AI Agents FAQ & Troubleshooting
Last updated September, 2025, for Yokohama Patch 6, and Zurich Patch 1
*Safe harbor applies for forward-looking statements*
AI agents are autonomous systems that interact with their environment to gather data, make decisions, and complete tasks that would otherwise need to be done by a human.
Customers must have Now Assist installed with a Pro Plus/Enterprise entitlement. They must be on Yokohama Patch 1+ or Xanadu Patch 7+, with the Now Assist for ITSM, HRSD..., and AI Agents store apps installed. Admins must have the sn_aia.admin role. AI Search and Now Assist Panel must also be enabled.
An agentic workflow is the overall business problem or goal you’re trying to solve. Think of it as the why—the reason you’re deploying AI agents. An agent is the who—the virtual worker that performs specific tasks to help achieve the use case. Tools are the operations run to perform actions. They are usable by AI Agents to perform their tasks and achieve the use case.
AI Agents uses OpenAI's GPT-4o for its orchestration layer on ServiceNow managed Azure servers. AI Agent tools may use Now LLM. The context window for the AI agent is 128K tokens (exceeding this may lead to unpredictable behavior). The orchestrator LLM can also be configured to use other models such as Now LLM, Claude, and Gemini, as of Yokohama Patch 6.
The max length for an AI Agent Role: Max Length = 2,000. For AI Agent Instructions: Max Length = 8,000
- My instance is located in an Asia Pacific data center and I'm receiving a "Global Routing Option" pop-up. What is that?
If your instance is hosted in our Asia Pacific region data center, GPT-4o is not yet supported in the data center. As a result, in order to use AI Agents, customers must agree to use Global Routing. This may send data outside the Asia Pacific geography, but still within the ServiceNow ecosystem.
AI Agents can connect to 3rd party LLMs and other systems (including agents) via the subflows/actions tools as an integration. More seamless communication between ServiceNow agents and external agents is on the roadmap.
Assist consumption pricing is based on number of actions that are run, not including the orchestrator or communicator agent actions. AI Agents are bucketed into "Small", "Medium", and "Large". See our assists overview for more information: ServiceNow Assists Overview
- Do skills and Virtual Agent topics run in agentic workflow still consume assists?
No. If a skill, Virtual Agent topic, or conversational catalog item is run in an agentic workflow as a tool, it does not consume assists. The agentic workflow run itself would still consume assists per the "Small", "Medium", "Large" guidelines. (Skills are not counted when run in Now Assist in Virtual Agent or Now Assist Panel either.)
No.
Humans are still involved when AI Agents are running. AI Agents introduce the concept of "human on the loop" systems where humans primarily monitor the AI's actions and only intervene when necessary, taking a more supervisory role rather than actively contributing to decisions. For example, human service agents can review an AI Agent-generated plan and provide feedback or explicit approval before execution or changes. We also provide prompt injection protection via Now Assist Guardian. Our grounded prompt templates tie the prompts to platform data and increase the accuracy of the response, as well as RAG. Customers can use the same in Now Assist Skill Kit for Skills as tools; and set the temperature to ensure the output is more grounded. Note that with any generative AI product, hallucinations are a possibility.
AI agent builders can create triggers when creating an AI Agent or Agentic workflow that calls the agent when a condition or objective is observed. Agents can also be trigger via database changes or a user query.
During an AI agent interaction, the AI agent may ask the human user for feedback and/or permission to proceed. The AI Agent then may change its approach depending on the feedback.
Yes, it can.
AI Agents is supported for English only in the Yokohama Patch 1 release. AI Agents may recognize and work with other languages, but your mileage may vary and support is not guaranteed. Official multi-lingual support is on the roadmap.
As of July 2025, AI Agents is supported for GCC/self-hosted environments.
Yes. The sys_domain field exists on all agent tables and will get the sys_domain_path if the domain separation is enabled in the instance.
Data in an agentic worfklow conversation is stored in short term memory and can be passed between tools via the Agent's instructions. It can also be passed between AI Agents via instructions. Note that the data is being transferred by the LLM and may be changed. To minimize this, in the instructions, tell the Agents to be exact in passing the data. Also in the Tools UI, change the LLM strategy to ‘None’.
Workflows such as ITSM, HRSD, and CRM are each shipping out-of-box agents to provide out-of-box content and experiences for customers. You can find the list of OOB AI Agents here: Link. More out-of-box agents will be shipped in future releases.
This is usually due to the Agent proficiency not being filled out or inaccurate. There could be several reasons: AI Search is not enabled, the token count for the orchestrator has been exceeded, the agent prompt(s) is not descriptive enough. Also check that the AI Agents and tools such as Skills are all Active.
- AI Agent states that it can no longer execute a tool
Check the property value for “sn_aia.continuous_tool_execution_limit”. This property controls the maximum limit for continuous, uninterrupted executions for the same tool.
You can trigger an agentic workflow to appear in NAVA (create an INC, then agentic workflow will run in NAVA), but you can't discover an agentic workflow in NAVA (if you type "help resolve INCxxx" in NAVA, nothing will happen.) Yes, Now Assist in Virtual Agent can run AI Agents. Be sure the AI Agents skill is activated in the “Assistants > Skills” setup page, and that the Virtual Agent experience is toggled on for the AI Agent. Also make sure that your AI Agent description and role is detailed so that it can match the user query. The model provider for AI Agents and Now Assist in VA is recommended to be something over than Now LLM.
- The Now Assist Panel times out after 2 hours
Check the value for “Conversation Idle Timeout” for the Now Assist Panel (sys_cs_channel) record. Also check the property value for “com.glide.cs.conversation_idle_timeout”. Note that this property is Global and not specific to the Now Assist AI Agents, but rather to Virtual Agent as a whole.
- How do I ensure that my AI Agents are performant?
If you are building agents that need to analyze a lot of data, use Skill Kit to build a skill that the agent can use to speed up the agent and limit it’s LLM calls. The more tools and agents contained within a use case, orchestration performance may differ (we don't recommend >15).
- My AI Agent isn't consistent and does different things each time.
AI Agents and generative AI are non-deterministic systems. Running the same thing twice might give different results due to the LLM and instructions.
- How do I learn more about ServiceNow AI Agents?
View our documentation at: Link. A Now Learning course on AI Agents is available here: Now Learning
