Information Regarding AI , Especially Agentic AI

[rudraabhishek5]

Hi Everyone,
My name is Abhishek. I am a RiseUp graduate. I am looking for case scenario where Agentic AI was used along with ServiceNow..It can be a full fledged application or a micro-applications. I am at a learning state and don’t know much so i kindly request to give references or citation is simpler language..

The Following template can be followed
1) Application name
2) Use of that application
3) What AgenticAI is contributing
4) What is base or core concept or what is the underline Tech Framework working including Agentic RAG, AutoGen, LangGraph, Semantic etc
5) I am looking for open source references including white papers….

Any suggestion will be highly useful.

[rahulswami]

Hi Abhishek,

Before the examples, one clarification that will save you a lot of confusion later.

AutoGen, LangGraph and Semantic Kernel are open source frameworks used to build agents in Python, outside ServiceNow.

ServiceNow does not run on any of them. It has built its own agentic stack on the platform. So when you look at a ServiceNow use case, the framework question has a different answer than you are expecting. The underlying ideas are the same, only the names change. I have mapped them below so the connection is clear.

Here are two examples in your template.

EXAMPLE 1 (full application)

1) Application name
Now Assist for Financial Services Operations (FSO), specifically the payment dispute agents.

2) Use of that application
When a bank customer says a card or ACH transaction is wrong, a dispute case is raised. Today an analyst manually pulls transaction history, checks the rules, decides whether the customer gets money back, and writes the reply. It is high volume and heavily regulated.

3) What agentic AI is contributing
ServiceNow ships prebuilt agents such as the ACH Dispute Agent and the Friendly Fraud Agent. They gather the evidence, work out a likely resolution path and prepare the communication. The analyst still makes the final decision. In banking that human approval step is not optional, so this is a good example of what people call human in the loop.

4) Core concept and tech framework
Built with AI Agent Studio, which is the low code place where you define an agent, its instructions and its tools. AI Agent Orchestrator coordinates multiple agents when one task needs several of them. Now Assist Skill Kit is for building the smaller single action pieces. Tools reach out to core banking systems through IntegrationHub, REST APIs and MCP. Grounding on knowledge articles uses RAG.

Mapping to what you already know:
AI Agent Orchestrator plays the role LangGraph plays, deciding which agent runs next.
AI Agent Studio tool definitions play the role of tool calling in AutoGen.
MCP is the same MCP you have read about, standard way for an agent to reach an external system.
RAG is the same RAG.

EXAMPLE 2 (micro application and a better first build for you)

1) Application name
Resolution Notes Coach, a small custom skill on Incident.

2) Use of that application
Support engineers often write poor resolution notes. Six months later nobody can understand what was done. This reads the note before closure and suggests a better version.

3) What agentic AI is contributing
It reads the incident work notes and description, judges whether the resolution note is actually useful, and rewrites it. Small, but it is a genuine reasoning task rather than a rule.

4) Core concept
Built in Now Assist Skill Kit. You write a prompt, define input fields from the record, choose the model, and deploy it as a UI Action. No Python and no external framework.

One warning that catches everyone. In AI Agent Studio an agent runs either as a Dynamic User, meaning it inherits the roles of whoever triggered it, or as a dedicated AI User with roles you assign. If that identity does not have the right roles, your agent returns nothing and no error tells you why.

5) Open source references and papers

This is the part I think will interest you most. ServiceNow Research has published real open source work on AI agents, and it is free to read and run.

WorkArena. A benchmark that tests how well AI agents can do normal knowledge work tasks on an actual ServiceNow instance. Paper at arxiv.org/abs/2403.07718, code at github.com/ServiceNow/WorkArena. Best part for you: the tasks run on a free Personal Developer Instance, so you can try it yourself.

WorkArena++. The harder follow up, focused on planning and multi step reasoning. NeurIPS 2024.

BrowserGym. The environment those agents run inside. github.com/ServiceNow/BrowserGym and the ecosystem paper at arxiv.org/abs/2412.05467.

AgentLab. A framework for building and testing your own web agents on those benchmarks. github.com/ServiceNow/AgentLab.

If you read one thing, read the WorkArena paper. It shows honestly how far agents still are from doing our jobs, and it will give you a much more grounded view than most marketing content.

My suggestion for a starting point: build one Skill Kit skill on a PDI first, something small like the resolution note example. Once you have seen a skill work end to end, agents make far more sense.

Happy to help if you get stuck.