Engage Digital | Intro to AI engine

The AI engine is an automatic and self-learning algorithm to categorize or ignore messages . The AI engine learns how agents manually categorize or ignore messages and uses that knowledge to predict the action an agent would take on incoming content, saving agents the trouble of doing it themselves. The AI engine allows you to control its precision, and trigger only when it is confident enough.

There are two types of AI engine algorithms:

AI routing: Determines the category of a message.
AI filtering: Determines whether a message should be ignored.

AI routing

AI routing determines the category to which a message thread belongs. When a new thread is created in the system, AI routing selects the most probable category that appears to be linked to the content using the model that it has built. If you have configured routing interactions based on skills, this will allow you to distribute this new content in real time to an agent who can process them, thus reducing unnecessary manual dispatching which is slow and expensive.

Note the following usage guidelines for AI routing:

AI routing works best if there are few categories with similar volume.
AI routing will not be accurate if there is too much disparity in the volume of each category. In this case, it will mostly categorize into the largest category (owing to probability rules).
AI routing analyzes only the first message of each thread. For this reason it is not intended to analyze brand messages on social networks.
AI routing analyzes only the message body, associated metadata (forum room for example), and source information to build its model.

Also note that the rules engine will not alter AI models. The AI engine builds its model only on contents that have been categorized manually or when an advisor replies to a message after recategorizing it or not recategorizing it. It is therefore important to correct wrong categorizations.

Building models

AI routing builds models to be able to categorize contents. Those models contain words or word combinations along with probabilities that those messages belong to such or such category.

Those probabilities are calculated based on a body made of X contents, which have been categorized manually by agents (therefore correctly categorized or validated), recategorized, or ignored (with or without recategorization). It is then crucial to correct the AI engine when it fails to categorize correctly so that it can learn from its mistakes. AI routing learns from 90% of the body and trains on the remaining 10%. Given that all messages already have a validated category, AI routing can determine whether it was right or wrong.

AI routing uses a maximum of 30,000 contents to build its model. The AI model is recalculated every night by adding the actions of the previous day (if the model status contains synchronize model). As a result, the AI engine continuously learns and changes its model over time, adapting to changing business contexts (new offers or new categories).

Learning mode

Over the first phase, AI routing or filtering must be launched in learning mode (status: synchronize model, no autocategorization) so that it can learn from agents’ actions to build its model. During this period, agents must manually categorize as many threads as possible to help the AI engine build its model.

AI filtering

AI filtering determines whether content should be automatically ignored. AI filtering is mainly useful for brand posts on social media. Indeed, on social channels with a considerable amount of noise, it will prevent your agents from spending too much time reading content that you generally do not have to reply to (such as community support, thank you messages, and conversations between users) and have them focus on content that needs to be replied to instead.

Unlike AI routing, AI filtering builds its model based on all non-agent messages that are not the first messages of the thread. AI filtering also allows you to ignore messages that contain only uppercase characters, very short messages, thank you messages, emoji-only messages, etc.

Minimum precision

For AI filtering, the minimum precision that you specify must be as high as possible. We recommend not to set it under 97%. For AI routing, an incorrect categorization is not too much of a problem because the message will still arrive in an agent’s inbox. However, if a message is mistakenly ignored, no agent will see the message and you will not be able to recover it.

The performance of the AI engine depends on the balance between coverage (the number

of categorizations) and precision (the rate of correct categorizations within this volume of content). If you choose a low minimal precision, the AI engine will categorize a greater amount of content at the expense of precision (global precision is lower because of the greater amount of prediction errors).

In contrast, if you choose a high minimum precision, fewer contents will be categorized but the AI engine will make fewer errors as well (global precision is higher because of the fewer number of prediction errors).

Iterating on model results

Once your model has completed its learning by analyzing the contents of at least ten messages, the results files are generated in the AI engine configuration page. These files include overall precision coverage, precision coverage by class, and detailed model training statistics. After reviewing these results, you can adjust the model parameters, or add keywords, to continue to iterate and improve the model.

Estado do sistema

Informações legais

Preferências de cookies

Aviso sobre cookies

Não vender nem partilhar os meus dados pessoais

Aviso de privacidade

Contacte-nos