their ML models to inform campaign execution and decision making.
- Inadequate AI risk management: Only 14 percent of banks have a specific AI governance framework. Instead, many AI risk-management practices are driven by traditional-model risk-management conventions such as those developed for know-your-customer and credit risk processes. But these conventions are poorly adapted to the iterative nature of the AI environment, which requires a more dynamic risk-management approach.
From hand crank to flywheel
At-scale personalization needs more than great analytics. It needs an integrated infrastructure with clear mission alignment around high-priority opportunities, use cases that cover the whole customer value lifecycle, an asset library equipped with ready-to-deploy case code, and a uniform set of practices to guide teaming and execution. We have outlined five core elements that make up this flywheel (Exhibit 2). Using this structure, initiatives flow across functions, forcing silos to collapse. Insights, tools, and practices are folded into playbooks that are deployed on successive campaigns, reducing launch times and continually improving outcomes. Equally important, the focus is on CLV rather than on purely transactional gains. Companies that have adopted this model have seen substantially improved customer outcomes.
- Identify high-value opportunities: Starting with one or two simple, high-impact journey use cases can help organizations roll out personalization initiatives faster while delivering value as they go. Those use cases should be determined based on customer and commercial impact and feasibility—taking into account the needs of high-value segments and microsegments, operational complexity, and the amount of new data or skills required. For example, a retail bank that was just starting out focused on the “deposits churn journey” within its affluent segment, then extended that to the “investments churn journey” for the affluent segment, and moved from there to the other segments. Once they were further up the maturity curve, the bank challenged its personalization teams to identify mass segment customers whose near-term value might be low but whose potential lifetime value was double or triple the average, making them good candidates to be upgraded to the affluent segment. Teams then developed campaigns targeting these new segments and developed relationship metrics to gauge whether the shift in segmentation strategies was effective.
The analytics factory within the personalization model
Understanding the factors that drive CLV requires massive amounts of historical data on multiple areas of customer behavior, from transaction histories to online footprints to demographic and psychographic characteristics. And it has to cover a sufficiently long period of time to enable pattern analysis. These needs alone explain why personalization programs cannot operate within discrete functions. Instead, organizations have to establish mechanisms to extract that data from their finance, marketing, channel, and product teams.
Even then, it can take months of trial and error to test thousands of different data combinations to arrive at a reliable algorithmic model. Once done, companies must figure out what actions they could take to drive CLV higher. Orchestrating these steps for a single major customer campaign can feel like a remarkable accomplishment, never mind doing it repeatedly and across all campaigns.
This is where the asset library comes into play. The library sits at the heart of a high-performing personalization model, centralizing, standardizing, and modularizing data and analytics and integrating decisioning and execution. It consists of four layers (exhibit).
The data layer takes inputs, known as data fields, from a company’s customer data and transforms them into a set of variables or features, such as a customer’s debt-to-asset ratios over a two-year timespan. These features are standardized with appropriate formatting, labeling, and categorization, and made available to serve as inputs to the machine-learning algorithms.
The analytics layer houses predictive intelligence. Algorithms (built with features from the data layer) test the strength and causality of different CLV factors and solve for customer questions, such as the risk of churn or propensity to buy. They are designed in a standardized, modular way and rigorously tested before being placed into the library to catch errors when they can be more easily addressed and enable greater sharing and reuse. Reuse, in turn, drives better campaign outcomes, since machine-learning analytics become “smarter” the more they are deployed in the field. For example, the retail bank mentioned earlier developed a sophisticated CLV forecast model in about a week by invoking routines that had been developed in their cross-selling, upselling, and value-reduction models. The model delivered better-than-expected accuracy in projecting a customer’s CLV growth because the algorithmic components had the benefit of prior rounds of field testing.
The decisioning layer tests which actions to take based on the predictive intelligence. But where heuristics and A/B testing may only be able to accommodate a handful of business rules, ML engines can test billions of parameters, enabling prescriptive interventions to be dialed in for a specific individual. The retail bank’s decision layer, for example, crunched through 50 billion parameters in different combinations, testing for the optimal channel, day, and time to send communications to specific high-value customers. The effort led to a 20 percent increase in the bank’s commission revenue across campaigns within the first eight months after implementation.
The execution-and-measurement layer takes these “decisions” and integrates them with relevant third-party tools to activate the campaign. Performance management is baked into this layer, creating a continuous information loop that feeds data from the field directly into the ML engine to refine subsequent analytics. Measurement and feedback requirements can vary significantly from one campaign to another, and even over time. So the modularity given by libraries is key to helping teams reuse previous components and speed delivery. After completing its first customer use case, for example, a bank had 1,500 features. While some elements were hyper-specific to the campaign, many others could be applied to new use cases—and because the algorithms underlying the features and data models are continually tested in the field, they became ever more predictive. Reusable models and interfaces can then be deployed to different parts of the business.
- Enable rapid activation: The asset library plays a vital role in helping personalization teams scale quickly (see sidebar, “The analytics factory within the personalization model”). The ready-to-deploy algorithms and automated decisioning and execution this structure enables make it easier and quicker to test new ideas, ingest data from the field, and issue ever-more-refined outputs.
In addition to the advanced machine-learning “brain,” rich customer-data set, and marketing automation tools, organizations need a robust set of performance metrics. An Asian bank, for example, introduced a battery of AI protocols, processes and tooling that enabled them to cut implementation time by 50 percent for more than 150 analytics use cases.
Creating playbooks of best practices can also shorten the learning curve. What goes into these guides can vary, but at minimum they should include campaign catalogs that document effective lead management and other tactics, triggers and performance metrics that have proven particularly predictive, and agile rituals such as how to break large projects into smaller cycles. They should also include recommendations on the ideal team size and composition and protocols for autonomous decision making.
Organizations have to be willing to invest in building momentum around the activation effort and sustaining it over time. A European bank, for example, was committed to driving its personalization agenda forward aggressively. They brought in or trained several hundred data scientists to help with the effort. These experts huddled with counterparts from marketing and the product-and-channel organization to create granular profiles of their customers’ financial lives and isolate drivers of CLV. Initial use cases focused on adding rich media imagery to campaigns, with the selection and messaging tailored to customer profile data. Subsequent use cases sought to allow customers to stop and start an application with different devices. Over time, the bank developed more than 200 use cases, adding them to a common asset library and sharing playbooks with lessons learned. Since launching its factory model, the bank has improved conversion rates ninefold. - Invest in fit-for-purpose martech enablement: An integrated technology stack is essential to orchestrate the insights loop that connects enterprise data to the ML models, feeding the resulting inputs into campaigns, then running the field data back through the engine until desired customer outcomes are achieved. Figuring out the best way to build this stack can get complicated quickly. Many companies fall into the trap of “throwing technology at the problem,” but instead of an integrated tech stack, they end up with a bloated one that adds cost and complexity. Best-in-class organizations let value dictate their spend. They align martech resources around their highest-priority use cases and tease apart the data, design, decisioning, distribution, and measurement dimensions they’ll need to meet their customer goals. This approach creates a martech road map based on value capture, which is more effective and sustainable. Given the expertise required to navigate the often-fluid martech environment, some companies find it helpful to collaborate with specialists, particularly when first building out their stack.
- Commit to creating a truly agile operating model: Many organizations apply agile development in pockets, but only a few manage their personalization efforts in a way that cross-cuts through the whole business. To help make that leap, some banks find it helpful to build integrated teams comprised of multiple agile pods, each tasked with a specific use case. One company, for example, broke campaign-related teams into “customer pods” and “enabling pods.” The customer pods served as the ideation center, responsible for capturing CLV insights and generating new ideas to test. The enabling pods took those insights and turned them into reusable assets and approaches. Both employed rapid test-and-learn practices and flexible teaming—pulling diverse talent together into specific project teams, then expanding, collapsing, and regrouping them as projects reached different milestones.
- Invest in talent and capability building: Leading banks take a data-driven approach to building their teams. They identify the skills needed to support personalization at scale within their business and train and hire for specific acumen, be it in advanced analytics, digital banking, digital marketing, and other areas. Given the high demand for talent, they also invest in developing knowledge champions who can carry lessons from team to team and raise the overall skill level. This train-and-scale approach helped the bank referenced earlier roll out its personalization model to six divisions, ultimately delivering over $120 million of value.
Breaking through data-architecture gridlock to scale AI
Leadership must also be in alignment
Given the cross-cutting nature of the flywheel model, organizations will need to secure strong senior-leadership alignment and define a large enough scope for campaigns to have a material performance impact. Some have found it helpful to focus on an individual business line to start, then expand across the enterprise. One retail bank set up its personalization model within its personal-banking division. That domain was small enough to help teams manage the learning curve, but large enough to contain multiple customer segments and product categories around which to design customer outcomes and campaigns.
Senior leadership must also be open to leading the case for change. We recommend establishing a steering committee comprised of individuals from analytics, marketing, and major product divisions who together have joint accountability for delivering on agreed customer outcomes. Authorized delegates can manage the committee’s day-to-day responsibilities, but it’s important that these individuals have sufficient organizational clout and decision-making authority to ensure that needed changes take effect.
Tailoring to markets of one requires a bank to function as a collective rather than as a collection of functions. By focusing on the right building blocks and prioritizing integration and alignment, financial institutions can create the machinery to personalize at scale, gaining the breakthrough value that they have long sought.
