ServiceNow Agent Intelligence provides customers access to a patented machine learning engine that further categorizes, routes, and assigns work in order to automate processes based on the unique characteristics of each customer.
Why is it important?
As workplaces are getting more digitized and more IOT solutions are coming into picture, companies today are struggling to keep up with new work, which stifles their ability to innovate.
This influx of new work causes manual work to create thousands of incidents/cases. This task consumes a lot of time and resources, thereby introducing human errors and slowing resolutions. As a result, customer loses productivity and have a lower CSAT.
Agent Intelligence eliminates the bottleneck created by manually triaging tasks. Service desks struggles to keep pace with the burden of re-routing. Requests become more efficient with Agent Intelligence while requesters get satisfying and fast resolution of their requests.
Machine Learning Core Concepts
Machine Learning uses generic algorithms (learning algorithms ) that has the ability to tell something interesting about a set of data without you having to write any custom code specific to the problem.
Instead of writing code, one can feed data to the generic algorithm and it builds its own logic based on the data.
We have 2 types of Learning in Machine Learning
Supervised Learning involves an algorithm based on known input variables, and a known output value. This set of known input variables and output values is known as the training dataset.
The goal is to approximate the mapping between input variables and output value so well that when you have new input you can predict the output value.
Unsupervised learning is used in situations where you only have input data and no corresponding output value.
The goal for unsupervised learning is to model the underlying structure or distribution in the data in order to learn more about the data. These are called unsupervised learning because unlike supervised learning above there are no correct answers and there is no teacher. Algorithms are left to their own devises to discover and present the interesting structure in the data
Supervised Machine Learning – Classification and Regression
Supervised learning problems can be further grouped into classification and regression problems.
Classification: A classification problem is when the output variable is a category (or discrete value), such as “red” or “blue”, or “disease” and “no disease”, or Assignment Group “A”, “B” or C.
Regression: A regression problem is when the output variable is a real value (or continuous value), such as “dollars” or “weight"
So what kind of Machine Learning solution does Agent Intelligence provide?
Agent Intelligence provides a Classification Supervised Learning solution.
In practice, Agent Intelligence uses a training dataset to derive an algorithm that predicts the value of an Output Field based on a set of Input Fields.
Because it builds a Classification algorithm, the Output Field must be a field type that holds a discrete value (such as a Choice or Reference Field), and not a field the contains a continuous value, such as a Currency, or Duration field.
Agent Intelligence Architecture
The solution is comprised of two main components:
Training Service: Responsible for extracting the dataset from the customer’s instance, building up
a Machine Learning Solution, and then pushing this Solution back into the customer’s instance
Prediction API: Leverages the pre-built Solution to predict the value of an Output Field based on Input Fields.
Building a machine learning model is compute-intensive, the resources allocated to a customer instance are not suited to this kind of workload.
The Training Service is provided as a shared service to customer instances. It is responsible for extracting the raw dataset from the client’s instance, splitting the data set into a training dataset, and evaluation dataset, building up a Solution, evaluating the solution against the evaluation dataset, and then pushing the Solution and related statistics back into the instance. Once the Machine Learning Solution has been loaded back into the client’s instance, the prediction occurs entirely on the instance itself. Unlike other Machine Learning services available in the market, this entire process is carried out without the customer needing to build any integrations to export the data, or leverage the prediction engine.
Two important concepts related to training are the Solution Definition and Solution.
The Solution Definition is a configuration record that specifies how to train a predictive model. All Solution Definitions specify these values.
The records used to train the model. For example, training based only on incidents that have been resolved or closed within the last 6 months.
The input fields the model uses to make predictions. For example, use the incident short description to make a prediction.
The output field whose value the model predicts. For example, set the incident category based on the short description.
The frequency to retrain the model. For example, retrain the model every 30 days.
The Solution is produced by Training Service based on the Solution Definition. Agent Intelligence uses this object to predict a target field value given one or more input field values. All solutions specify these values.
The solution precision is the aggregate percentage of correct predictions. For example, a precision of 50 means that out of 100 predictions, half of them should have the correct value.
The solution coverage is the aggregate percentage of records that receive a prediction. For example, a coverage of 50 means half of all eligible records actually receive a prediction.
The solution classes are the output field values for which the model can make predictions. Each class is an output field value with a list of possible precision, coverage, and distribution metrics to choose from. For example, the Incident Categorization solution has a class for each category such as
software, inquiry, and database.
The class distribution is the percentage of records from the entire table that have this particular output field value. For example, a distribution of 50 for the inquiry class means that half of incidents have the inquiry category.
The Prediction API Leverages the pre-built Machine Learning Solution to predict the value of an Output Field based on Input Fields.
The result of calling the Prediction API is a predicted value and confidence score that provides insight into how likely the predicted value is correct.
There are two ways to interact with the prediction API:
Scripted API: The Machine Learning Predictor script include provides an API that can be used from a server-side Script. The most common use of the script include is to use it in a business rule to apply predictions to a record.
REST API: The REST API allows client-side logic (or external agents) to execute a prediction. The REST client must provide the ML Solution identifier, as well as the input parameters, and it receives back a predicted value and confidence score.
When the Incident solution is enabled, a business rule is automatically created (Default Incident Based Prediction). This Business Rule uses the Scripted API to iterate through the ML Solutions available for the target table and apply the predicted value.