Setting Up and Using ServiceNow Machine Learning

Introduction

Machine Learning, the long-awaited functionality of the Now platform, has come out with the Kingston release. The new application in ServiceNow providing machine learning features is called Agent Intelligence. The application is now available on personal development instances; so to give ServiceNow experts and customers interested in the topic an insight, in this blog post we are going to take a look at setting up and using Agent Intelligence features.

The Main Features of Agent Intelligence

Agent Intelligence helps companies drive automation by using a machine learning solution that can learn and improve. The solution creates machine learning models that automatically populate selected fields of records; for example, they can categorise and assign incidents based on the short description with a continuously improving efficiency.

Agent intelligence is a customer-specific solution, which gives predictions based on each customer’s own set of data that is already available in ServiceNow, such as incidents, requests, problems, changes, or practically any other records.

The solution model behind Agent Intelligence is built on three elements:

  • A query from a requester to predict the value of certain fields.
  • A machine learning model prediction request to check trained solution models.
  • An intelligent update action based on the prediction coming from a trained solution model.

(picture source: ServiceNow Docs Site)

When, for example, a new incident record is submitted, a before insert business rule in the background calls the MLPredictor Script Include, which checks the available solution models for a prediction and then populates target fields automatically when a record is submitted.

The solution model’s inputs are the records that are already in the ServiceNow instance. The input fields of the model are used to make predictions; for example, the incident’s short description can be used to make a prediction for the assignment group, as an output field.

Agent Intelligence can be extended to any process in ServiceNow by creating custom solution models and training those models on existing customer data.

Setting Up Agent Intelligence

The activation of the Agent Intelligence plugin can be requested from ServiceNow’s HI customer system.

After activating the new plugin, a solution definition should be set up, that means, filters, inputs and outputs need to be defined. Filters will determine the set of data that will be used for training a predictive solution model. Inputs are the fields that will then be used as an input of a prediction request from a new record, and finally the output field is predicted based on the input fields and a related, pre-trained solution model.

After setting up a solution definition, it needs to be trained and the resulting solution model needs to be tested to see if the accuracy and coverage of the solution are good enough. After that, the solution model is ready to be used immediately.

Using Agent Intelligence

The usage of Agent Intelligence is automatic for the records determined in the solution definitions, so agents who work with Agent Intelligence only need to be aware that certain fields are expected to be predicted automatically.

Trained solution models have a version, meaning that solution definitions can be retrained any number of times, and only the latest solution model will be active, so the system does automatic versioning for the trained solution models.

Reports can be reviewed periodically, and if needed solution definitions, solution model precision and coverage levels can be adjusted continuously to fit expectations.

Good to Know about Agent Intelligence

  • As a rule of thumb, Agent Intelligence is recommended to be used only if there are over a hundred thousand records in a ServiceNow instance; but even then, the data quality is an important factor, so there can be a need for significantly more records to get predictions with high coverage and precision.
  • Agent Intelligence and the related reporting features consist of two separate plugins, so they need to be activated separately. When requesting the activation of the Agent Intelligence feature both plugins should be added to the request.
  • After activating the Agent Intelligence plugins, the trainer service’s system property needs to be set up.
  • There are a number of hidden application menus related to Agent Intelligence, which can be activated from Application Menus.
  • Unless there is a specific reason, resolved and closed records should normally be used in solution definitions to train a solution model because those are the most likely records to contain accurate information for the value of a field that is expected to be predicted as a result of the solution model training.
  • Training times are usually fast; a training for a basic solution normally runs for around 5 minutes per 100 thousand records, which will be dependent on the complexity of the solution model defined.
  • Solution models should be reviewed periodically to ensure that they are efficient and up-to-date. Re-training can be set to automatic, or it can be triggered on demand from the solution definition. It is recommended to re-train them once every few months or after major changes when there are enough new records in the instance.
  • There can only be one predicted output field per solution definition, but there can be any number of solution definitions and solution models set up for the same record types and same set of data.
  • In each solution model, the solution coverage and precision can be updated for each class that belongs to the solution model, which will then have an impact on the overall precision and coverage of the solution model.
  • When updating precision and coverage in a solution model, keep in mind that a higher level of precision will result in a lower level of coverage, and the other way around, so a higher level of coverage will result in a lower level precision.
  • Also keep in mind that Agent Intelligence does not always come back with a prediction, and if it does, the prediction is not always accurate. These two factors are dependent on the coverage and precision of each solution model on overall and the classes of those solution models in details. In the end, your predictions will be as good as the quality of the data is in your instance.
  • Agent Intelligence licensing information and prices can be requested from ServiceNow account managers; licensing is based on the number of predictions.

For further reading

Author: Zalan@esmAlliance

Updates

Blog