The next stage in AI to reach and retain customers

There’s rising awareness among both insurers and their customers of the potential of advanced analytics to enhance engagement, and ultimately sales. In fact recent research has shown customers identify artificial intelligence (AI) as a key driver of better experience. Yet it’s often less clear how to apply AI in a way that delivers on this promise.

In my previous article, I discussed the value of using AI to personalise customer interactions. In this second instalment of this three-part series, I’ll explore how insurers can get the best results from AI-powered tools, in terms of retaining customers and improving the quality of interactions.

Most insurers use AI primarily to identify the customers most likely to let their policies lapse. Single-purpose propensity models are highly effective when it comes to identifying a specific subset of customers at risk of being lost.

Applying such a targeted approach makes sense when customer interactions are relatively costly. However, if the cost of outreach is low and the subset of customers identified is a significant proportion of the total customer base then the impact of a propensity model becomes less meaningful. After all, why zero in on a subset of customers only to end up interacting with a majority of customers anyway? Conversely, why ignore most of your customers if communication is cheap?

Propensity models may also be of less relevance when it comes to responding to situations such as inbound inquiries. For all these reasons, it’s important that instead of relying on the results of a single solution, insurers leverage multiple AI models to realise a greater return on investment. 

Behavioural segmentation models have emerged as a particularly powerful tool. These divide customers according to behavioural patterns and formulate insights accordingly, as opposed to traditional marketing approaches that assign customer groups personas based on demographic characteristics.

In general, we have found that demographic-based approaches underperform behavioural models in terms of customer response rates. By analysing customer behaviours, behavioural models provide visibility into motivations, and allow insurers to deliver messages that speak to these directly.

Behavioural models can reveal stark differences within the customer base that may not be apparent along demographic lines. For one of our clients, customers in a behavioural segment indicating high activity had a 33 times greater likelihood to take action than customers in the low activity segment. Insights like these have clear implications for where, and when, outreach should be targeted, and the most relevant content to deliver.

Behavioural analysis can also highlight cases where customers take the same action, but for different reasons. For example, new homeowners considering a competitor’s offering and customers who are simply price-sensitive might both have high lapse propensity, but different motivations require distinct approaches, even though a propensity model might categorise both customer groups the same way. Interactions that recognise behavioural and motivational differences and are tailored accordingly tend to deliver superior results.

A different type of propensity model can make personalisation possible primarily by selecting the optimal message to send each customer from a ‘menu’ of prepared messages. Using this method for SMS renewal messages led to a 0.8% increase in retained premiums for one of our clients. These models can also incorporate reinforcement learning: with ongoing testing, the AI program can learn which content is most effective for each customer, as well as the ideal channels and times of day for interactions, to maximise their commercial impact.

That said, these models raise possible ethical issues which need to be factored into any responsible company’s strategy. Unlike with behavioural segmentation, it is not always clear why a propensity model chooses a particular message, and the difficulty of explaining results can raise questions. For this reason their usage needs to be monitored carefully.

Guiding principles for AI, from Swiss Re article:

Another important consideration is the frequency of communication. A ‘pestering’ approach that insists on touching base every day demonstrates a lack of understanding of customers, and a lack of respect for their time. In insurance, where ongoing relationships are vital, it can cause premium-paying customers to stop engaging, or even cancel their policies. These considerations need to be front of mind regardless of what a propensity model may recommend.

Propensity and behavioural segmentation models should ideally play complementary roles. For example, an insurer could use a propensity model to determine the optimal channels for different customer groups, and a behavioural segmentation model to optimise the content sent to each customer. This ensures coverage of all customers, maximises the return on investment in more expensive communication channels, and responsibly positions personalisation to deliver the best possible outcomes.

Another flaw of commonly deployed AI models is the embedded assumption that customers will either have a consistent propensity to act throughout the year, or only take action once a year (e.g. policy renewal). In contrast, we’ve observed that customer propensity to act frequently changes. Customers have many possible triggers, and it is important to understand what each means.

Another way to frame this is to look at the customer journey. Traditionally, insurers view this as a series of functional stages, such as onboarding, policy renewal and claims. However, the customer’s perspective is often very different: buying a new home, starting a family, or working towards financial goals. The right models can identify patterns in data and trigger events that reflect which of these life events customers are likely to be going through.

One  approach which has been shown to be successful in the past is to use models to understand what individual customers may do in the next three months. By applying behavioural models to analyse past patterns of behaviour for each customer, we can understand the most likely next action each customer may take. Studying the behaviour of similar customers following often complex patterns of trigger events can provide insights into where a customer on the brink of a life change will go next in their journey.

Journeys should therefore be on the minds of insurers looking to move away from the limitations of a single propensity model. Combining behavioural segmentation with propensity models  provides guidance on interactions with the right customers, at the right time, and with relevant content, to create consistent and relevant experiences across all interactions.

While personalisation has been proven to be effective, how AI models are deployed is important. Responsible use of AI is critical to consumer trust and, ultimately, the long-term viability of AI as a tool to improve customer experience. Our next article will discuss common challenges in developing and deploying these models, and what it will take to overcome them – a task that can’t be solved by data scientists alone.