TLDR: A lot of work has been done in recent years to develop customer health scores.  In B2B SaaS, even more work has been done to automate the process. Interestingly, the underlying approach that most companies use is flawed.  

We can draw a lot of lessons from sales forecasting in order to get health scores right.

Learning from sales forecasts

If a health score is meant to predict the future of an account (such as retention, growth or churn), then we can learn a lot from our sister function of Sales, where forecasting is used to predict the future of a sales lead or deal.

Sales forecasting is a well established practice.  A salesperson describes each sales opportunity according to its stage, its probability to close and likely close date.  They use guidelines provided to them, such as checklists that a deal must satisfy before it can move to a given stage. It’s kind of a dark art, that uses a mix of quantitative and qualitative measures.

But forecasting doesn’t end with what the salesperson submits. Their sales manager often overrides the probability and close date when submitting their regional forecast to their boss in turn.

Why do sales managers do this? Because they know every salesperson on their team is a bit different.  Some are optimists, some are sandbagging conservatives, and some are just plain inexperienced.

The VP of sales does the same thing for the same reasons, when rolling up each sales manager’s forecast into a consolidated forecast for the CEO and CFO.

By the time this is all done, 10-30% of the forecasted business from the salespeople has been removed when the forecast hits the CEO’s desk.

The most important takeaway from the sales forecasting process?  There are multiple forecasts, used by different levels of the organization for different purposes.

What’s wrong with health scores?

I’ve seen multiple problems with customer health scores.

The first problem is that Customer Success Managers are asked contribute their own evaluation as an input to the score. People are different, and the optimist and pessimist personalities manifest in the scores.

Compounding that problem is the attempt to use one score for multiple purposes

• The health score is used to prioritize the Customer Success Manager’s time
• The health score is used to forecast future revenue, including in the roll-up forecasts for the department

These mistakes lead to a rejection of the health score model because it can usually be disproven using a few outlier accounts as examples.

You need 2 health scores

If you were to adopt health scores the way sales has used forecasting, the following would be true:

• Customer Success Managers use health scores to guide the prioritization of which accounts need their attention right now. They get to use their own subjective inputs in addition to whatever quantitative data might help them better score their accounts.
• Departmental leaders use a different health score model based on quantitative data alone to produce a risk-based forecast of customer renewals, expansion and churn. Their model has different inputs and/or weightings, since they are forecasting revenue, not prioritizing CSM time

Is there a role for predictive analytics?

Predictive tools have taken hold in many sales organizations. What can we learn from sales that could apply to customer success?

Predictive technologies are being purchased to improve managers’ and VP’s ability to override the salespersons’ forecasts with a more accurate version of the same.

Said another way, VP’s buy predictive technology because they don’t want to miss a quarter and get in trouble with their boss. For now, they are less interested in predictive analytics as a way to teach reps how and where to spend their time deal by deal.

Predictive models for customer health are going to be wrong as long as we judge them according to multiple criteria.  If a model has to both accurately forecast customer revenue and prioritize CSM workload, it’s got no master.

We should also recognize that the data sets that feed the customer success health models are often too scarce.

Whereas sales teams have had success with predictive models, it’s because they had 100 leads for every closed customer. They were scoring the leads dataset to forecast the customer close.

Customer Success only has the customers, so the predictive model is working on 1/100th the size of the sales data set.  This smaller data set can mean a lot less accuracy in the model.

If you’re going to make a customer health score, start by giving it a single purpose. I recommend using it to develop risk-based revenue forecasts, where the Customer Success leader can use quantitative data to see relative differences in health scores across the entire customer base.