Preparing for Einstein Reply Recommendations

Understand the Data Requirements

To get started with Einstein Reply Recommendations, Einstein needs at least 1,000 closed chat transcripts in a given language, that contain four or more chat turns. A “turn” is a line in the chat conversation with terms or phrases that can help Einstein learn. Here’s an example of a chat with four turns.

• Customer: Hello. Can you tell me where life is sweet?
• Service Rep: I believe it’s on the sunny side of the street.
• Customer: Thanks! And what do I do with my worries?
• Service Rep: We recommend leaving them on your doorstep.

If you’ve enabled bots, Einstein begins counting conversation turns only when an agent joins the chat. Turns don’t include bot utterances.

If you don’t have enough closed chats, we let you know when you try to build your predictive model. You can also run a query to determine the number of chat transcript records. This query doesn’t detect duplicate chats, which aren’t counted toward your total number of cases for training. Chat data from 3rd party services isn’t supported.

While Einstein can generate a list of common replies with 1,000 transcripts, the more usable closed chats that you give Einstein to learn from, the better. To identify common replies, Einstein can analyze up to 200,000 chat turns from closed chats that have four or more turns, starting with your most recent chats. If you give Einstein 1,000 transcripts initially, when your org has 10,000 usable transcripts, Einstein rebuilds the model and performs a one-time refresh of your reply templates. We notify you when Einstein completes the refresh so that you can review and publish new and revised replies.

To build a predictive model, Einstein Reply Recommendations supports only unencrypted LiveChatTranscript fields.

Take Reply Recommendations for a Test Run

While you can set up Einstein Reply Recommendations in a sandbox org, migrating your chat data from production can be time-consuming. For this reason, we recommend turning it on in your production org and giving a small group of service reps access to reply recommendations. Choose knowledgeable service reps who could answer customer questions without the help of Einstein Reply Recommendations. Those service reps are able to boost accurate recommendations and let you know about any issues. When you’re comfortable with the way recommendations are working, open them up to your larger team.

If you decide to use reply recommendations in a sandbox, follow these guidelines.

• Match your sandbox licenses to your production licenses.
• To meet the data requirements and ensure realistic recommendations, make sure that your sandbox org’s chat transcripts are identical to your production org’s. Einstein Reply Recommendations examines chat transcript records. The body in a chat transcript contains four parts: the data and time that the chat started, the chat origin, a service rep name, and the chat conversation content. To analyze a chat transcript and identify language patterns, Einstein Reply Recommendations looks for “Chat Started:” in the chat body.