The best practice for using DistillerAI in your review is to review with Re-Rank enabled in the Level Settings until you have found 95% or more of includes, and then use AI Review to bulk exclude the remaining references based on what the highest remaining score is. This method ensures that the AI has actively learned enough to find the majority of your includes.
However, there is no "recommended" score to rely on. It will depend on the number of references you have reviewed (your "training data"). For example, if you review only 10 references and then check the scores with AI Preview and Rank, references with scores below 0.01 considered to be excludes wouldn't be as certain as if you had reviewed 1000 or 10,000 references. The larger the training data, the more likely the scores are to be accurate. It is up to your team's discretion to determine what is a sufficient amount of training data.
Finally, the scores assigned to references by DistillerAI are all relative to each other. A Project in which you're including and excluding 50% of references respectively will generate different scores than in one where you're including only 1% of references.