AI in customer retention

As my previous articles have probably made clear, I’m a big believer in the potential of artificial intelligence (AI) and data analytics to help insurers foster customer loyalty with more pertinent, personalised services. That said, I believe it’s equally important to be transparent about technology’s limits, and the challenges insurers may face in achieving their AI ambitions. 

For all the touted success of AI, insurers can find it hard to realise the benefits. The reality is that extracting value from AI extends far beyond building or adopting a model; it requires us to integrate and operationalise those models in the course of everyday business. Clients may not struggle to build an AI model - but reliably delivering results day in, day out is much harder.

Many models implicitly assume that customers can be ’pestered’ with repeated interactions until they’re prompted to respond.  However, this approach may have dire consequences for customer experience. One obvious answer is to embargo repeated customer interactions on the same topic and move to running campaigns on a regular cycle, such as every six months, but this is far from ideal.

An alternative and superior customer-centric approach is made possible by models designed to understand customer motivations. To create interactions at the optimal time for the customer, models need to grasp customer attitudes or trigger points, and to target specific customer journey phases. Rather than running periodic campaigns at a time that suits the insurer, why not run a continuous campaign that directs content to individual customers at the optimal times for them?

Customer-centric and more traditional models may statistically yield the same theoretical performance. But models designed to nudge customers at the right time are better positioned to deliver exceptional business results, since each customer is more likely to respond favourably when the timing is more meaningful to them.

The prospect of better business results should encourage insurers to personalise customer experiences in an even deeper way. The number of models deployed can be increased, with each model designed to address specific business processes and customer touchpoints. Different models can deliver more granular insights into customer behaviour, such as propensity to buy or lapse, channel preference, renewal prediction, behavioural segmentation and responses to specific campaigns. And each model can provide even more insights to further personalise customer experiences and enhance outcomes – a powerful virtuous circle.

As I see it, there are five key elements to successfully deploy a multi-model AI approach to drive customer engagement:

1. Automating data processing from multiple source systems to provide an updated and holistic picture of each customer
2. Delivering timely scoring of individual customers with multiple AI models
3. Using multiple models to automate decision-making and determine the optimal personalised content and channel for each customer
4. Integrating AI and analytical outputs into downstream systems so they can seamlessly inform and enhance business users’ execution of campaigns
5. Observing the outcomes of customer engagement and integrating this feedback into future model predictions and customer content to drive continuous learning and improvement

I’ve seen many clients build AI models that deliver meaningful insights, but achieving these five goals is a more complex and involved undertaking. Given our work across the industry and the technological capabilities we’ve developed, many clients choose to partner with us to benefit from the best practices and platforms we’ve established. These can provide a foundation to extract value from analytics more efficiently, and with greater confidence.

Even as models deliver results, it’s important to remember they can be victims of their own success. The most common challenges are saturation effects and negative feedback, both of which can quickly render an initially well-performing model unproductive. Before deploying a model, consider how it will deliver sustainable outcomes over the longer term.

Saturation effects occur when previous campaign success reduces the ability of the model to identify suitable customers in the future. For example, customers who pay premiums via automatic payments often have lower lapse rates.  In an automatic payment campaign, a lapse model could be used to prioritise nudging customers towards automatic premium payments. Yet a successful campaign would mean that, over time, the model recommends fewer and fewer customers that are likely to lapse and eligible for this kind of outreach. Consequently, the campaign's return on investment reduces over time.

Inevitably these models will have to be retrained and may suffer from negative feedback. In the above example, the model learns new customer behaviour by identifying customers who lapse and don't respond to the automatic payment campaign, while customers that did respond and didn't lapse are no longer available for the model to learn from. The model is retrained with negative results and may therefore perform worse than the original model for the automatic payment campaign.

Such examples show how lasting business outcomes from AI depend on careful model design and consideration of long-term performance.

It’s only natural to chase - or to use a technical term, ‘overfit’ - the best-performing models, so they end up mirroring too closely to limited sets of training data. Such models can precisely replicate what happened in the past, but the results may not be generalised into the future. The further we move from the actual customers and time period that a model was trained on, the worse the model tends to perform.

There are many approaches to addressing this issue, but the three I view as most useful are:

• Using human knowledge of customers to design model features that accurately reflect how customers see the world and make decisions. For example, customers may be sensitive to premium rate increases, and a model feature that accounts for the change in insurance premiums from previous years can be more powerful than one that is ignorant of this human behaviour
• Training and testing the model using data from different customer groups, preferably over different time periods
• Embedding a control group or otherwise designing the model so that it is unaffected by the negative feedback loop described above.

Through this series of articles I’ve explored how advanced analytics can help shift customer experience from being transactional to personalised. More personalised experiences may deliver superior business results but also increases complexity. We need to develop a better understanding of customers, their attitudes and motivations. This requires shifting models designed around the insurer to those designed to reach customers at times that are meaningful for them.

We also need to be cognisant of the many challenges in designing sustainable models that enable ethical and thoughtfully curated customer experiences, and to acknowledge that to fully realise the value of personalised customer experiences, an entire suite of well-orchestrated analytical models may be required.

All this has to be based on careful planning and coordination, technological capacity and a certain appetite for innovation. Yet I believe insurers who join us on the journey can unlock the transformative potential of AI systems for their customer relationships, making those relationships, and ultimately their business models, more resilient.