Each time we build a model for a customer, we use a feature correlation process so that only the most important features that contribute to renewal likelihood are included in the model.

It's possible for more features to be added or existing features to be removed from Hook's model. This can happen during implementation or for scheduled model rebuilds.

Contact the Hook administrator for your organisation to give feedback on any features you'd want included in a future remodel.

Machine learning models improve their performance when they are built using a large training set of data. Therefore we expect new data on renewals and churns to improve the accuracy of Hook’s model.

We update models for our customers each quarter. This ensures that Hook’s churn predictions remain at a best in-class level of accuracy.

Hook's engagement models typically achieve 75% or above accuracy in predicting customer churn*, which is above best-in-class for the CS industry.

The other 25% will be from external factors that data can't predict such as company strategy changes. This information is discovered during conversations with customers and therefore not included in the data we use to build Hook's model. Signals can be used to flag positive or negative notes on accounts that cannot be seen using data alone and can be configured to override Hook's Engagement Level.

Our data science team often performs analysis to compare Hook's accuracy to your organisation's previous health score when presenting new models to customers.

Subscription information is included in the model because it’s necessary for model performance. This is not always key information for a CSM but is important for building Hook's model. An example feature would be subscription ARR.

There are two other buckets of data we use for Hook's model: insightful data and actionable data. Insights are helpful for CSMs to understand a customer's health. For example, number of support tickets. Actionable data is information a CSM can take action on to improve the Engagement Level of an account. For example, increasing the number of active users.

We would not showcase suggested actions for a customer for subscription information, since this is not realistic for a CSM to change.

How are edge cases handled by the model? (e.g. CSMs being aware of upcoming M&A events for customers, core champion soon to leave the business, customer has emerging budget constraints etc)

If these edge cases are captured in company data, we can include them in Hook’s model. Otherwise, signals can be used to flag edge cases for each customer. In signal settings, these can be configured to override Hook’s engagement level.

Related Article: Hook's Engagement Level

Engagement Levels for customers update weekly. This can be seen in the Engagement timeline on the Overview page

Each key metric has a different level of importance in contributing to renewal likelihood. If a metric that is not an important indication of renewal likelihood hasn't changed for a customer, this will not significantly change the engagement score.

Hook's model will be unique to your organisation and makes renewal predictions using various different data sources. Product usage data informs the engagement score, but so does meetings, support and subscription data. One of the most important features of the model could be data within one of these other categories.

We only include data in Hook’s model from your company’s data. Unless this information is available in your company’s database and has been provided for implementation, our model will not include it for making churn predictions.

We encourage CSMs to add Signals for customers in these cases. Find out more about signals here.

Signals in Hook enable CSMs to indicate positive or negative flags for an account that cannot be seen using data alone. For example, a customer sharing in confidence that they're considering another software provider. Signals can be configured in signal settings to override Hook's Engagement Level. This will be reflected in the Overall level field, and the original level in the Engagement field in the Customers table.

A customer may have an Inactive Engagement Level if they have are either zero users or no users with activity. Scoring takes place weekly, so any updates to user activity will be reflected in the next engagement score for an account.

A customer will have a No Data Engagement Level if they did not receive a score in the most recent scoring. This could be for a couple of reasons:

The current subscription was created in your CRM after the account renewed, so we have not included the account in this scoring.

Yes, Suggested Actions deliberately leave those factors out, but the model itself does not. It helps to split the data into three kinds:

All three feed the engagement score. Suggested Actions, though, only ever draw from the actionable data, because those are the things a CSM can realistically move. You can run a play to grow active users, but you cannot change how long an account has been a customer, or what they pay, through a conversation.

So factors like tenure and price still genuinely affect the score. They just never show up as a suggested action, because there is nothing a CSM could do to change them.

Potential Renewal is not a prediction of what that specific account will renew at, and it is not tied to the account's risk score. It takes the account's current ARR and applies the typical change in spend seen from accounts at the same Engagement Level that went on to renew.

Because renewing accounts tend to hold flat or expand rather than shrink, that typical change usually sits at or slightly above current ARR, even at lower levels. So an at-risk account, or one whose ARR has been declining year on year, can still show a Potential Renewal above its current ARR. The figure reflects how accounts at its level typically renew, not the account's own recent trajectory.

It is best read as a benchmark for "what a renewal at this level usually looks like", rather than a forecast for that one account.

Two accounts have the same low metric (for example the same low active-user rate), but very different Engagement Levels. Why?

The Engagement Level comes from all of the model's inputs working together, not from any single metric.

One account can sit high despite a weak metric because other strong factors, such as a long tenure or a healthy subscription profile, outweigh it. Another account with the same weak metric can sit low because it has nothing offsetting it. So a metric that is clearly the thing holding one account back may carry much less weight for another.

The same input can matter a lot for one account and very little for another, depending on everything else the model sees for that account.

*Disclaimer:
​

The reported accuracy of approximately 75% reflects the model's performance during internal testing and may vary depending on specific use cases or data inputs. While we strive to improve model quality continuously, predictions generated by this system are not guaranteed. Users are encouraged to review outputs and apply human judgment where appropriate.