nav[aria-label="Primary Navigation"] { padding: 0; & ul { list-style: none; width: 100%; display: flex; flex-direction: row; justify-content: start; align-items: start; gap: 30px; padding: 0; & li { margin: 0; } & ul li { list-style: none; } } }

Your CSMs are overriding the health score. How long before they stopped?

lachu536

Lachu here, new to GGR, joining from Bhutan, which I'm told makes me the furthest-flung member by some margin.

I build retention tooling, which means I spend most of my time looking at how health scores get constructed. But the thing I can't get a straight answer on is what happens after they're built.

Everyone benchmarks time-to-implementation. Six weeks, four months, whatever the vendor promised. Nobody seems to track the number that actually matters: how long until your CSMs stopped quietly ignoring the score and going with their gut instead.

I've seen teams a year into a platform where the score is technically live, technically accurate, and functionally decorative. The CSM opens it, disagrees, and works their own list. Which means the org paid for a system of record and got a dashboard nobody drives from.

My working theory is it comes down to explainability. A score that says "62" gets overridden. A score that says "62, because logins across the account dropped 40% and your champion changed roles three weeks ago" gets acted on, because now it's telling the CSM something they didn't already know, instead of grading something they thought they knew better.

So, genuinely asking the people who've lived it:

How long did it take before your team trusted the score? And what specifically closed the gap, better inputs, better explanations, or just time and a few cases where the score turned out to be right and the gut was wrong?

Happy to be told my theory is wrong. That'd be more useful than being told it's right.

Find more posts tagged with

Customer

Retention

Revenue

CS Team

Comments

Ed Powers

Involuntary, meaning a dunning problem? I guess the question becomes whether a failed payment increases the odds of customer churn in the particular case of a buyer of your new technology. If it does, verified by a simple 2x2 contingency table, Pearson's chi-square and Yule's Q, it would be good reason to test the variable in their health score model and see if it increases precision or recall. If it does, the ROI would be easy to quantify.

lachu536

Ed, this is exactly the kind of answer I was hoping for. Reading now.

The "watermelon / passion fruit" framing is a good one, and I think it sharpens my theory rather than replacing it. My instinct is that explainability and predictive accuracy solve different failure modes, a score can be genuinely accurate and still get ignored if the CSM can't see why, and it can be beautifully explained and still get ignored if it's wrong often enough that trust never builds. Sounds like your second link is the piece I'm missing on the accuracy side.

For context on where I'm coming from, I build Exeechain, a retention intelligence tool mostly focused on the involuntary side of this right now, flagging accounts lapsing on failed payments rather than trying to score everything at once. Partly because it's provable in days rather than months, partly because I don't yet trust myself to build the broader score well, which is exactly what your post is warning me about.

If you're ever curious to poke at it rather than just hear me describe it, happy to set you up with an account, no cost. Would genuinely value your eye on it given how deep you clearly are in this. No pressure either way, the conversation's useful regardless.

Going to read both properly before I say anything else useful.

Ed Powers

Hello Lachu,

This might be of interest: https://edpowers80525.substack.com/p/how-to-measure-health-score-accuracy?r=v8i0v. In my experience, CSMs don't trust health scores because they're terribly inaccurate, nowhere close to being useful. The key is to measure predictive performance, describe what the scores mean, and make CSMs aware of the limitations. Just like it is with a weather forecast, health scores offer no guarantees but inform people about a likely outcome based on the evidence at hand and the capability of the model. CSMs will always need to dig deeper to diagnose the situation and make their own decisions since not all data can be observed or collected, and every account is a little bit different.

Managers, however, must continually improve health score predictive performance in order to reduce the frequency of unpleasant surprises ("watermelons") and reduce wasted time ("passion fruits"). That process is described here: https://edpowers80525.substack.com/p/how-to-improve-customer-health-scores?r=v8i0v