Predictive Intelligence (I will refer to it as "PI" in this article) is a key feature to take your organisation to the next level. It is a platform function that provides a layer of artificial intelligence that empowers features and capabilities across ServiceNow applications to provide better work experiences.
In this article i will try to cover some best practices and some pitfalls to avoid that can hopefully speed up your implementation.
Phased Approach
PI is a technology that can help you in several areas. Several frameworks exist in ServiceNow
- Classification : used to categorise and assign tasks, incidents, and cases at scale accurately.
- Clustering: Group similar records into clusters so you can address them collectively or identify patterns.
- Similarity: Surface similar tasks and content to predict new major issues and recommend critical actions.
A big bang approach to design and build models for the three areas at the same time is very challenging because lots of iterations are need to converge to the right model.
The implementation should be done step by step starting preferably by the classification framework (if incidents are not routing to the correct team it is not even worthy to start looking how to optimise the resolution by grouping similar since the assignment group is not even correct at the first place..). A pragmatic roadmap would be (but not the only way, can be adapted)
- Incident/Case Routing And/Or Categorisation
- Clustering
- Similarity
DATA management (Bad Data in >> Bad Data out)
Data is key in this features. PI is light on the process side since it is only routing, classifying or grouping records. It is relying on your current process IT, HR or CSM processes.
You can leverage the full Power of PI is when your data is CLEAN. And when the right set of data is used for the training.
It is important to feed the training engine with enough data without overwhelming it with all kinds of data.
Start to identify the challenging area you are facing and start training the model on it. for example if you are using classification framework to route incidents, you might need to exclude incidents created by monitoring tools since by design you know where to route them.
You might also want to start with one language and include other language a little later down the line.
Start by 100k -> 200k records. A higher volume would need lot of time to train and will slow you down in your first iterations.
You should note that the time is highly impacted by the number of possible values of the 'output' field i.e the field you are trying to predict.
in case the default training timeout (6 hours) is not sufficient, you can increase it -reasonably- with this system property: glide.platform_ml.training_timeout
Smooth implementation
Don't be afraid of starting Predictive Intelligence journey even if you have already spent time on building the Routing logic using other features such as Data Lookup. Predictive Intelligence can Co-Exit with your existing logic as shown in the picture below.
 |
When the PI model is trained, the prediction can be triggered programatically. The "Template" example is done using an onBefore business rules which means that the Logic which means that it runs before data lookup. Please check the precedence between data lookup, assignment, and business rules here. This means that the system will try to assign the group predicted by the PI and if not found it can still take the group defined in the lookup. You can even still configure the system to take the logic defined in the lookup in the transition phase if PI precision is not high enough. This can be done by setting the "Always Replace" property to true in the data lookup definition. Assuming the business rule doing the prediction is an "onBefore" business rule with order = 100, the picture on the left would be the sequence used by the system to assign the group. |
Data model
When the PI plugin [com.glide.platform_ml] is activated, this below data model will be generated in your instance.
ML Solution Definition and ML Solution tables are respectively the configuration of your solution and the result of the trained solution. You can see that for each framework a set of tables is also generated to manage the results and some advances properties.
Photo taken from This NowLearning Training
Don't hesitate to share in this thread your experience with this module and your tips and tricks for a successful implementation.
