Problem with automated Einstein Case Classification

Question

We have a custom Case field Cause, which is a picklist. This field is a good candidate for Einstein Case Classification. We have created and trained a model for several months. The suggestions seem OK, so now we want to automate the predictions. But this is what we see in the Case History:

After creation, Einstein (the dedicated user that is associated with the automated prediction) switches between empty and a prediction. This is annoying and we don't understand why it happens three times.

A few minutes later, a human user opens the case, sees no value for Cause and picks one herself (usually the value suggested by Einstein Case Classification).

Combined, this does not make any sense: we start with an empty Cause, it is being changed to a value, back to empty, back to value, but in the end it is empty.

Questions:

1. What could cause the repeated assignments by the Einstein process?
2. How to explain that - at the end - the field is still empty?

We notice that for cases that have been created by human users it seems to go well. But cases that are the result of email-to-case, exhibit the behavior described above.

We have some case related flows and Apex triggers. Several use the value of Cause, but none of them change it or initialize it. Also: Subject and Description, the fields used by Einstein Case Classification to predict Cause, are not changed by flows or Apex triggers.

The debug logs do not show the actions of Einstein Case Classification.

Answer 1

After creating a sandbox with all relevant data and configurations, I was finally able to solve this.

I deactivated all flows, the single Case Apex trigger and all Case validations. The combination of Email-to-case and Einstein Case Classification worked perfectly.

By gradually activating flows, trigger and Case validations, I was able to locate the culprit and - much to my surprise - it turned out to be a validation!

The validation is supposed to ensure that at least one of two fields is filled, but not for new cases created via Email-to-case (those creations should always succeed).

When Einstein Case Classification starts working on the case, it is no longer new, yet both of the fields that are covered by the validation, are empty. So Einstein Case Classification should fail, it should not be able to save the updated Case. What we see in the Case history is Einstein Case Classification "fighting" the validation and ultimately failing (of course).

The solution is to take the Einstein Case Classification process into account in the validation. And now it works.