Getting Smart About Intelligence

Artificial Intelligence (AI) has been building in capability, acceptance, and adoption over the past few years, and that trend has rapidly increased with the emergence of Generative AI (GenAI). ServiceNow has been building up AI-enabled capabilities on the platform for a while now and made it a key topic for Knowledge 24 soon after the launch of GenAI capabilities on the platform. 

With so much attention given to AI and all the different terms associated with the technology, understanding exactly what AI capabilities ServiceNow has to offer can be a challenge. Here, we look at the current state of AI offerings on the ServiceNow platform to gain a better understanding of the tools at our disposal. 

Intelligence on the ServiceNow Platform

AI, ML, PI, GenAI–these are all terms and acronyms you most likely have heard before. Are they all the same thing? Not quite. 

Artificial Intelligence (AI) is the umbrella term for all of the other capabilities. Machine Learning (ML) is a type of AI that analyzes data to find patterns and derives models that can be applied to assist with decision making. Predictive Intelligence (PI) is a subset of Machine Learning that takes data (typically structured or semi-structured) as an input and uses it to predict an output. 

Generative AI (GenAI) is an expansion of Machine Learning that adds to models the ability to create new data based on existing data given an input. This is a significant difference from PI models, where the possible outputs or range of outputs are known in advance. For example, a PI model can predict what product a person is going to buy based on past purchases, the time it will take for a Case to be resolved by a customer service agent, or a relevant Knowledge Article based on a search query or data from an Incident.  

GenAI can go further to do things like build a tailored product description highlighting the features that are most important to a buyer, recommend steps that can be taken to resolve a Case by extracting troubleshooting and resolution steps from similar cases, or presenting a unique response based on content from one or more relevant Knowledge Articles. Predictive Intelligence will display the product or the article, whereas GenAI will synthesize more targeted content based on those products and articles. This is a subtle but powerful distinction. 

So what capabilities does ServiceNow offer? The ServiceNow Documentation lists all AI-enabled capabilities and, as of May 2024, even includes AI capabilities, with a Beta launch of summarization and code explanation on nearly every page. For now, we focus on two of the most widely used capabilities: Predictive Intelligence and Now Assist, ServiceNow’s GenAI solution. 

The main similarity between the AI-enabled capabilities used on the platform is to help replace low-value, repetitive activities and to surface the right information in front of users in order to accelerate decision making, complete work faster, and become overall more accurate and efficient.

Code Generation Example - Now Assist

Now Assist for Code Generation

Predictive Intelligence

As of the Washington DC release, ServiceNow includes three predictive frameworks for creating, training, and using PI models on the platform: classification, similarity, and clustering. 

The Classification framework can predict a field value based on structured and semi-structured data. The most common example, and one of the earliest used capabilities, is predicting the category of a record based on inputs like the short description. It can be used to set nearly any choice or reference value on a record, provided there is a set list of possible values and a strong correlation between the fields used to make the prediction; this makes automatic assignment a particularly valuable use case, considering that reassignment is one of the main causes of delay in Incident and Case resolution. 

The Similarity framework identifies other existing records that are similar to a given record. A common use case for this framework is to identify Knowledge Articles that relate to an Incident or a Case. Another use case is to find similar Case or Incident records that can be used to identify troubleshooting or resolution steps that may apply to the current Case or Incident. It can also be used to find a Change Request that caused or may have contributed to an Incident. These capabilities all help to accelerate issue resolution, minimize impact to users and customers, increase efficiency, and boost user and customer satisfaction. 

The Clustering capability identifies groups of records that are similar to one another. The key distinction from Similarity is that we do not start with a given record to make a comparison. Instead, we look at a population of records and divide them into groups of similar records. This is useful to help identify common issues, potential common causes, potential Major Incidents or Issues, or other shortfalls like knowledge gaps or single points of failure. The most common use case is Major Incident and Major Issue Management, where Clustering can identify the impact and help locate a common cause. It is also helpful for proactive Problem Management, to help identify and resolve common causes of minor issues before they have a larger impact. 

Although deprecated as of the Washington DC release, the Regression framework works like Classification but is used to predict a numeric value, such as time to resolve. Its limited utility is likely why it is no longer supported. 

Now that we understand the PI frameworks and a few use cases, keep in mind these considerations when implementing them. 

Most important is identifying the outcomes you want to achieve with these capabilities (as is the case when implementing any capability). Is there value in automatically assigning a category to your records? You can assess this by thinking about the amount of time spent on categorization, whether incorrect categorization causes any issues, and what other processes or decisions rely on categorization. Despite heavy attention on defining the right categories, most organizations (although they include category in reporting) have no actionable use for it. Consider instead whether automatically assigning a record to the right group based on a predictive model rather than static assignment rules may have better outcomes. 

Data quality is almost as important as outcomes. Predictive models need to be trained to produce good, usable solutions—and they need to be trained on your data. The quality of the predictions relies solely on the quality and accuracy of the data used to train the models. You can’t make good predictions with bad data. Do you currently have processes in place to ensure the correct data for your models (both the input fields used to make the prediction and the output field you want to predict)? If not, then focus on improving data quality first in order to be successful with Predictive Intelligence. 

You must also have sufficient quantity of data to train the models. The recommended number of records is at least 30,000, with a maximum size of 300,000. More records are not always better, as the data you train must be accurate to what you want to predict. However, the more quality data you provide to the models, the better the predictions will be. 

Another key consideration is to manage expectations. When training a predictive solution, there will be a tradeoff between the coverage of the solution (i.e. the expected percentage of records that will be able to produce a prediction) and accuracy (i.e. the percentage of predicted values deemed correct). No solution will be able to predict a value for 100% of all records with 100% accuracy, so you need to decide which is more important and tune your models accordingly. 

Finally, ensure you have a plan in place to continually train your solutions. As you bring in new data, it is critical to continue training the solutions so that they remain relevant to the issues you are addressing. A solution that had good accuracy and coverage 12 months ago may no longer be able to make good predictions today due to different applications being deployed, changes in groups or organizational structure, migration to the cloud, or other factors. Be sure that you allocate adequate time and resources to continually retrain the models and review their performance. 

You can find more information on training, tuning, and maintaining predictive solutions in the ServiceNow Documentation.

Generative AI

There is a lot of excitement for and attention to GenAI right now, but also a lot of skepticism and caution. Workers in a knowledge economy are rightly concerned that some or all of their job functions may be replaced by AI. Current AI tools do not yet rise to that level; instead, they stand to replace tedious and repetitive tasks and help work get done faster. 

ServiceNow clearly recognizes the best possible use for these tools, a fact made clear by naming the product Now Assist (rather than Now Replace). The current GenAI capabilities (referred to as “Skills” by ServiceNow when also delivered as a use case) included in Now Assist generally fall into two broad categories: summarization and content creation. New capabilities are being introduced quite rapidly, so be sure to keep tabs on the ServiceNow Now Assist Documentation for updates. 

Summarization skills are useful for quickly getting up to speed on an issue or other type of ticket when a handoff occurs, whether from another agent or from an automated process like an Alert or the Virtual Agent. It also includes use cases like generating Resolution Notes for an Incident or Work Order based on activity from the record history. This can save huge amounts of time as users no longer have to manually review the activity, highlight key points, and distill that information; the summary is available for you to pick up work or to glean insights and draw conclusions. 

Content generation skills cover such use cases as drafting Knowledge Articles as well as code, Playbook, and Flow generation. In these cases, the system presents a recommended artifact for the user to then adjust and use if they wish. Often getting to this starting point saves crucial time, and these items can then be knitted together to provide the desired outcomes. 

There are also capabilities that combine both summarization and content generation. Now Assist for AI Search provides actionable answers to search queries that are generated or selected from the content of relevant search results. In other words, Now Assist can present a uniquely generated answer or can surface an existing answer that it deems the best match. Now Assist for the Virtual Agent simplifies topic discovery by eliminating the time and effort needed to set up and train NLU models or refine keywords and by using AI search capabilities to provide more meaningful and actionable results during the chat conversation. This results in a faster time to value for search and Virtual Agent through easier setup and more relevant results. 

As with PI models, it is important to consider the outcomes you want from GenAI solutions. Fortunately, Now Assist provides more targeted use cases so that outcomes are clear from the start when you select which components you want to implement. 

Even with Generative AI, the quality of data that the system uses to generate a response is critical to the utility of the response. Detailed notes must be kept on tasks and other activities that are meant to be summarized. For example, if you do not record actions taken or activities that occurred in the work notes for an Incident or a Case, and you have not used the Additional Comments to communicate, then there will not be any content for Now Assist to summarize. 

The output of GenAI solutions is a starting point and not a final product. AI is still developing rapidly, and as such, effective governance must be put in place to ensure that the right solutions are put in place and producing results, that their performance meets realistic expectations, and that the outputs are used in the right way. We cannot blame the AI model if it generates code that we copy and paste into place but do not validate that it functions and performs the way we want.

Conclusion

The ideal use of AI is to replace low-value work in order to enable us to focus on the aspects of our jobs that matter most. GenAI capabilities are not meant to displace any of the PI capabilities but rather to complement them and provide a richer, more intelligent experience throughout the platform.

Armed with our new understanding, we are ready to define our outcomes and make smarter choices about how we use intelligence to achieve them.

Date Posted:

May 30, 2024

Share This:

Leave A Comment

Categories

Tags

Loading

Fresh Content
Direct to Your Inbox

Just add your email and hit subscribe to stay informed.