A Perspective on AI’s Use in Service Delivery and in ServiceNow
Introduction
I guess it’s time to write something about AI. The topic seems unavoidable these days. Everywhere you turn there’s another article either breathlessly extolling the virtues of AI, or warning of the dangers, or predicting a future where AI either has given us lives of leisure or made us its slaves. What I don’t see is a lot of analyses of the short-term practical uses of AI, and where it may or may not make sense for business use. In this vein, I will give you my thoughts on how AI fits into a ServiceNow strategy for the next 1-5 years, and highlight what I think are both the realistic benefits and the unrealistic hype.
A Brief History (and Where We Are Now)
There has been some version of AI in the ServiceNow platform for years now. The early iterations were machine learning (ML) models that attempted to fill in “ticket” details based on what was learned from previous tickets. For example, an Incident comes in from an end-user with a short description of “Can’t access SAP expense module”. The ML uses its database of learning similar Incidents with keywords of access”, “SAP”, and “expense module” to auto-assign and prioritize the Incident automatically, bypassing initial human triage, saving time and human intervention.
The second phase of AI incorporated Large Language Models (LLM) into specific toolsets (currently branded Now Assist). These tools can take data from the system and build natural language outputs for a variety of uses. The most common example is using LLM to summarize a triage type “ticket” (Incident, Case, etc) into a resolution notes field, saving a human from having to read through the entire history and start writing a summary from scratch. They can also be used to write emails and code blocks.
The next phase is Agentic AI; there’s quite a bit of “press” about this today. Agentic AI is meant to mimic a human’s process / decision flow – analyzing a problem and taking steps to remediate the problem automatically using intelligence it has built up over time analyzing similar problems. It is meant to work largely if not wholly autonomously.
The AI journey to-date can be summarized thusly:
- Informational: The AI provides information – and hopefully answers – to human interactions based on learning from previous similar interactions and “grading” of the information provided by the AI.
- Problem Solving: The AI runs workflow automations based on input from the human to automatically provide solutions to the user’s needs. The AI bases its decision making on “grading” the success of previously suggested solutions.
- Automatic detection and remediation: The AI learns when issues are occurring and runs workflow automations based on learned knowledge of the issue to automatically resolve the issue.
Much of the AI available in ServiceNow today is still ML-based. You’ll notice this as you read about and set up the AI tools: there are minimum number-of-record thresholds that must be met in order for the “machine” to learn enough to be able to provide viable suggestions and solutions. Viable in the sense that the AI can determine it is statistically likely that the suggestion is accurate.
Recently, ServiceNow has added large language model capabilities to leverage “ChatGPT-like” functionality. This allows for auto-creation of chat and “ticket” summaries, notification text, and other customer communications. This is the next phase of AI in ServiceNow, currently branded as Now Assist.
As we venture into mid-2025 (and the annual Knowledge conference), it is likely we will be hearing quite a bit about ServiceNow’s plans for Agentic AI, as the promise aligns so well with the Service Management paradigm ServiceNow has excelled in.
Considerations for AI in Service Delivery
The marketing around these new AI solutions certainly make them sound like the answers to all your problems. While there is promise for automating facets of work that are tedious and human-intensive, I believe that using AI specifically for Service Delivery needs to be carefully considered. Let’s look at a tiered approach to using AI for a Service Delivery situation, then delve into the concerns and challenges.
Consider an example of the layered approaches to Service Request using AI:
- A chatbot that attempts to answer user questions based on knowledge gleaned from analyzing the available data in the system. This is the most basic level and functions much like ChatGPT does using the broader internet. Simple to turn on and implement, but many users don’t want this as a solution, any more than they want an automated phone tree when they call a company’s customer support number.
- Automatic ordering of a Catalog Item based on questions and prompts to the user. This expands on the solution above, but adds the intelligence to understand when a Catalog Item will provide the required solution, and uses prompts to gather the required information to begin the Service Request process without the user having to navigate to the Catalog and order it themselves.
- Automatic fulfillment of the process started in part 2. This would include the AI initiating an RPA or orchestration to complete the Service Request without human intervention.
The layers go from initial “call” deflection with targeted knowledge, to full end-to-end delivery with AI making all the determinations for the correct process steps along the way.
Concerns/Challenges
What are the concerns for letting AI take over the end-to-end delivery? Typically they are allowing AI to make all the decisions along the way. Regarding the full process, you can ask yourself:
- Has a human determined the process can be automated?
- Has a human built the workflow that the AI can initiate?
- Can the Request be fulfilled without human approval or other fulfillment work done by a human?
If you say that a human does not need to make these determinations, there are potential organizational challenges that arise. They center around what KPIs a company uses, and how they rank the importance of these KPIs. In a Service Management environment, consider:
- Customer satisfaction risk. Regardless of the AI solution, is a company willing to risk customer satisfaction, and potentially customer retention, by turning customer service over to a non-human entity? In the scenarios above, what if the AI makes an incorrect determination and there is no human error-proofing along the way? What if AI misinterprets a password reset request as a system reset request and reinitializes a server VM?
Most of us in the business world live in risk-averse environments. Leadership and management are more likely to preserve the status quo than take large risks against customer satisfaction and retention.
There’s a further challenge I don’t hear about much: AI works best with large volumes of data to analyze. The bigger the better. In an ideal world, ServiceNow’s AI would be analyzing data across all customer instances, not just one customer. But companies are not likely to sign up to share their data. Additionally, there is some data even within a single customer environment that will not be made available to AI: Think sensitive HR data. It certainly is a different scenario than say a Google AI agent scanning the entire internet for useful data.
Reality Check
What I see for most ServiceNow customers is a “tiptoe into the AI water” approach, rather than wholesale adoption. The approach often incorporates these aspects:
- Turning on an AI tool in a subproduction environment and observing the results
- Using LLM to auto-create text passages for notes and emails that get reviewed and edited by a human
- Letting AI make a suggestion for a resolution but having a human execute the action, or at least approve of the action prior to AI executing them
- Building orchestration flows that AI (the system) will execute, but a human designs, or co-designs with AI
Smart companies use their best employees to spearhead these initiatives, running in lab-like environments before turning on in production, and becoming champions of the initiative throughout the organization to drive adoption.
Conclusion
There’s great promise in the features and functionality of AI. Being able to use machine learning and large language modeling to remediate issues and fulfill requests automatically, freeing up precious human resource time and capital, is appealing to corporate executives in all industries. As with most new technologies, the theory and the marketing pitch run far ahead of the reality and challenges of implementation and management of the technology. Given the scope of AI for automating work, much consideration must be given as to when, where and how to implement. Smart organizations have knowledge of themselves and their processes, and plan for an intelligent, phased approach to AI implementation, aligned with their KPI objectives. These same organizations measure against these KPIs and continually tune along the way. The alternative risks both customer satisfaction and retention; eliminate these risks as you go on your AI journey!
