7 common mistakes in predictive care and how to avoid them | NTT DATA

Thu, 28 February 2019

7 common mistakes in predictive care and how to avoid them

Predictive care is a relatively new area of predictive analytics. It typically promises to dig out and fix customer issues before they even happen. The ultimate customer experience is within reach. Would call centres even be needed anymore?

However, not all customers who contact the call centre have issues that can be detected. Some customers will always end up calling, regardless of any preventative actions. On the other hand, not every customer affected by an issue will make contact.

This is one of the first lessons of predictive care - predicting which affected customers will contact a call centre. Dealing first with those customers most likely to make contact is a clear way to reduce inbound contact and operational expenses (OPEX).

NTT DATA has looked deeply into the issue and discovered seven of the most common predictive care pitfalls and how to avoid them:

  1. Data models are all that matter in predictive care. - "FALSE" While the predictive model indicates which customers affected by an issue will make contact, this alone isn’t enough. Benefits accrue by acting to address these customer issues.
  2. Predictive care is all about avoiding or deflecting contact. - "FALSE" Predicting who will call because of an issue allows action to be taken to avoid these contacts in the first place or to divert callers to a low-cost channel, such as self-serve. However, some customers will always make contact. In these cases, predictive care helps organisations resolve issues first time while delivering a great customer experience
  3. The most frequently occurring contact reasons are always the priority - "FALSE" Some of the most common reasons for making contact include lost phones and changes of direct debits and these are difficult to predict. Instead, focus on those that can be predicted and where action can be taken to solve them. Contact reasons that are easy to predict and have automated solutions will deliver the biggest benefits
  4. Data scientists are the only resource you need to be successful with predictive care - "FALSE" For fully effective use cases, business owners need to work with data scientists to improve predictive models and show which actions produce the best results. Staff members need the expertise to analyse the output of predictions, recommend how to improve the model, and prioritise actions for customers with issues. Cross-functional teams are essential and business owners need to work alongside data scientists to maximise value from predictive care use cases
  5. I need sophisticated data and a “perfect” model before I act on predictions - "FALSE" Building and refining a predictive model requires time (and extra data) to improve predictions. However, a perfect model is not needed to start making predictions: start with the available data, train the model on historical data and, if the prediction is “good enough”, start testing on new data. Then, refine further. Also, don’t wait to integrate and automate your models before you start making predictions. Prove the value of use cases with simple models (sometimes manual) and when happy with the result, automate and integrate into your systems
  6. Accurate predictions always enable action - "FALSE" While prediction accuracy is important, it’s not the only metric. Organisations should also consider “precision” and “recall” before deciding which actions to take. If the model has low precision (<50%), there is a risk that taking action, such as sending a notification intended to avoid contacts, could “wake up” customers that were not planning to call in the first place. Beware of accuracy and always ask your data scientists about the precision and recall of a model – if precision is low, prepare properly and wait for customers to call you
  7. I must build generic models for predictions - "FALSE" Generic models are often too difficult to action. Knowing that a customer is affected by a billing issue is an example. You need to know the specific issue affecting customers to take the right action. Also, remember that customers with multiple issues are more likely to call. Build specific models for each of the contact reasons and “fuse” together the output of the models to increase overall precision and the ability to predict which customers are most likely to call.

Simon Driscoll
Head of Data Consulting

Michele Biron
Strategy Director

Read more about how predictive analytics can be applied in different areas of business

How can we help you

Get in touch