Can AI manage my customers better?

We were asked this question by an insurer. 

This was in the context of whether AI and Machine learning (ML) has advanced to the level that it can be dropped in, and work out for itself how and where to improve the management of existing insurance customers.

This is unfortunately more marketing hype than reality today.  In this post we will outline why we see it that way.

There are three reasons why this doesn’t work (yet) …

Reason 1 – Weak signal in the data today

In insurance, the current data contains patterns (a signal) that helps us target retention, servicing, cross-sell activities to the right customers. But this data signal is weak compared to many other industries like banking or mobile telco.  The reason is obvious – insurers have few interactions with our customers today and therefore less relevant data in their current environment.

Machine learning can nevertheless be very effective and we have proven this many times. But data scientists have to work hard to build ML models that are good enough. These models require a strong dose of human business knowledge, and human feature engineering to shape the data in an optimal way and get the most out of the signal that is there.

The promise of purely automated machine learning models or off-the-shelf AI that can be dropped in, does not deliver such good results today. It needs more relevant data, and models need to be tailored.

Reason 2 – Avoid the spam trap

Many of our re-engagement interventions with customers, require a member of staff or distributor to contact the customer. In many cases a human-to-human interaction is much more effective than a digital email or SMS interaction. The majority of customers (with perhaps the exception of motor or home insurance) do not interact digitally. 

With human-to-human actions, the bar for success needs to be high. Staff and distributors will tolerate only a few failed actions before they will give up and believe that the actions do not work.  We need to be aiming for at least 20-30% hit rate.

By way of illustration, one of our insurance clients had created ML models that were not quite up to the task. They pushed the actions out to the team via their CRM.  The call centre staff we spoke to explained that they ‘ignored that bit of the screen because the actions did not work’. The actions probably did work, but not often enough, and therefore most staff had already given up before they had a success. The campaign was dead before it started.

Digital channels appear to have a lower success bar.  An email or SMS campaign might have a 5% success rate if it is doing very well.  Digital campaigns have little direct cost, and systems do not lose motivation if they have a low hit rate.  But there be dragons here also. Customers are clearly human, and failed digital engagement attempts can switch off customers and train them to ignore future interventions. 

Spamming customers through any channel is probably a losing game. To avoid this, we need the best models and campaigns we can achieve.

Reason 3 – Codify business knowledge

There’s a lot of traditional insurance knowledge, actuarial modelling, and analytics required to create effective retention campaigns.  It is not all about predictive models and machine learning.

For example on premium retention, it is often important to know whether a customer needs to be underwritten again or can proceed without health evidence.  This helps select the right customers. These rules are delivered by creating a small lump of code written together with the business experts that really understand the products.  That’s the quickest and most accurate way to codify this knowledge into models.

In theory an AI could learn this in time but the humans would have to spend significant time and money training it, creating sample datasets, and so on.  Better, we think, is to write this as a piece of human code and plug it into the models. Then in the future, the AI or automation engines can be told to apply these rules.  

Combining AI and HI (human intelligence)

The bottom line is that we need to be creating the most effective campaigns we can today in the first place.  And this is best done using a combination of human knowledge and machine learning. Fully automated AI and ML cannot achieve the same results today.

Creating the most effective campaigns possible creates a virtuous circle: 

  • Staff and distributors want to do more actions, because the actions work
  • More actions means we capture more data and relevant feedback data
  • The feedback data is itself very powerful – it helps improve predictions, iterate campaigns and helps us improve models and campaigns for the next round 
  • We push out new, even better actions next time …

Automated AI and ML is developing fast, and in time we will have built enough further data to enable it to be more effective.  But that will take some time.


Get in touch with Matt