Implementation office hours Q&A - Now Assist for ITSM

[aastha14]

Listed below are the questions raised during our Now Assist for ITSM Q&A session with answers. Use this as a starting point for your implementation planning. 

Q. What is the difference between AI Agent and Now Assist? Which is more customizable for agent interactions?

Now Assist and AI Agents are distinct but complementary building blocks. Understanding both is essential before deciding what to build.

• Now Assist is ServiceNow's generative AI product layer. It delivers contextual, in-workflow AI experiences — summarization, reply recommendations, resolution notes, search-generated answers, and more. Custom capabilities are built using the Now Assist Skill Kit (NASK).
• AI Agents are autonomous digital workers configured in AI Agent Studio. They receive a goal, reason through a plan, execute tools (scripts, subflows, knowledge lookups, Now Assist skills), and adapt all with minimal human intervention.
• They work together: a Now Assist custom skill can be added as a tool inside an AI Agent, allowing you to reuse prompt-engineered skills within a larger agentic workflow. AI Agents do not simply 'execute' Now Assist skills — they are a separate orchestration layer that can optionally use skills as modular building blocks. An agent can operate entirely using script tools, subflow tools, and knowledge graph tools with no Now Assist skill involved at all. The skill becomes a tool only when you deliberately add it to the agent's toolkit.

NOTE: For customizing agent (fulfiller) interactions: Now Assist Skill Kit gives you fine-grained control over prompts and output formatting. For autonomous, multi-step task execution: AI Agent Studio is the right tool. For maximum flexibility, combine both.

Q. After deployment, do AI Agents need to be maintained and monitored by the deployment team, or can process owners adjust them?

Monitoring is required and both the deployment team and process owners play a role.

• ServiceNow provides built-in analytics within AI Agent Studio and the Now Assist Admin Console to track skill usage, assist consumption, and agent performance.
• Process owners with the appropriate admin role (sn_aia.admin) can review agent configurations, adjust instructions, and modify tool assignments.
• Platform best practice is to monitor usage trends before making significant changes to prompts, triggers, or tool configurations — but the timing should be driven by observed data, not a fixed calendar period.
• Governance of agent changes should follow your standard change management process, including lower-environment testing before production updates.

Q. What are best practices for modifying out-of-the-box skills or agentic workflows?

• Clone OOB skills and agentic workflows before customizing — never edit OOB records directly.
• Deactivating OOB records can break upstream dependencies; always work on copies.
• Test all changes in a sub-production environment before promoting to production.
• Avoid creating duplicate agents or skills that perform the same task — this creates maintenance overhead and unpredictable routing behavior.
• Use prompt engineering within the Skill Kit or AI Agent instructions to refine behavior; skills should be modular so they can be reused as tools across multiple agents.
• When adding a custom skill as a tool inside an AI Agent, use a descriptive name for input variables — agents use these names to understand how to populate inputs autonomously.

Q. Can Now Assist prioritize certain catalog items or topics based on frequent user requests?

Prioritization works differently depending on whether you are talking about search results or Virtual Agent topic routing. These are two separate mechanisms.

• For catalog item and knowledge ranking in search results: This is managed through AI Search. Configure Result Improvement Rules, search profile source weighting, and ML-learned click signals to influence how catalog items surface.
• For Virtual Agent topic prioritization: Topic routing is controlled within Virtual Agent Designer through NLU confidence thresholds and topic configuration — not through AI Search. A topic with a higher confidence match will be selected over a lower-confidence alternative.
• For Now Assist in Virtual Agent (NAVA): The system uses AI Search profiles to retrieve contextual content, so ensuring your AI Search profile is well-tuned with the right sources and relevance settings directly improves the quality of Now Assist responses.
• Escalation to a live agent remains available when AI results are insufficient — configurable within the VA topic design.

Q. Can AI Agents be tied to a Service Catalog workflow, e.g., to open a Teams message dialog with a customer?

Yes — but with an important caveat on AI Agent-specific Teams support

• Now Assist integration with Microsoft 365 Copilot is generally available, allowing users to initiate Now Assist self-service tasks (catalog requests, policy lookups) directly within Microsoft Teams.
• AI Agents can invoke ServiceNow subflows and Flow Designer actions that include Teams webhooks, notifications, or outbound integrations as part of their toolset.
• AI Agents in Microsoft Teams currently have limited native support. The Copilot for M365 integration does not yet support AI Agents directly. Validate your specific Teams + AI Agent use case against the current store app version before committing to this architecture in a production design.
• For Teams-based interactions today, the more reliable path is: Virtual Agent (or Now Assist in VA) handling the conversational layer, with AI Agents executing backend tasks triggered from that conversation.

Q. Can Virtual Agent with Now Assist search both internal KBs and the web to provide combined results like an L1 agent?

Now Assist in Virtual Agent can search both internal knowledge bases and configured enterprise external content sources — but this is not the same as open internet search.

• What 'external content' actually means: Specific enterprise systems (SharePoint Online, Confluence, other indexed sources) that have been connected to AI Search via pre-built connectors, the External Content REST API, or Flow Designer ingestion flows.
• What it does not include: Live web/internet search. Now Assist does not browse the internet or retrieve public web results like a search engine.
• To enable blended results from internal KBs + SharePoint/Confluence: configure AI Search profiles to include indexed external content sources, then ensure Now Assist in Virtual Agent is pointed at the same search profile.
• SharePoint federated search is supported from the Washington DC release onward; Confluence support followed in Xanadu. These are the primary supported external enterprise sources.

Setting up external content connectors requires additional configuration (schema tables, indexed sources, ingestion flows) and should be treated as an integration project, not a simple toggle.

Q. How can we monitor or restrict external content that Now Assist returns to end users?

ServiceNow provides a dedicated safeguarding feature called Now Assist Guardian.

• Now Assist Guardian monitors both prompts sent to LLMs and the responses generated in real time. It detects offensive content, prompt injection attempts, and sensitive topic areas.
• Administrators can configure Guardian in: Now Assist Admin Console > Settings > Now Assist Guardian.
• For each guardrail, you choose the response mode: Log Only (default) or Block and Log. Blocked responses are replaced with a standard message and the user is redirected to a fallback topic.
• Guardian applies at the instance level across ITSM, CSM, HRSD, and other Now Assist-enabled workflows.
• For PII/PHI protection in prompts and responses, the companion feature is Data Privacy for Now Assist — which masks sensitive data before it leaves your instance for LLM processing. Available from Yokohama onward.
• AI Control Tower (premium add-on) provides broader governance, lifecycle management, and risk visibility across your AI footprint.

Q. Why does AI Search return only internal KB data when Virtual Agent is enabled?

• AI Search results are determined by the Search Profile assigned to the application. If only internal KB sources are added to the active profile, only internal data will surface.
• To include additional sources (catalog items, external content, incidents, etc.), add the corresponding indexed sources to the Search Profile and republish.
• Ensure the Virtual Agent's Search Application Configuration is using the same Search Profile as your portal. Mismatched profiles between the portal and Virtual Agent are a common root cause of inconsistent results.
• From Washington Patch 3 onward, the Now Assist in Virtual Agent search profile can also be found and updated directly in the Now Assist Admin Console under 'Information Sources'.

Q. Can Virtual Agent mimic an L1 agent by providing troubleshooting steps before escalation?

Yes — and the most powerful way to achieve this today is through AI Agents in Virtual Agent, not just AI Search-driven responses.

• AI Agents in Virtual Agent (available from Yokohama Patch 3 / Xanadu Patch 9): AI Agents are now surfaced directly within the Now Assist Virtual Agent experience. The Virtual Agent uses Now LLM to identify the best AI Agent for the requestor's query and routes to it automatically — no explicit trigger required.
• Prebuilt ITSM AI Agent use cases include incident triage, root cause assembly, and resolution guidance — providing genuine multi-step troubleshooting before escalation.
• For knowledge-driven troubleshooting, configure AI Search profiles to include the relevant internal KBs and any connected enterprise content sources. Now Assist in VA will use that profile to generate contextual answers.
• Live agent escalation remains configurable within the VA topic design when AI results are insufficient.
• AI Agents in Virtual Agent for autonomous troubleshooting, backed by a well-tuned AI Search profile — is the closest current equivalent to an L1 agent experience.