Use PDIs? Take our 5-minute survey to help shape the PDI roadmap.

ServiceNow AI Agents — Build, Operations & Governance Guide (Australia release)

Varun_Gupta1
ServiceNow Employee

an hour ago

ServiceNow AI Agents: Build, Operations & Governance Guide

Australia release · Varun Gupta

Most AI agent problems in the field are not model problems. They are the same four questions asked too late: what should I have checked before building, why is this burning assists, where do I look when it fails, and who was supposed to approve it. This guide answers all four in one place, across the full lifecycle: build, operate, govern.

What it is. A single working reference for anyone who builds, runs, or approves AI agents on the ServiceNow platform, current to the Australia release. It pulls scattered guidance into one place so you are not hunting across a dozen articles mid-build.

Who it is for. Four roles, each with its own reading path in the navigation index below, so nobody has to read all the pages:

  • Power User — activating a first agent, no code
  • Admin — configuring, monitoring, and governing
  • Developer — building custom agents and tools, and debugging
  • Architect — design, guardrails, and the governance framework

What you get from it. Pre-build guardrails that stop a bad agent before it exists; a clear out-of-the-box versus custom decision rule; loop prevention and trigger design with the real property defaults; the exact debug table sequence to walk when an agent misbehaves; how tool count drives assist consumption and how to forecast and control it; and a full governance layer covering approval gates, AI Control Tower, data privacy, and an AI incident response runbook.

How to read it. Content is either documented ServiceNow platform behaviour, or recommended practice I have built where no official equivalent exists. Where a section is my own framework rather than product documentation — the incident severity model, the response sequence, the RACI extension, the proposed governance model in Section 11 — it says so in the text. Treat those as guidance to adapt, not as ServiceNow product capability. Every table name and system property in the reference sections was verified before publishing.

The governance sections are opinionated, especially on incident response and the approval model. That part is my view, not ServiceNow's. Tell me where you disagree, and tell me what the debug section is missing.

Guardrail: Use this guide before creating or modifying any AI Agent on the Now Platform. Applicable to Architects, Developers, Admins, and Power Users.

Persona-Based Navigation — Start Here

Find your role below and click the links to jump directly to the sections most relevant to you. Each link takes you to the exact section — use "Back to top" at the end of each section to return here.

Power User — Building Your First AI Agent — No code required. Start here if this is your first AI Agent.

Admin — Configuring, Monitoring & Governing — Platform configuration, Now Assist setup, analytics, and governance.

Developer — Building & Debugging AI Agents — Custom agent development, tool configuration, loop prevention, and debugging.

Architect — Governance, Design & Decision Making — Platform architecture, governance framework, design patterns, and guardrails.

Full Table of Contents

Click any section title to jump directly to it. Use "Back to top" links throughout the document to return to this index.

How to Navigate This Guide

This guide serves four audiences. Each has a recommended reading path — use the audience indicators throughout to know which sections are most relevant to you.

Audience Role Start Here Key Sections
Power User Building first AI Agent with no-code tools Sections 1 → 3 → 5 → Bonus Guardrails, Admin Setup, Build Process, Quick Start
Admin Configuring and monitoring Now Assist Sections 1 → 3 → 6 → 8 Guardrails, Admin Setup, Loop Prevention, Assists
Developer Building custom agents, tools, integrations Sections 2 → 4 → 5 → 6 → 7 OOB Decision, Architecture, Build Process, Loops, Scenarios
Architect Governance, design patterns, guardrails Read all — focus on 1, 4, 6, 8 Guardrails, Architecture, Loop Prevention, Decision Flow

Throughout this guide, each section and scenario includes an audience tag like this:

For: Architect · Developer · Admin · Power User — All audiences — foundational content

↑ Back to navigation index

Section 1 Pre-Build Guardrails

Stop and answer every question before touching the platform

For: Architect · Developer · Admin · Power User — Required for everyone — no exceptions

Answer all questions below before building anything. Any 'No' or 'Unknown' = stop and resolve before proceeding.

Platform & Licensing

  • Instance on Australia (current release, GA 5 May 2026). Required for AI Agent Fabric, MCP, A2A, Role Masking, and Agentic Playbooks.

  • AI Native tier active (Foundation, Advanced, or Prime); custom AI skill and agent creation requires Prime. Legacy Pro Plus or Enterprise Plus remain valid until renewal and reached end of sale on 1 July 2026.

  • sn_aia.admin role granted to admin user

  • ServiceNow Otto panel and AI Search enabled on instance

Design & Governance

  • OOB agent checked in AI Gallery AND Autonomous Workforce catalog BEFORE any custom build

  • Use case fully documented as a process chart

  • Data quality confirmed — tickets, KB articles, CMDB records

  • Governance / change board approval obtained

Risk & Cost

  • Human-in-the-loop defined for high-risk or irreversible actions

  • Trigger recursion risk assessed (see Section 6)

  • Expected Assists consumption estimated and approved

  • Sub-production test environment ready — never first-run in production

  • Data classification completed for every table and field the agent reads or writes (see 11.13)

  • Autonomous tool-execution override reviewed: sn_aia.enable_usecase_tool_execution_mode_override

↑ Back to navigation index

Section 2 Core Concepts & OOB Decision

Know the terminology — then always check OOB first

For: Architect · Developer · Admin · Power User — Foundational terminology for all audiences

Key Terminology

Term Definition
Agentic Workflow The 'Why' — the overall business scenario being automated. Defines the objective.
AI Agent The 'Who' — a modular worker that performs one specific task within a workflow.
Tool The 'What' — an action the agent executes (flow action, script, KB search, record op).
Orchestrator The LLM coordination layer (GPT-4.1 by default) that plans and sequences agents.
AI Agent Studio No-code/low-code UI for building and managing agents. Navigate: All > AI Agent Studio.
ServiceNow Otto panel (formerly Now Assist Panel) Chat interface where users interact with AI Agents in real time.
AI Control Tower Platform-native AI governance layer, bundled in every AI Native tier. Five dimensions: Discover, Observe, Govern, Secure, Measure. Covers AI asset inventory, runtime monitoring, risk assessment against NIST AI RMF and EU AI Act frameworks, access control, a real-time kill switch, and cost plus value tracking. See section 11.12.
Short-Term Memory (STM) Temporary in-session memory shared across agents within one conversation. OOB enabled.
Long-Term Memory (LTM) Persistent cross-session memory stored in sn_aia_memory table (vector format).
Assists Cost unit. Consumed per tool action executed — not per conversation.
AI Agent Fabric Interoperability layer enabling ServiceNow AI Agents to connect with external agents and tools via MCP and A2A protocols. Available Zurich P4+.
Model Context Protocol (MCP) Open-source standard (Anthropic) for connecting AI agents to external tools, data, and systems. ServiceNow supports MCP Client (consume external tools) and MCP Server (expose AI skills). Zurich P4+.
Agent-to-Agent (A2A) Google-developed protocol enabling communication and context sharing between ServiceNow AI Agents and external AI agents. Zurich P4+.
AI User New sys_user identity type for AI Agents. Pre-configured with specific roles independent of the invoking user. Use when agent needs permissions greater/different than the invoking user. Zurich P1+.
Role Masking Required security configuration (Zurich P4+, store app v6.0+) that limits AI Agent execution to a subset of the invoking user’s roles. Applies hierarchically: Agentic Workflow → Agent → Skill Tool.
Agentic Playbooks Zurich feature that weaves AI agents into individual playbook tasks, blending human-in-the-loop decisions with autonomous steps. Users can modify playbooks via natural language.
Autonomous Workforce Product line (Australia release, GA Q2 2026) deploying AI Specialists with defined roles, permissions, and business context to execute end-to-end workflows. First OOB: L1 Service Desk AI Specialist.
Build Agent Zurich conversational AI agent in ServiceNow IDE that creates apps, writes scripts, generates files, and manages the full development lifecycle via natural language (vibe coding).
EmployeeWorks Moveworks-powered conversational AI front door (GA Feb 2026) providing a single entry point for employees via Teams, Slack, or browser. Routes natural-language requests through governed ServiceNow workflows.
AI Admin Center Australia-release single control hub for AI. Adds natural-language (conversational) configuration, instance-specific recommendations, and embedded usage, performance, and business-value analytics on the home page. Supersedes the scattered Now Assist admin consoles.
AI Agent Advisor Capability inside AI Admin Center (GA May 2026) that analyzes your instance data (incident and case tables by default) to surface high-impact automation opportunities and matches them to out-of-the-box AI agents.
Action Fabric Governed interface that exposes the ServiceNow system of action (flows, playbooks, approvals, catalog) to external AI agents through a GA MCP Server. Identity-verified and auditable. Included in every Now Assist and AI Native SKU.
AI Native SKUs (Foundation / Advanced / Prime) Three AI-native license tiers that replaced the legacy Standard, Pro, Pro Plus, Enterprise, and Enterprise Plus tiers on 9 April 2026. Now Assist, Moveworks, Workflow Data Fabric, and AI Control Tower are bundled into every tier. Custom AI skill and agent creation requires Prime.

OOB vs Custom — The Rule

Rule #1: Always check the AI Gallery before building anything custom. Custom agents should only be built when OOB agents cannot satisfy the use case — even with configuration.

Situation Recommended Path
Use case matches OOB agent (ITSM, HR, CSM, SecOps) Activate OOB → configure → test. Done.
OOB agent close but needs minor tuning Clone OOB → modify instructions/tools only
No OOB agent; process is well-defined Build custom via AI Agent Studio
Process is unclear or undocumented STOP. Document the process first. Return here after.
Use case spans multiple modules Multi-agent agentic workflow with orchestrator

AI Gallery (primary OOB repository): applibrary.service-now.com/ai_gallery

↑ Back to navigation index

Section 3 Admin Setup — Now Assist Configuration

For: Admin · Power User — Admins configuring Now Assist | Power users getting their environment ready

Admin Console — Your Command Center

Navigate: All > AI Admin Hub. This is your central point for activating and managing all AI capabilities. Available OOB in Vancouver and newer releases.

New in the Australia release: AI Admin Center is the single control hub for all AI capabilities and supersedes the scattered Now Assist consoles. It adds natural-language configuration, instance-specific recommendations, and an AI Agent Advisor that reads your instance data to surface high-value automation candidates. Usage, performance, and business-value analytics sit on its home page. Use it as your day-to-day entry point on Australia; the Admin Console steps above still apply on earlier releases.

6-Step Activation Sequence

Step Action Navigate To
1 — Plugin Authorization Authorize required plugins from ServiceNow Store ServiceNow Store > Authorize plugins
2 — Install Install authorized apps in your instance All > System Applications > Application Manager
3 — Skill Activation Filter by workflow; activate via guided wizards Admin Console > Skills tab
4 — Experience Config Configure where skills appear (on-record, Otto panel, context menus) Admin Console > Experiences tab
5 — Role Assignment Restrict/grant skill access by user role Admin Console > Permissions
6 — AI Agents Enable Grant sn_aia.admin role; enable AI Search and the Otto panel User Admin > Roles | All > AI Search Status

Analytics Dashboards — What to Monitor

  • Skill adoption by department

  • Self-service resolution rates

  • Guardian monitoring — prompt injection detection

  • Value Insights — time saved per AI skill

  • Assist Consumption per Agentic Workflow

  • Status of Execution Plans — real-time view

 

Loop Monitoring via Analytics

Navigate: AI Agent Studio > AI Agent Analytics. Use two key tabs:

  • Status tab — real-time Execution Plans and statuses over time

  • Assist Consumption tab — graphical view + Top 10 Agentic Workflows by Assists consumed

Warning Signs of a Loop: Multiple invocations from a single test | Sudden Assists spike | Single workflow consuming disproportionately high Assists in short window

Pro Tip: Use analytics data to build ROI case for leadership. Enhancements ship every 2-3 months — review settings after each store update.

↑ Back to navigation index

Section 4 Architecture Principles

Design for modularity, security, and cost control

For: Architect · Developer — Architects designing agent topology | Developers implementing agents

4.1 Modularity Rules

  • One agent = one responsibility. Never build mega-agents.

  • Limit agents per use case to ~10 or fewer for clarity and maintainability.

  • Every agent must have at least one tool — agents without tools produce no reliable output.

  • Prefer Flow Designer over scripts — upgrade-safe and auditable.

  • Set tool Output Transformation Strategy to 'Concise' unless verbose output is required.

4.2 Context Window — Critical Limit

128K Token Limit: Exceeding causes unpredictable agent behavior. Token consumption includes system prompts, tool outputs, conversation history, and instructions. Keep agents focused.

4.3 Orchestration Strategy

Strategy Behavior Best For Avoid When
Base Planner One-time static plan, no replanning Simple, linear, deterministic workflows Branching logic or error recovery needed
ReActive Planner Iterative, handles dynamic conversation Chat-driven, multi-turn, unknown flow Batch processing of many records
Batch Planner Groups tasks, mid-execution replanning Parallel subtasks with dependencies Simple single-step operations

Configure under: AI Agent Studio > [Your Use Case] > Strategy field (sn_aia_agent → strategy).

Australia Release: The AI Agent Orchestrator is enhanced as a centralised command centre allowing specialised AI agents to collaborate across departments. The Autonomous Workforce product line introduces AI Specialists with pre-assigned roles, permissions, and business context for end-to-end workflow execution.

4.4 Security & Governance

  • Each agent runs under the Run-As user's ACLs — access controls are always enforced.

  • Apply the principle of least privilege to the run-as user.

  • All agent actions are logged in AI Control Tower — every tool execution is audited.

  • Use AI Guardian for prompt injection protection.

  • Domain separation is supported (sys_domain exists on all agent tables).

  • Zurich P1+ introduces AI User identity type — a new sys_user that runs with pre-configured roles independent of the invoking user. Choose between Dynamic User (inherits invoking user’s permissions, default) or AI User (fixed role set).

  • Role Masking is a REQUIRED configuration on Australia (store app v6.0+). It limits the roles under which AI Agents execute to a defined subset. Applies hierarchically across Agentic Workflow → Agent → Skill Tool.

  • Role-based access configuration is now mandatory in AI Agent Studio — you must define who can discover/use an agent and the run-as entity before saving.

  • Use GlideRecordSecure() and addUserEncodedQuery() in Script tools and Flow Script steps for security-hardened data access.

  • Use Supervised mode for any high-risk, irreversible, or sensitive tool actions. Supervised mode also mitigates potential negative impact of role-masking misconfigurations.

  • Kill switch (Australia): AI Control Tower can disable a model and its tool access in real time through the AI Gateway. Use it to stop a misbehaving agent immediately, without waiting for a code change or deactivation flow.

4.5 LLM Selection

Release Default LLM Alternatives Available
Australia (current) GPT-4.1 (OpenAI on SN Azure) NowLLM, Claude, Gemini
GCC / Self-hosted NowLLM (supported July 2025+) Confirm with account team for other models

Australia Update (Official): GPT-4.1 remains the default orchestrator LLM. NowLLM, Claude, and Gemini are available alternatives. BYOK or models not hosted by ServiceNow may incur additional charges. The Australia release introduces AI Agent Fabric with MCP/A2A support, enabling external LLM and agent connectivity governed via AI Control Tower. Action Fabric extends this by exposing the ServiceNow system of action (flows, playbooks, approvals, catalog) to external AI agents through a GA MCP Server, identity-verified and auditable, and is included in every Now Assist and AI Native SKU.

↑ Back to navigation index

Section 5 Step-by-Step Build Process

Discovery → Design → Tools → Memory → Triggers → Test

For: Developer · Power User — Developers building custom agents | Power users using AI Agent Studio guided interface

Phase 1 — Discovery & Design (before touching the platform)

  • Map the complete use case as a process chart before opening AI Agent Studio.

  • Identify all system actors: which tables are read/written, which integrations are called.

  • Define success criteria — what does 'resolved' look like for this use case?

  • Check AI Gallery for OOB agents (Section 2 rule always applies).

  • Validate with stakeholders and process owners before building.

Power User Tip: Use Now Assist's built-in authoring tool to generate agent role and instructions — just type a description of what your agent should do. Now Assist generates the prompt for you.

Phase 2 — Agent Design

  • Define the Agentic Workflow (the 'Why') — the overall business scenario.

  • Break into individual Agents (the 'Who') — one responsibility per agent.

  • Write the Agent Role: domain, scope limits, tone, audience — be specific.

  • Write Agent Instructions: step-by-step guidance, name each tool explicitly.

  • Example: 'If incident priority is P1 or P2, immediately call [Notify On-Call] tool.'

Phase 3 — Tool Configuration

Tool Type Use For Key Setting
Flow Action Calling existing SN flows (preferred) Set Input/Output as String — upgrade-safe
Record Operation CRUD on ServiceNow tables Scope to minimum required fields
KB Search (AI Search) Searching knowledge articles via RAG Use Hybrid search; confirm Search Profile published
Subflow / Spoke Integration Hub calls to external systems Integration Hub spokes are AI Agent-ready
Script Custom logic not available via Flow Designer Minimize — use only when no other option

Zurich P4+ Tool Addition — MCP Server Tool: Connect AI Agents to external tools, data, and systems via Model Context Protocol. Configure in AI Agent Studio → MCP Server Console. Requires sn_mcp_server plugin. External tools consumed via MCP Client do not consume Assists; inbound Now Assist invocations by external agents do consume Assists.

For each tool, define: Execution Mode (Supervised vs Autonomous), Display Output (Yes/No), Output Transformation Strategy.

Phase 4 — Memory Strategy

Memory Type Scope Use When How to Enable
Short-Term (STM) Single conversation Sharing data between agents in one use case OOB — reference in agent instructions
Long-Term (LTM) Across sessions (user-linked) Personalization, reducing repeat questions Configure in AI Agent Studio; stores to sn_aia_memory

Phase 5 — Trigger Setup

See Section 6 for the full trigger deep-dive and loop prevention patterns.

  • Triggers are optional — Otto panel manual invocation does not require one.

  • Set trigger conditions as granular as possible to prevent unintended executions.

  • CRITICAL: Assess recursion risk before activating any record-based trigger.

Phase 6 — Test & Validate

  • Always test in sub-production first — never first-run in production.

  • Test with at least 5-10 records to surface loop behavior early.

  • Review debug tables after each test run (see Section 8 for the sequence).

  • Validate Assists consumption — confirm within expected bounds before promoting.

Manual Testing and Automated Agentic Evaluations

New in the Australia refresh. Deepens Phase 6 above.

Phase 6 is the checklist. This is the out-of-the-box tooling behind it. ServiceNow ships two complementary evaluation layers natively, so you do not build a custom test harness. Manual testing drives fast iteration during build; automated Agentic Evaluations validate the whole agent at scale before you deploy. The product loop is Configure, Evaluate, Optimize. Both consume assists (see 9.6).

Surface Type What it does Where
Testing tab Manual Single-execution, interactive run. Exposes the agent reasoning, each tool call, and the decision path so you see what the agent does on one scenario. Validates instructions, reasoning, and tool selection. AI Agent Studio > Testing
Agentic Evaluations Automated An LLM-as-judge reads each execution log and scores the full decision sequence (every tool selection, parameter, and step), not just the final output. Runs across many logs at once. Role: sn_aia.admin. AI Agent Studio > Testing > Automated evaluations, or All > AI Skill Kit > Agentic Evaluations
Auto-Evaluation for AI Skills Automated Judge-based scoring of a single skill in isolation, for example a summarization or classification skill: does this one capability produce quality output, separate from agent orchestration. AI Skill Kit

Running an automated Agentic Evaluation, in order:

  • Configure: pick the agent or agentic workflow, name the run, then Start automated evaluation.

  • Choose metrics: Overall Task Completeness is the default. Each execution log is rated fully, partially, or not completed. Tool-calling accuracy and step-level quality are the other judged dimensions. Confirm the exact metric set in your instance.

  • Build the dataset: three out-of-the-box options. Existing execution logs give a real-world distribution for agents already live. Manual testing produces fresh logs with full control over scenarios and edge cases. Automatic generation has an LLM drive the agent conversationally across many records to produce 50 to 100+ scenarios fast, best for at-scale pre-deployment validation. AI Agent Advisor auto-builds a dataset from its opportunity clusters, pre-selected in the tool, so you test against the exact records that flagged the opportunity.

  • Evaluate: read the Evaluation Results Dashboard. Overview tab for headline scores, Detailed Results for per-record drill-down. Exclude unrepresentative records, tune thresholds to your risk tolerance, clone the run for before and after, and export to CSV.

  • Optimize: fix the pattern, not the record. If many runs fail for the same reason, for example the agent picks the wrong tool, correct that tool description rather than the individual cases. Re-run after each change. Most teams iterate two to four times to reach deployment-ready scores.

Results flow to AI Control Tower > Evaluation tab for a cross-agent governance roll-up (see 11.12), so test evidence sits in the same control plane as risk and value.

Honest limit: Agentic Evaluations is decision support, not a deployment gate on its own. Scores are AI-generated and probabilistic. Pair it with human sign-off: the sub-production validation in Phase 6, the ARB gate in 11.5, and manual review of high-risk paths.

Testing discipline that separates real confidence from false confidence:

  • Negative testing: deliberately feed contradictory, out-of-scope, and edge-case inputs. Passing only the happy path is false confidence.

  • Dataset diversity: test data that all looks alike overstates readiness. Mirror the real spread of production inputs.

  • A/B on changes: for any model swap or prompt change, run parallel evaluations to measure the impact before you switch.

  • Regression cadence: schedule evaluation runs on every upgrade, use case change, and adoption of a new out-of-the-box agent. Validate ServiceNow OOB use cases in your own environment too.

  • Cost awareness: evaluation runs consume assists, roughly 1 assist per record processed for agentic evaluation. Budget it into your assists plan (9.6).

For the deterministic platform an agent depends on (business rules, client scripts, flows), regression-test with the Automated Test Framework (ATF). Australia adds ATF Code Coverage and an ATF Troubleshooting Agent, a Now Assist for Creator skill that analyzes a failed test, finds root cause, and suggests fixes. Agentic Evaluations judges probabilistic agent behavior; ATF covers the deterministic scaffolding around it. To confirm evaluation prerequisites in one place, run the Now Assist Readiness Evaluation store app.

↑ Back to navigation index

Section 6 Loop Prevention & Trigger Design

For: Architect · Developer · Admin — Anyone designing or activating trigger-based agents

Recursive loops — where an agent's actions re-trigger itself — are the #1 cause of runaway Assists consumption. Proper trigger design prevents them.

6.1 Trigger Types — When to Use Each

Trigger Type When to Use Key Caution
On Record Created Agent fires once on new record creation Safe alone. Dangerous combined with On Record Updated without filter conditions.
On Record Updated Agent reacts to changes in existing records High recursion risk if agent writes to the same record it watches.
On Record Created and Updated Agent handles both creation and updates Requires strict filter conditions. Test with 5-10 records first.
Scheduled (Daily/Weekly) Analyzing patterns across many records Do NOT loop inside Agentic Workflow. Use a Flow to iterate; call agent per record.
Inbound Email Agent processes incoming emails NEVER combine with record-based triggers — email may fire both simultaneously.

For: Architect · Developer — Preferred pattern — no custom fields, built-in audit trail, AWA/PI compatible

  • Assign the record to a Virtual Workers assignment group when the agent should process it.

  • Set trigger filter: Assignment Group = [Virtual Workers Group]

  • Agent processes the record (adds work notes, resolution plans, etc.).

  • Agent tool reassigns the record to a human fulfiller group as its final step.

  • Reassignment makes the filter condition FALSE → loop broken automatically.

Benefit Detail
No custom fields required Uses existing Assignment Group field — no schema changes
Native audit trail ServiceNow logs all assignment group changes automatically
AWA/PI compatible Works with existing Advanced Work Assignment and Predictive Intelligence rules
Clear ownership model Virtual Workers group = AI processing | Human group = human review

6.3 Loop Prevention Pattern 2 — Custom True/False Flag

  • Add a boolean field 'AI Agent Processed' to the table.

  • Set trigger filter: 'AI Agent Processed = false'

  • Agent sets the field to TRUE as its FIRST action — before any other processing.

  • Subsequent record updates no longer satisfy the trigger condition.

Caution: Do NOT modify the OOB State field — other AI skills depend on it. Use a custom boolean field instead.

6.4 Built-In Recursive Check Properties

Navigate to: All > sn_aia_property.LIST. Requires Now Assist AI Agents plugin (sn_aia) v4.0.38+. Store app v6.0+ adds role masking as an additional security layer.

Property Trigger Scope Default Limit Time Window
recursive_check.query_for_create_record Scheduled, API pathway, batch 50 executions (create_max_executions) 15 min (create_time_window)
recursive_check.query_for_update_record On Record Updated 5 executions (update_max_executions) 15 min (update_time_window)

Design Rule: These properties are safety nets — not design strategies. If you are relying on them to stop loops, your trigger design needs to be revised.

6.5 Emergency Loop Response — Live in Production

  • Step 1 — Deactivate trigger: AI Agent Studio > [Agent] > Trigger tab > Set Inactive.

  • Step 2 — Cancel stuck: All > sn_aia_execution_plan.LIST > Filter State = In Progress > Cancel.

  • Step 3 — Analyze: Review which fields the agent updates and why the filter condition stayed true.

  • Step 4 — Redesign with Pattern 1 or 2. Test with 5-10 records. Re-enable only when loop is confirmed resolved.

↑ Back to navigation index

Section 7 Scenario Flow Guides

Real situations — DO and AVOID guardrails

For: Architect · Developer · Admin · Power User — Use these to navigate common decisions and situations

Scenario: 1 — New Request: Build vs OOB [All Audiences]

Situation: Team requests a custom AI Agent for password resets via Virtual Agent.

DO: Check AI Gallery first. ServiceNow ships a Password Reset agent in the ITSM collection. Activate, configure the run-as user and Identity Provider integration, test, and deploy. No custom build needed.

AVOID: Building a custom agent from scratch. Duplicates OOB capability, creates upgrade risk, and increases Assists consumption without added value.

Scenario: 2 — OOB Agent Modification [Developer / Architect]

Situation: The OOB Incident Categorization Agent doesn't match your internal CI taxonomy.

DO: Clone the OOB agent. Modify instructions to reference your CI naming conventions. Add a Record Operation tool to look up your internal classification table. Scope within a custom scoped application.

AVOID: Editing the OOB agent directly in place. ServiceNow upgrades will overwrite your changes. Always clone and scope within a custom application.

Scenario: 3 — Multi-Agent Design [Architect / Developer]

Situation: Complex employee onboarding spanning HR, IT, and Facilities — all requiring automation.

DO: One Agentic Workflow with 3-4 specialized agents (IT Provisioning, Device Order, Facilities Access). Use STM to share the new hire record sys_id across agents. Use Batch Planner for parallel execution. Keep each agent under 5 tools.

AVOID: One mega-agent with 15+ tools. Bloats context window (>128K token risk), makes debugging impossible, and creates a single point of failure.

Scenario: 4 — Trigger Design: Loop Prevention [Developer / Admin]

Situation: Incident Triage Agent triggered On Record Created. It adds resolution notes — which triggers On Record Updated again.

DO: Use Assignment Group pattern: filter = 'Assignment Group = AI Triage Group'. Agent processes, then reassigns to human group. Filter condition becomes false. Loop broken with a built-in audit trail.

AVOID: 'Active = true AND State = New' without a loop-breaking field. Stays true after work note update. Hits the 5-execution recursive limit and wastes Assists.

Scenario: 5 — Supervised vs Autonomous [Architect / Admin]

Situation: SecOps Agent detects a security alert and wants to automatically isolate a compromised endpoint via CrowdStrike.

DO: Set the isolation tool to Supervised mode. Agent generates the isolation plan and presents it to the on-call security analyst for approval before execution. High-risk, irreversible actions always require human-in-the-loop.

AVOID: Autonomous mode for endpoint isolation. Can cause business disruption and violates most security governance frameworks.

Scenario: 6 — Updating a Production Agent [Developer / Admin]

Situation: Production Change Management Agent needs a new Jira notification integration added.

DO: Clone the production agent to a dev scope. Add and test the Jira Integration Hub spoke tool in dev. Validate with ATF and test conversations. Promote via standard change process. Monitor AI Control Tower for 48 hours post-deploy.

AVOID: Editing the live production agent directly. Even minor prompt changes alter non-deterministic behavior.

↑ Back to navigation index

Section 8 Debug, OOB Agents & Decision Flow

Where to look when things go wrong + OOB catalog + decision guides

Debug Sequence — Follow in Order

For: Developer · Admin — Debugging agent failures or unexpected behavior

Step Table / Location What You'll Find
1 sn_aia_execution_plan.LIST Which agents were planned, in what order, and their current status
2 sn_aia_execution_task Individual agent execution tasks and statuses
3 sn_aia_tools_execution.LIST Tool-level execution logs including inputs and outputs
4 sys_gen_ai_log_metadata Raw generative AI logs — requires admin role
5 sn_aia_message.LIST Full sequence of system messages in the conversation
6 sys_cs_message.LIST Conversation messages including all user inputs

Top Pitfalls & Fixes

Pitfall Symptom Fix
Overly complex agents Inconsistent or hallucinated outputs Reduce tools per agent; split into focused agents
Missing tool descriptions Orchestrator selects wrong or skips tools Write detailed Name and Description for every tool
Context window overflow Unpredictable outputs; agent stops mid-task Reduce tools; set Concise output; use STM
Recursive trigger loop Agent fires repeatedly; Assists spike Apply Assignment Group or Custom Flag pattern (Section 6)
Missing run-as user ACLs Agent response not visible; action fails silently Verify ACLs for all tables/fields the agent touches
Vague prompt instructions High variance in outputs across runs Add explicit step-by-step instructions; name tools directly

OOB AI Agent Catalog

For: Architect · Developer · Admin · Power User — Check before building any custom agent

Domain Key OOB Agents Install Path
ITSM Incident Categorization, Change Management, Post-Incident Review, Resolution Plan, Password Reset Now Assist for ITSM store app
HRSD Employee Onboarding, Policy Q&A, Leave Request, Benefits Inquiry Now Assist for HRSD store app
CSM Case Resolution, Customer Inquiry, Escalation Management Now Assist for CSM store app
SecOps Security Incident Lifecycle, Threat Triage, Phishing Response Now Assist for SecOps store app
CRM / SPM Lead Management, Project Status, Risk ID, Resource Allocation Now Assist for CRM / SPM store app
Voice Agents IT Voice Agent, HR Voice Agent (EN, DE, ES) sn_itsm_voice_aia collection

Additional OOB Agents (Zurich / Australia):

  • Creator Agents (Zurich): Build Agent (conversational app builder in ServiceNow IDE), UI Builder Agent (in-context page configuration assistant), ATF Troubleshooting Agent (automated test framework debugging).

  • Autonomous Workforce (Australia, GA Q2 2026): L1 Service Desk AI Specialist — first OOB AI Specialist. Autonomously diagnoses and resolves common IT requests (password resets, software access, network troubleshooting) end-to-end. Operates 24/7, escalates when human intervention needed. ServiceNow internal deployment: 90%+ of employee IT requests handled autonomously, 99% faster than human agents.

  • Agentic Playbooks (Zurich): AI agents woven into traditional ServiceNow playbooks, completing tasks autonomously while keeping humans in the loop. Users can modify playbooks via natural language.

Decision Flow — Every New AI Agent Request

For: Architect · Developer · Admin · Power User — Follow every step in sequence — no shortcuts

Step Question YES → Action NO → Action
1 Process documented and mapped? Continue to Step 2 STOP. Document the process first.
2 OOB agent exists in AI Gallery? Activate + configure OOB. Done. Continue to Step 3
3 OOB agent can be cloned and adapted? Clone + modify only. Go to Step 5. Continue to Step 4
4 Use case scoped and <10 agents? Proceed to AI Agent Studio. Split into smaller use cases first.
5 Governance approval + recursion risk assessed? Continue to Step 6 Obtain approvals + complete Section 6 assessment.
6 Sub-production tested (5-10 records)? Promote to production. Return to testing.
7 Monitor AI Control Tower for 48 hours? Agent confirmed stable. Done. Set monitoring alerts in AI Control Tower.

↑ Back to navigation index

Section ★ Power User Quick Start

Building your first AI Agent — start here, no code required

For: Power User — If this is your first AI Agent — follow these 5 steps

AI Agent Studio provides a guided, no-code interface. Now Assist even writes the agent role and instructions for you. Here is the simplest path to your first working agent:

Visit applibrary.service-now.com/ai_gallery. Search for your use case. 90% of common needs already have an OOB agent ready to activate. If you find one, jump to Step 5.

Step 2: Complete the Guardrail Checklist (Section 1)

Make sure your instance, license, role (sn_aia.admin), and process documentation are all confirmed before opening AI Agent Studio. Skipping this is the most common cause of failed first builds.

Step 3: Use Now Assist Authoring to Write Your Agent

In AI Agent Studio — describe what your agent should do in plain language. Now Assist generates the Role and Instructions for you. Review, refine, and save.

Step 4: Add Tools Using the Guided Tool Picker

Use Flow Actions first — they appear as guided options and are upgrade-safe. Set Execution Mode to Supervised for your first agent. This means the agent will ask for human approval before taking any action, giving you a safety net while you learn.

Step 5: Test in Sub-Production Before Going Live

Run 5-10 test records in a sub-production environment. Check AI Agent Studio > AI Agent Analytics > Status tab. Confirm no unexpected loops or errors. Only then promote to production.

Key Rule: When in doubt, use Supervised mode. The agent asks for human approval before acting. It is always safer for first builds and can be switched to Autonomous once you are confident.

Platform Navigation for Power Users

What You Need Navigate To
Build your first agent All > AI Agent Studio > Create and Manage
Find OOB agents applibrary.service-now.com/ai_gallery
Check if your instance is ready All > AI Search Status | All > AI Admin Hub
Monitor your agent after go-live AI Agent Studio > AI Agent Analytics
Get help from Now Assist authoring AI Agent Studio > Create Use Case > Use AI to Generate

↑ Back to navigation index

Section 9 Dashboards, Reporting & Value Measurement

For: Architect · Admin — Measuring adoption, tracking execution, and quantifying ROI

Before launching any use case, identify exactly which metrics you will track. Instrumentation must be in place BEFORE go-live — not after.

9.1 The 5 Standard Adoption Metrics — Track All of Them

ServiceNow measures these five metrics across every Now Assist use case. Apply the same framework to AI Agents.

Metric What It Measures How to Track
Adoption Monthly Active Users (MAU%) — Now Assist users as % of total potential users AI Analytics Dashboard > Adoption tab
Usage Who is using the feature — triggers like summaries generated, agents activated AI Analytics > Usage tab | AI Agent Analytics > Status tab
Sentiment Positive feedback as % of total feedback — thumbs up/down in ServiceNow Otto panel AI Analytics > Sentiment tab | Guardian monitoring
Accuracy Search success rate, conversation success rate, code acceptance rate (60-100%) AI Analytics > Accuracy tab | AI Search Preview
Hours Saved Basis of all value calculations — agent AND requestor. Convert to $ value. AI Analytics > Value Insights tab | AI Agent Analytics > Assist Consumption

9.2 Platform Dashboards — Where to Find Every Report

Dashboard Navigate To Key Data Points
AI Analytics All > AI Admin Hub > Analytics 7 dashboards: Adoption, Usage, Sentiment, Accuracy, Hours Saved, Guardian, Value Insights
AI Agent Analytics — Status AI Agent Studio > AI Agent Analytics > Status tab Execution Plans over time, agent statuses, execution counts
AI Agent Analytics — Assists AI Agent Studio > AI Agent Analytics > Assist Consumption tab Assists consumed per workflow, Top 10 workflows by cost, consumption trends
AI Control Tower All > AI Control Tower Real-time agent status, all tool executions, full audit log; plus the five AICT dimensions (Discover, Observe, Govern, Secure, Measure), including cost and value tracking. See 9.5 and 11.12.
AI Search Analytics All > AI Search > Search Analytics Query volumes, result relevance scores, Genius Result acceptance rates
Guardian Monitoring All > AI Admin Hub > Guardian Prompt injection attempts detected, content policy violations
AI Skills Usage All > AI Admin Hub > Skills tab Activation status, per-skill usage volumes, user access

Pro Tip: Export dashboards and assign read-only viewer roles to leadership stakeholders. Track adoption metrics, analyze for trends, then layer qualitative and quantitative data into your value story.

9.3 Now Assist Value Calculation — 5 Proven Value Drivers

ServiceNow's own methodology for quantifying the financial impact of Now Assist. All calculations convert time saved into dollar value using defensible assumptions.

Value Driver 1 — Agent Productivity (Summarization & Resolution Notes)

The formula: Now Assist uses × Acceptance rate × Time saved per use × Cost per agent = $ value per day.

  • Agent reading time saved: Summary word length ÷ median technical reading rate

  • Agent writing time saved: Summary word length ÷ median technical writing rate

  • Total time saved per use: ITSM agents save 4-6 minutes per use | CSM agents save 12-16 minutes per use

  • Acceptance rate baseline: 60-100% depending on implementation maturity

Value Driver 2 — Requestor Productivity (Now Assist in Search / Portal Q&A)

  • Each successful Now Assist search saves approximately 2 minutes vs traditional click-through (30 seconds vs 2.5 minutes)

  • ~5% of successful search interactions would have otherwise become a low-complexity support case

  • Formula: Successful searches × Time saved per interaction = Total requestor hours saved

Value Driver 3 — Virtual Agent Self-Service

  • Successful session = employee resolves issue without human agent assistance

  • Average time on successful VA topic: 3.68 minutes vs ~15 minutes for a live case

  • Fully automated Topics and Conversational Catalog Items: 100% of successful interactions counted as cases avoided

Value Driver 4 — Deflection Impact

ServiceNow's deflection definition: user is offered Now Assist response AND does not submit a ticket within 24 hours of interaction.

  • Positive engagement signals that validate deflection: thumbs up, scroll/expand, or 15+ seconds on response without scrolling

  • Deflection rate = tickets created after Now Assist ÷ total requests initiated with Now Assist

  • Now Assist uplift = Now Assist deflection rate minus baseline deflection rate (without Now Assist)

Key Finding: Targeted Now Assist use cases showed deflection rates twice as high as baseline. Consistent self-service metric definition is critical — ServiceNow internally found 4 different deflection definitions that would have blocked measurement.

Value Driver 5 — Developer Productivity (Code Assist)

  • Formula: Daily Now Assist uses × Acceptance rate × Time saved per use → Annual hours saved → $ value at hourly wage

  • Track via Now Assist for Creator analytics: code acceptance rates, flow generation volumes

9.4 What Good Looks Like — ServiceNow's Own Benchmarks

Use Case Benchmark Result Timeframe
Resolution notes generation ~80% reduction in agent documentation time per note Months 1-2 of rollout
Now Assist in Search (deflection) 14% increase in employee self-service deflection rate Portal go-live
ITSM incident summarization 4-6 minutes saved per agent use Ongoing measurement
CSM case summarization 12-16 minutes saved per agent use Ongoing measurement
VA topic conversation vs live case 3.68 min (VA) vs 15 min (live case) per issue Ongoing measurement

9.5 Value Management in AI Control Tower

New in the Australia refresh.

AI Control Tower is where value now lives. The Measure dimension consolidates cost and ROI tracking for every AI asset, so you manage value in the same place you govern the agent, not in a separate spreadsheet.

What Measure gives you out of the box:

  • Realized value, adoption, and ROI tracked per AI asset across its full lifecycle, from the day it goes live. The AI Asset Owner is accountable for that value, not just for uptime.

  • Cost tracking across model providers (OpenAI, Anthropic, Google), so token consumption and spend are visible per asset and per team.

  • Per-team and per-business-unit Assist quotas, so one workflow cannot drain the shared account-level Assist pool.

  • Value and adoption reporting that ties spend to business outcomes rather than raw usage counts.

How this guide’s metrics roll up: the five adoption metrics in 9.1 and the five value drivers in 9.3 are the inputs. AICT Measure is the aggregation and reporting layer on top of them. Keep the metric definitions in 9.1 consistent across teams; four different definitions of deflection will break any aggregate value number.

Where you read it day to day: AI Admin Center shows usage, performance, and business-value analytics on its home page for the operational view. AICT Measure is the governed, cross-asset value record you report to sponsors.

Timing: the base AI value and adoption tracking is GA now. Further Measure and governance enhancements reached GA in the June 2026 release. Confirm what is live in your instance before you commit dashboard numbers to a steering committee.

Scale reference: ServiceNow has reported running its own AI estate on AI Control Tower, tracking 1,600+ AI assets and roughly $500M in cumulative AI value in 2025. The same Measure tooling is what you get out of the box.

9.6 Assists Consumption: Estimate, Forecast, and Control

New in the Australia refresh.

9.5 covers where value lives. This is the hands-on layer: how assists are counted, how to forecast spend, and the out-of-the-box controls that stop runaway consumption. AI Agents charge for value as assists, not tokens, so cost tracks the work done rather than raw model calls.

Agentic assists per execution scale with tool count:

Tools run per execution Assists consumed
0 to 4 tools 25 assists
5 to 8 tools 50 assists
9 to 20 tools 150 assists

This is why over-tooling is a cost problem, not just a design one. Fewer, well-scoped tools per agent means lower assists. Where a Virtual Agent topic or conversational catalog can cover the use case, it is cheaper than an agentic execution.

See actuals in three places:

  • AI Agents Analytics dashboard: AI Agent Studio > Analytics > Assist Consumption.

  • Now Assist subscription management: Subscription management > Subscriptions > Now Assist usage.

  • AI Control Tower consumption dashboards for proactive monitoring (see 11.12).

The underlying data lives in sys_gen_ai_usage_log (assists per Now Assist execution) and sn_aia_execution_plan (agentic executions by workflow, with an Assists reference into sys_gen_ai_usage_log).

Forecast actual spend with Platform Analytics:

  • Total assists: build an Indicator Source on sys_gen_ai_usage_log summing the Assists field, add a Historic backfill job and a Daily job, then turn on the Forecasting tab (Auto, or Linear, Exponential Smoothing, or Holt-Winters) and set a Threshold with users to notify. Show it as a line chart with Show forecast and Show forecast range, the 95% band, on a dashboard.

  • Agentic only: point the indicator at sn_aia_execution_plan with a scripted sum of assists and a breakdown by agentic workflow, so the dashboard can filter by workflow.

  • Watch-outs: assists reset annually on the contract anniversary, so reset the accumulation start date each year. Sub-production consumes assists too, so track it.

Estimate before you deploy with Performance Analytics. Use candidate volume as the driver. For incident triage: an Indicator Source on incident (Count), then a Formula Indicator of incident count multiplied by 25 for the small tier, optionally by an adoption factor such as 0.80 since not every record is agent-handled, and optionally a second formula by 50 for the medium tier. Forecast the formula indicator to project spend before go-live. Add assignment-group, priority, or keyword filters so the estimate reflects what agents will actually take.

Control runaway consumption with out-of-the-box properties in the sn_aia_property table:

Property (sn_aia_property) Purpose Default
kill_switch.mode off, warn_only, or enforce. enforce warns on Days 1 and 2, then deactivates the trigger on Day 3. warn_only
kill_switch.max_fires_per_window Fires on one record that mark it as breaching. 5
kill_switch.min_distinct_records Breaching records needed for a window to count as runaway. 25
kill_switch.window_size Length of one observation window, in minutes. 1440 (24h)
kill_switch.consecutive_windows_duration Total look-back span, in minutes. 4320 (3 days)
recursive_check.create_max_executions Max matching executions creating records. 50
recursive_check.create_time_window Window for create checks, in minutes. 15
recursive_check.update_max_executions Max matching executions updating a record. 5
recursive_check.update_time_window Window for update checks, in minutes. 15
alert.assist_spike_hours_to_check Hours between spike-check jobs. 3
alert.assist_spike_usage_percentage_threshold Percent increase that triggers the spike alert. 0.5 (50%)
alert.assist_spike_usage_threshold Minimum assists to trigger the spike alert. 5000

To cap traffic to a GenAI capability or LLM provider, by instance or by user, configure rules in sys_one_extend_rate_limit_rules. Example: rate-limit agentic executions by targeting Capability Definition = AIA ReAct Engine.

  • Keep alerts on: do not disable consumption notifications, even in sub-production.

  • Stay current: keep the Now Assist AI Agents store app at version 6.0 or later.

  • Design first: these properties are the backstop. Correct trigger design (Section 6) is the primary control against loops that burn assists.

Availability: the sn_aia_property throttling and the rate-limit properties above are out of the box as of Australia Patch 3 (June 2026). Confirm the properties and their defaults in your instance.

↑ Back to navigation index

Section 10 Admin & Developer Reference — Verified Tables, Logs & Troubleshooting

For: Architect · Developer · Admin — Reference for building, migrating, and debugging AI Agents

10.1 Core Agent Configuration Tables

For: Developer · Admin

Table Name Purpose When to Use It
sn_aia_agent Stores core definitions of AI Agents — name, role, instructions, LLM configuration.

Migration: Verify agent exists in target environment.

Dev: Confirm correct role and instructions saved.
sn_aia_agent_config Stores configuration details including activation state per agent. Dev & Validation: Verify agent is active and properly configured after deployment.
sn_aia_agent_tool_m2m Manages the many-to-many relationship between Agents and their Tools.

Migration: Verify tool associations are intact after transfer.

Debug: Confirm agent has its expected tools assigned.
sn_aia_usecase Defines the specific use cases (agentic workflow objectives) supported by AI Agents. Deployment: Validate use cases are correctly mapped to the right agents and active.
sn_aia_tool Stores tool definitions — all tools available for agents to invoke. Dev: Review tool configuration, execution mode, name, and description.

10.2 Orchestration, Teams & Trigger Tables

For: Architect · Developer —

Table Name Purpose When to Use It
sn_aia_team Defines Agentic AI teams (groups of agents) and their orchestration strategy.

Migration: Validate team-level structures are intact.

Dev: Confirm orchestration strategy (Base/ReActive/Batch Planner) is set correctly.
sn_aia_team_member Stores individual agent membership within a team. Troubleshooting: Confirm which agents are participating in a specific team's workflow.
sn_aia_trigger_configuration Stores trigger configurations for Agentic Workflows and AI Agents — type, table, conditions, run-as user.

Troubleshooting: First place to check when agent execution did (or did not) start.

Dev: Validate trigger conditions and run-as user before activation.
sn_aia_property System properties for AI Agent behaviour — including recursive check limits.

Debug Loops: Check recursive_check.update_max_executions (default 5) and create_max_executions (default 50).

Governance: Review and adjust anti-loop thresholds.

Nav: All > sn_aia_property.LIST

10.3 Runtime & Execution Tracking Tables

For: Developer · Admin

These are the primary tables for debugging live agent interactions. The debug sequence below is the order recommended by ServiceNow employees in official community documentation.

Table Name What It Reveals When to Use It
sn_aia_execution_plan The high-level plan generated for each agent request — which agents were called, in what order, and overall execution status.

STEP 1 in debug sequence. First place to check for any execution issue. Confirms system acknowledged the request and formed a plan.

Loop detection: Multiple plans for same record in short window.
sn_aia_execution_task Individual tasks and steps inside an execution plan — granular step statuses and results.

STEP 2 in debug sequence. Find exactly which step failed or stalled.

See which tool was invoked at each step.
sn_aia_tools_execution Tool-level execution logs — inputs passed to each tool and outputs returned.

STEP 3. Debug tool failures: see exact inputs the LLM passed to a tool.

Verify tool output was correct before LLM processed it.
sn_aia_message Full sequence of system messages exchanged during a conversation — orchestrator and agent messages.

STEP 4. Deep debugging: see the complete message flow.

Verify data passed correctly between agent steps.
sys_cs_message Conversation messages including all user inputs and agent responses as displayed in ServiceNow Otto panel.

STEP 5. See exact user input and agent response as surfaced in the Otto panel.

UX debugging: trace full dialogue for a session.

10.4 GenAI & LLM Log Tables

For: Admin · Architect

Table Name What It Contains Access Level When to Use It
sys_gen_ai_log_metadata A non-sensitive subset of fields from the full generative AI log. Contains model used, skill name, token counts, response metadata, and latency information. Designed for admin reporting and analytics. Admin role — accessible to platform admins. Does not contain prompt content or sensitive case data.

Auditing: Confirm which LLM model was invoked for an agent action.

Analytics: Build Now Assist usage reports by skill.

Debugging: Confirm AI was invoked when agent produces no output.
sys_generative_ai_log Full generative AI log including prompt content and response text. Contains sensitive HR, Legal, and case data. MAINT ONLY — ServiceNow internal staff only. Admins cannot access this table even after ACL modification (SN Support confirmed this is by design to protect sensitive data). For maint operations: raise a ServiceNow support (HI) ticket. Do not attempt to modify ACLs to gain access — SN Support policy.
sys_gen_ai_usage_log High-level GenAI usage events platform-wide. Confirms LLM was invoked. Used for skill usage counting and adoption reporting. Admin role — accessible.

Monitoring: Confirm LLM was actually called during agent execution.

Adoption: Count Now Assist uses per skill per time period.

Nav: All > sys_gen_ai_usage_log.LIST
sn_nowassist_skill_config Skill configuration records — maps skills to prompts, providers, and usage conditions. Admin role Skill debugging: review which prompt template and provider a skill uses.
sn_nowassist_skill_family Self-referring table categorising skills by workflow, product, and feature hierarchy. Admin role Reporting: Build Now Assist usage reports segmented by department or product.
sn_entitlement_genai_assist_counts Aggregate overview of Assist consumption — total purchased vs consumed. Admin role

Licence governance: Track Assists consumed vs entitlement.

Nav: Subscription Management > Account Level Entitlements > Now Assist Usage

10.5 AI Agent Memory Tables

For: Developer · Admin

Table Name Purpose When to Use It
sn_aia_memory Long-term memory (LTM) records — stores individual facts tied to users and memory categories. Each record has a relevance score that increases when the memory is recalled. User-specific, not session-specific.

LTM debugging: Check what was stored across sessions for a specific user.

Data governance: Review LTM retention and what is persisted.

Properties that control LTM: sn_aia.ltm.enable_long_term_memory and sn_aia.ltm.use_memory_for_ai_agent
sn_aia_ltm_category Defines the types/categories of information that can be stored in long-term memory (e.g., Device Preferences, Location). Dev: Review and configure which categories of information your agent is permitted to remember.
sn_aia_ltm_category_mapping Controls which AI Agents have access to which long-term memory categories. Governance: Restrict which agents can read or write specific memory categories for data minimisation compliance.

10.6 Now Assist AI Search — QnA Log (Search Queries & AI Responses)

For: Admin · Developer —

When a user searches and receives a Genius Result (Now Assist AI-generated answer), the interaction is logged in sn_ais_assist_qna_log. This is the primary table for reviewing what users searched, what the AI answered, and whether the answer was correct.

Table Name What It Captures When to Use It
sn_ais_assist_qna_log The user query (search utterance), the AI-generated Genius Result answer, and the source documents used to generate the answer.

Quality review: See what users searched and what AI answered.

Source investigation: Identify which

Feedback analysis: Review positive/negative feedback on Genius Results.

Nav: All > sn_ais_assist_qna_log.LIST
sys_search_signal_result_event Search queries alongside the documents users clicked (click-through signals). Shows what users settled for when searching.

Search quality analysis: Understand which documents users find useful.

Complementary to QnA log — QnA shows AI answers; this shows click behaviour.

Note on QnA log field names: The CEG AI CoE article confirms the table name sn_ais_assist_qna_log but does not enumerate all field names. Navigate to the table in your instance (All > sn_ais_assist_qna_log.LIST) to view the actual fields available in your release.

 

 

10.7 Key System Properties for AI Agents

For: Developer · Admin

Property Name What It Controls
sn_aia.continuous_tool_execution_limit Maximum limit for continuous, uninterrupted executions for the same tool within a single agent run.
recursive_check.query_for_update_record Enables recursive check on record update triggers. Default: 5 executions within 15 minutes.
recursive_check.update_max_executions Maximum number of executions before the update record recursive check triggers. Default: 5.
recursive_check.update_time_window Time window (minutes) for the update record recursive check. Default: 15 minutes.
recursive_check.query_for_create_record Enables recursive check on record create triggers. Default: 50 executions within 15 minutes.
recursive_check.create_max_executions Maximum executions before the create record recursive check triggers. Default: 50.
recursive_check.create_time_window Time window (minutes) for the create record recursive check. Default: 15 minutes.
sn_aia.ltm.enable_long_term_memory Enables long-term memory for AI Agents. All previous user interactions used as context for the LLM.
sn_aia.ltm.use_memory_for_ai_agent Enables long-term memory retrieval for individual AI Agent runs.
sn_aia_property table Navigate here to view and adjust all AI Agent system properties.

 

10.8 Quick Reference — Tables by Lifecycle Phase

For: Architect · Developer · Admin

Phase Tables to Check What to Validate
Phase 1: Development sn_aia_agent | sn_aia_agent_config | sn_aia_agent_tool_m2m | sn_aia_usecase | sn_aia_tool Agent definition saved | Active state | Tools associated | Use case mapped | Tool execution mode set correctly
Phase 2: Migration sn_aia_agent (exists?) | sn_aia_team | sn_aia_team_member | sn_aia_trigger_configuration Agent exists in target environment | Team structure intact | Agent still in correct team | Trigger conditions correct and active
Phase 3: Testing sn_aia_execution_plan | sn_aia_execution_task | sn_aia_tools_execution | sn_aia_message Plan was created | All tasks completed | Tool inputs/outputs correct | Message flow logical and complete
Phase 4: Production Monitoring sys_gen_ai_log_metadata | sys_gen_ai_usage_log | sn_aia_execution_plan (loop check) | sn_aia_property LLM invoked successfully | No execution spikes | Recursive check limits appropriate for your use case
Phase 5: Troubleshooting sn_aia_execution_plan → sn_aia_execution_task → sn_aia_tools_execution → sys_gen_ai_log_metadata → sn_aia_message → sys_cs_message Follow debug sequence in order (Section 8). Each layer reveals more detail about where the failure occurred.

10.9 Scenario-Based Troubleshooting (Verified Debug Paths)

For: Developer · Admin

Symptom Start Here Then Check Likely Fix
Agent produces no output — nothing happens

sn_aia_execution_plan

(was a plan created?)

sn_aia_trigger_configuration

(is trigger active?)

 Trigger not active or conditions not met. Check sn_aia_trigger_configuration.
Agent starts but stops mid-execution

sn_aia_execution_task

(which task has state = Error?)

sn_aia_tools_execution

(what inputs/outputs did the tool receive?)

 Tool permission error (run-as user ACL) or tool input mapping failed.
Agent fires multiple times — Assists spike

sn_aia_execution_plan

(count plans per record)

sn_aia_trigger_configuration

(filter conditions analysis)

 Apply Assignment Group or Custom Flag loop prevention pattern (Section 6).
Wrong tool called by agent

sn_aia_tools_execution

(which tool was invoked?)

sn_aia_message

(what did the orchestrator decide?)

 Tool description is too vague. Update tool name and description to be more specific.
AI Search not returning Genius Results

sn_ais_assist_qna_log

(any records for the query?)

AI Admin Hub > Analytics

(AI Search tab)

 Search profile Genius Result config not linked or not active. KB article quality issue.
Tool execution limit reached Check sn_aia.continuous_tool_execution_limit property value sn_aia_execution_task (task state) Increase limit if legitimate batch use case; reduce tool calls if design issue.

Section 11 AI Governance & Operations Framework

For: Architect · Admin — Governance framework for everyone involved in requesting, approving, building, or operating AI capabilities

How This Section Relates to the Rest of This Guide: Sections 1-10 of this guide cover HOW to build AI Agents technically. Section 11 covers WHO decides, WHEN approval is needed, and WHAT process to follow. Both must be read together — technical excellence without governance compliance is a governance failure.

11.1 Governance Principles [Proposed]

Proposed governance model — read this first. This section is my proposed operating model for governing AI agents, not an official ServiceNow position, standard, or instruction. It reflects one architect’s view of how to build governance around the platform. ServiceNow is not prescribing this model.

How to read it: subsections marked [Proposed] are my recommended approach with no official equivalent — principles, RACI, approval gates, the request lifecycle, the operating model. Everything else in this section describes documented ServiceNow platform capability. Treat proposed items as a starting point to adapt to your own policies, not as ServiceNow guidance.

For: Architect · Admin · Developer — These principles govern every AI capability — OOB, custom, or agentic

Principle Description Guardrail in This Guide
Business Value First Only approve AI developments with a tangible cost/benefit ratio. Dev effort + Assist licence consumption must be justified by measurable business benefit. Section 9 — Value Calculation Framework provides the ROI methodology
Reuse Before Build Validate ServiceNow OOB capabilities, existing VA LLM Topics, KB search, and agentic workflows before proposing new custom builds. Check AI App Library. Section 2 — OOB vs Custom Decision Rule + AI Gallery reference
Security & Compliance All AI capabilities must complete an AI Assessment prior to development and align with organisation security baselines and AI Governance policies. Section 4.4 — Security & Governance Architecture Principles
Transparency & Auditability AI logic, triggers, and decisions must be fully documented, explainable, and auditable. No black-box deployments. Section 10 — Admin Tables (all actions logged in AI Control Tower)
Sandbox First All experiments and POCs must start in Sandbox. Sandbox is never for production use cases. Section 5 — Phase 6 (Test & Validate) — sub-production always first
Continuous Improvement Deployed AI capabilities must be periodically reviewed against KPIs. Capabilities not meeting targets must be remediated or decommissioned. Section 9 — Adoption Metrics (5 standard metrics per use case)

11.2 Roles & Responsibilities

Role ServiceNow Role Key Responsibilities
AI Agent Developer sn_aia.admin Build and configure AI Agents per approved requirements. Follow lifecycle, AI Assessment, and governance guidelines. Cannot create update sets independently.
Virtual Agent Developer sn_aia.admin + virtual_agent_admin Build LLM Topics, design conversational flows, configure intents and triggers. Coordinate with Platform Team.
Platform Team / Admin admin Enable tools, monitor Assist consumption, publish metrics dashboards, enforce compliance, manage sandbox, facilitate AI Assessments, control deployments.
Product / Process Owner — Define business requirements, prioritise backlog, approve business cases, review KPIs post-deployment, authorise UAT sign-off.
ARB (Architecture Review Board) — Review and approve complex AI Agent and LLM Topic requests. Ensure architectural alignment. Enforce governance policies. See Section 11.5 for activities requiring ARB.
AI Assessment Lead — Facilitate and approve mandatory AI Assessments for all new AI Agent creations before ARB review proceeds.
Requester / Business User — Submit AI Agent creation requests through your intake process. Complete AI Use Case template. Support UAT testing.

11.3 RACI Matrix [Proposed]

For: Architect · Admin — R = Responsible | A = Accountable | C = Consulted | I = Informed

Activity / Phase Platform Architect Platform Admin AI Agent Developer Product / Process Owner Business User ARB
REQUEST & INITIATION            
Submit AI Use Case Request I I I C R/A I
Complete AI Use Case Template C I C C R/A I
Record the request I I I I R/A I
AI ASSESSMENT & DESIGN            
AI Assessment (ethics & compliance) C C I C R C
Design Review Board R/A C C C I C
ARB Review & Approval C I I C I R/A
SANDBOX & BUILD            
Provision Sandbox Environment I R/A I I I I
Conduct Sandbox POC C C R I C I
Plan and track build work C I R A I C
DEVELOPMENT            
AI Agent / LLM Topic Development C C R I I I
Update Set Management C A R I I I
Functional & Bias Testing I C R/A I I I
TESTING & DEPLOYMENT            
UAT Sign-off I I C A R I
Production Deployment C R/A C I I I
Change Record Management I R/A C I I I
MONITOR & GOVERN            
Assist Consumption Monitoring I R/A I I I I
KPI & Performance Review C R C A I I
Governance Board Review (quarterly) C C I C I R/A
Decommission Decision C C I R/A I C

11.4 Governance Process — End-to-End Request Lifecycle [Proposed]

For: Architect · Admin · Developer · Power User — All AI capability requests follow this process regardless of type or complexity

The governance process spans five phases (Plan, Design, Build, Deploy, Operate) with three parallel entry tracks. The flow below shows all decision points, approval gates, and escalation paths.

Figure: AI Governance Request Lifecycle — refer to Section 11.3 (RACI) for role assignments at each stage.

11.5 Activities Requiring Governance Approval [Proposed]

For: Architect · Admin · Developer — No development may commence on Dev without a completed AI Assessment AND appropriate approval

Activity Approval Required Complexity Determination Notes
Create / Activate / Deactivate LLM Topics Lightweight review (standard) or ARB (complex) Determined by Platform Team at intake Structural changes to existing topics require ARB
Modify existing LLM Topics Lightweight review Minor: copy/intent updates = lightweight review Structural changes require ARB review
Create / Modify AI Agents ARB mandatory All new AI Agents require ARB regardless of complexity No exceptions — AI Assessment also required
Create / Modify Agentic Workflows ARB mandatory All agentic workflows (multi-step, tool calls, orchestration) Includes trigger design — recursion risk must be assessed
Modify AI Search Profiles or Sources Lightweight review or ARB Changes to indexed sources reviewed for data governance Coordinate with Platform Team for re-indexing impact
Decommission AI Capability Lightweight review or ARB Must be documented with business rationale KPI underperformance is a valid decommission trigger

11.6 AI Use Case Template — Required for All Requests [Proposed]

For: Power User · Developer · Admin — All requests must include a completed AI Use Case document recorded with the request

Field Description & Guidance Example
Module ServiceNow module(s) in scope ITSM / SPM / CMDB / HR / CSM
Problem Statement Clear articulation of the business problem being solved Agents spend 12 min avg documenting resolution notes per incident
Concrete Example A real-world anonymised example illustrating the problem Agent resolves password reset but must manually type 3-paragraph resolution note
Steps & Tasks Process, data sources, systems, user interactions involved Who does what, when, with which systems — full process map
Conversational Use Case? Yes/No — does it require Virtual Agent / chat? Yes = LLM Topic route | No = AI Agent / Skill route
Business Justification Quantified expected benefits FTE hours saved, SLA improvement, deflection rate increase, cost avoidance
Target Persona Who uses this AI capability End User / Fulfiller / Analyst / CMDB Admin / Manager
AI Capability Type Specify the type of AI capability proposed LLM Topic / AI Agent / Agentic Workflow / AI Search / OOB GenAI Skill
OOB Validation Confirm AI App Library checked and no suitable OOB solution found AI Gallery URL: applibrary.service-now.com/ai_gallery
Est. Assist Consumption Estimated Assist units per transaction (see Section 11.8) Small workflow: 25 | Medium: 50 | Large: 150 | Summarisation: 1
Success Metrics Quantitative and qualitative KPIs Confidence score >0.80, deflection rate +10%, adoption >60% MAU, CSAT
Target Go-Live Expected deployment date and pilot scope Target quarter — pilot with agreed team scope before full rollout

11.7 Operational Model — Three-Tier Structure [Proposed]

Tier Owner Operational Responsibilities
Platform Governance Platform Team Assist licence management, environment provisioning (Sandbox/Dev/Test/Prod), governance policy enforcement, audit readiness, ServiceNow release upgrades, access control.
AI Development & Delivery Developers + Product Owners AI Agent and LLM Topic development, update set management, training verification, testing, deployment coordination, documentation maintenance.
Business Operations Product/Process Owners + Business Users Business case ownership, KPI monitoring, UAT, feedback collection, escalation of performance issues, prioritisation of backlog.

Instance Strategy

Environment Purpose Rules & Constraints
Sandbox POC & Experimentation Mandatory for all new POCs. Not for production configurations. Issues are non-urgent. Must be scoped and agreed with Platform Team before use. No formal Product/Process Owner approval required.
Development Active Development All new development starts here. Update sets must be created. No direct production promotion permitted. ACT scan before promotion.
Test QA & UAT Functional testing, bias testing, and UAT performed here. UAT sign-off required before promotion to production.
Production Live Operations Promotion only via approved update sets and change management. Platform Team controls all deployments. No direct production modifications.

Governance Review Cadence

Review Type Frequency Participants & Focus
Operational Review Weekly Platform Team: Assist consumption, active incidents, deployment queue, loop alerts.
KPI & Performance Review Monthly Platform Team + Product Owners: deployed Agent KPIs vs targets, feedback trends, low-confidence Genius Results.
Governance Board Review Quarterly ARB + Platform Team + Stakeholders: policy updates, backlog prioritisation, compliance audit, licence review.
Annual Strategy Review Annually Leadership + Platform Team: AI roadmap alignment, capability decommission decisions, budget planning.

Proactive Monitoring by Category

Category Metrics Tracked Review Cadence & Action
OOB GenAI Skills Usage counts, user satisfaction ratings, error rates per skill Monthly. Underperforming skills flagged to Product Owner for remediation.
AI Search / Genius Results Confidence scores, generative answer quality (sn_ais_assist_qna_log), source citation accuracy Bi-weekly. Low-confidence results trigger prompt and KB article review.
AI Agent Responses Task completion rates, Assist consumption per transaction (AI Agent Analytics), fallback rates Weekly. Agents consuming >2x estimated Assists flagged for architecture review.
Trigger & Loop Monitoring Execution plan counts per record, Assists spike detection Daily automated alert. Recursive trigger loops trigger immediate deactivation.

11.8 Assist Consumption Reference

For: Architect · Admin · Product Owner — Confirm current rates via AI Admin Hub > Analytics > Assist Consumption

Category Skill / Capability Assists Rationale
Core Agent Productivity Incident Summarisation 1 Simple summarisation — single LLM call
Core Agent Productivity Resolution Notes Generation 1 Single AI action
Core Agent Productivity Chat Summarisation (VA & Live Agent) 1 Standard summarisation
Core Agent Productivity Knowledge Article Generation 10 Complex multi-step generation
Core Agent Productivity Email Reply Recommendation 5 Quick AI response suggestion
Core Agent Productivity Case Summarisation 1 Simple summarisation
Advanced & Strategic Change Risk Explanation 5–10 Analyses multiple data points
Advanced & Strategic Incident Assist (Panel Q&A) 5–10 Multi-turn contextual Q&A
Advanced & Strategic AI Search with Generative Answers 5–10 Multi-step semantic search
Advanced & Strategic Suggested Next Steps 5–10 Multiple internal signals processed
Advanced & Strategic ServiceNow Otto panel Conversations 5 Standard panel interaction
Agentic Workflows Small Agentic Workflow 25 Simple or low-complexity — 0 to 4 tools per execution
Agentic Workflows Medium Agentic Workflow 50 Moderate complexity — 5 to 8 tools per execution
Agentic Workflows Large Agentic Workflow 150 High-complexity, multi-step — 9 to 20 tools per execution

Cost Governance Rule: Product/Process Owners must acknowledge and approve the estimated running Assist cost BEFORE ARB approval is granted. Agents consuming more than 2x the estimated Assists in production will be automatically flagged for architecture review.

11.9 LLM Topics vs AI Agents — When to Use Which

Dimension LLM Topic (Virtual Agent) AI Agent (Agentic Workflow)
Use Case Type Conversational, single-turn or guided dialogue Multi-step reasoning, autonomous task execution
User Interaction Chat-based, user-driven Background processing, minimal user interaction
Complexity Low to medium — predefined intents and flows Medium to high — dynamic decision-making and tool calls
Assist Consumption 1–10 Assists per interaction 25–150 Assists per workflow
Governance Approval Lightweight review (standard) or ARB (complex) ARB mandatory for all new AI Agents
AI Assessment Required? Only if using external data sources or triggering agentic workflows Always required before development commences
Example Use Cases Password reset, FAQ, incident status enquiry, KB search Incident resolution with CMDB lookup + knowledge search + ticket update + notification

Zurich P4+ Governance Addition: MCP Server registrations and external agent connectivity via A2A must be reviewed and approved through AI Control Tower. Assign clear ownership for approval/rejection of MCP server registrations and skill modifications as part of your AI governance model.

11.10 AI Development Best Practices [Proposed]

Prompt Engineering Standards

All AI prompts (agent roles, instructions, LLM Topic flows) must follow the RCOE format:

Component What to Define Example
Role Define the AI's persona and domain identity You are a ServiceNow ITSM analyst specialising in incident triage for [Organisation] IT.
Context Provide relevant background — what data is available, what constraints exist You have access to the incident record, the caller's assigned devices from CMDB, and the knowledge base.
Output Format Specify the expected structure Respond with: (1) Category, (2) Subcategory, (3) Suggested CI, (4) Confidence level. Use plain English, no technical jargon.
Examples Include few-shot examples where available Example: If description mentions "VPN", category = Network, subcategory = Remote Access.
Constraints Define tone, length, language, scope boundaries Always respond in English. Maximum 3 sentences. Do not suggest actions outside the ITSM module. If unsure, say so.

Design & Architecture Rules

Use Agentic AI Workflows for multi-step reasoning and orchestration — not LLM Topics.

Build modular, reusable agents — avoid monolithic agents doing everything (see Section 4.1 — max 10 agents per use case, 5 tools per agent).

Separate high-risk and low-risk tasks — pilot with low-risk, high-volume use cases first.

Apply loop prevention patterns before any record-based trigger goes live (see Section 6 — Assignment Group Pattern).

Validate against the ServiceNow AI App Library before building any custom capability.

Keep architecture modular — LLM model may need to change per use case (NowLLM vs third-party).

Security & Compliance Rules

Enforce role-based access for all AI capabilities — principle of least privilege for run-as users.

Maintain audit trails via AI Control Tower — all actions logged and auditable.

Conduct periodic fairness and accuracy validation — minimum quarterly KPI review.

All customisations must be within defined application scope — no global scope customisations.

Update sets are mandatory for all Dev → Test → Prod migrations. No direct production modifications.

11.11 Compliance, Audit & AI Assessment

Audit Trail Requirements

All AI Agent creation, modification, and decommission activities must be documented and traceable.

AI assessment records must be retained.

ARB approval decisions must be documented in ARB meeting minutes.

Deployment records must be retained per your change management process.

AI Assessment Process

Step Activity Owner
1 Requester notifies Platform Team when submitting the request Requester
2 Platform Team coordinates with AI Assessment Lead Platform Team
3 Requester completes the AI Assessment questionnaire (invitation sent automatically) Requester
4 AI Assessment Lead reviews and approves or requests clarification AI Assessment Lead
5 Approved AI assessment is recorded before ARB review proceeds AI Assessment Lead / Requester

11.12 AI Control Tower: Platform-Native Governance

For: Architect · Admin · Developer — The control plane every other part of Section 11 operates through

AI Control Tower (AICT) is the platform-native governance layer for every AI asset on the instance: agents, models, MCP servers, prompts, datasets, and skills. As of the Australia release it is included in every Now Assist and AI Native SKU and is no longer a separate purchase. Treat it as the control plane, not a dashboard.

The single most important change in Australia: governance moved from advisory to enforcing. AI Stewards can require formal approval before an MCP server is activated for use in any agent builder application, and unapproved servers are not visible to agent builders. The control is enforced in the tooling, not documented in policy. Your ARB decisions now have technical teeth. Design your approval process accordingly.

The Five Dimensions

Dimension What it does out of the box Where it maps in this guide
Discover Auto-discovers AI assets into the AI Asset Inventory. Beyond the ServiceNow platform it queries 30 third-party systems including AWS, Google Cloud, Microsoft Azure, SAP, Oracle, and Workday. Service Graph Connectors now cover Databricks, Snowflake, Hugging Face, and Salesforce. Inventory referenced in 11.1 and 11.6
Observe Continuous runtime monitoring with live metrics, alerts, and agent reasoning traces. Replaces periodic manual audits. Operational model 11.7; debug Section 8
Govern Risk assessment across agents, models, MCP servers, datasets, and prompts, with pre-built regulatory content. Risk assessment 11.11; approvals 11.5
Secure Veza access-graph technology extending identity access governance to hyperscaler AI environments, with scoped permissions, least-privilege enforcement, and auditable identity chains for every agent, model, and action. Plus a real-time kill switch via the AI Gateway. Roles 11.2 and 11.3; security 4.4
Measure Cost tracking across model providers and ROI dashboards tying spend to outcomes, with per-team and per-BU Assist quotas. Value management 9.5; assists control 9.6

What AICT Now Blocks, Not Just Flags

This is new and it changes your build process. Brief your developers before they hit it as a surprise.

Control Behaviour
MCP server approval mandate When enabled by an AI Steward, dropdowns in AI Agent Studio show only approved MCP servers. Unapproved, rejected, and paused servers are hidden from selection.
MCP PII detection AI Stewards can activate PII detection for any MCP server with a single toggle. Every call is scanned, and if sensitive data is detected AI Gateway blocks the entire payload before it reaches the requesting agent. No code changes are required from agent builders.
BYOK provider governance Data routing controls and approval enforcement at the Skill Kit level ensure only approved providers and models are available to skill and agent builders.
Managed vs Unmanaged scope Only Managed assets have governance workflows, risk assessments, value measurement, and monitoring enabled. Only AI Stewards can move assets between states. Historical data and audit trail are preserved when an asset leaves active governance.

Governance Rule: Decide who holds AI Steward before you enable the MCP approval mandate. Once enforced, an unapproved server is invisible to builders, and a developer with no visibility into why will raise a defect rather than a governance request.

Runtime Security Monitoring

Two distinct capabilities sit in the Security and Privacy tab. Agentic Threat Monitoring covers agent deviation detection, flagging when an agent strays from its authorised role or constraints including prompt injection attempts, role boundary breaches, and override attempts; and output screening, which scans agent outputs deterministically for PII leakage and embedded vulnerabilities including XSS, SQL injection, and remote code execution, without additional LLM calls. Data Model Integrity is separate and continuously scans LLM outputs across user, agent, and skill invocations against enterprise policy rules, detecting harmful content, jailbreaks, and refusal quality issues.

Each detection carries explicit evidence and reasoning for investigation, audit, and regulatory review. Sampling, thresholds, and coverage are configurable. Use these in your monthly KPI review in 11.7.

Risk and Compliance

Risk assessment questions are embedded in the AI asset intake form. High-risk or non-compliant submissions route automatically to the AI Steward before onboarding, so assets carry a High / Medium / Low classification from day one based on the asset’s use and purpose. This replaces part of the manual triage in your 11.6 template.

Pre-built compliance content packs are available for the EU AI Act, California AI Act, and Colorado AI Act, each including regulations, citations, control objectives, and risk statements. Cross-framework mapping means a single control can satisfy the California AI Act, Colorado AI Act, EU AI Act, and NIST AI RMF simultaneously. ISO 42001 content is in legal review and will follow in a later release.

Do Not Build What Ships: If your organisation is standing up EU AI Act controls in a spreadsheet, stop and activate the content pack first. Test a control once, satisfy multiple frameworks.

AI Asset Lifecycle Through AICT

  • Intake: every initiative starts as an AI Use Case, the business-facing charter and parent record for all assets built under it. This is the 11.6 template, captured as a record.

  • Assess: the AI Accountable Officer assigns risk, impact, and data-sensitivity assessments and designates an AI Asset Owner. Maps to 11.11 and the gates in 11.5.

  • Build and test: development happens outside AICT (Section 5), but evidence and approvals are enforced within it before anything ships. Agentic Evaluation results roll up here.

  • Deploy, monitor, measure: Observe and Measure take over once live. Value tracking in 9.5.

  • Decommission: assets are formally offboarded through a governed workflow, and a structured Use and Purpose field captures why an asset exists to support accurate risk classification. Audit trail is preserved.

Governance Roles AICT Enforces

Role Owns
AI Steward Day-to-day governance across the inventory. Sole authority over Managed / Unmanaged state and the MCP approval mandate.
AI Accountable Officer Assigns assessments, designates asset owners, owns the risk posture.
AI Asset Owner Accountable for a specific asset including its value realisation (9.5).
Risk and Compliance Analyst Compliance content packs, control testing, evidence for audit.

Map these onto 11.2 and 11.3. Reported OOB role identifiers are sn_ai_governance.ai_steward and sn_ai_asset_mgmt.ai_asset_owner. Confirm the exact role names in your own instance before wiring them into assignment rules. This guide does not enumerate AICT table or field names; open the workspace and read the live schema rather than hardcoding from a document.

Other Capabilities Worth Knowing

  • Anonymous reporting of AI misuse or incidents, supporting whistleblowing requirements.

  • Automated Rules that move matching assets into Managed state on a schedule. Up to 10 rules, 5 active at a time, three shipped out of the box.

  • Publishing a managed ServiceNow agent to the Microsoft Agent 365 directory directly from the asset record, so the agent is manageable in both registries without duplicate registration.

  • AI case management for logging and managing AI cases and inquiries to address issues.

Availability note: the five-dimension AICT shipped with the Australia release, and a further set of Discover, Govern, and Secure capabilities reached GA in the June 2026 release. Do not treat any of the above as roadmap. Verify which features are enabled in your instance, since availability depends on store app version and activation, not release alone.

Honest limit: AICT governs what it can see. Assets built outside the discovered estate, shadow AI on unconnected platforms, and anything a team declines to register remain invisible. Discovery coverage is a programme commitment, not a switch.

11.13 Data Privacy and Sensitive Data Handling

For: Architect · Admin · Developer · Power User — Required before any AI capability touches personal, HR, health, or commercially sensitive data

Every Now Assist and AI Agent interaction sends instance data to a language model for inference. Even where processing is transient, that is a data transfer, and it is the first question a privacy or security reviewer will ask. Sections 1 to 11.12 cover whether the agent works. This covers whether it is allowed to.

Three Layers of Control

Layer What it is Who acts
Platform default Zero persistence. Prompts and responses are processed in memory and are not stored in ServiceNow data centres; temporary data is discarded when the interaction concludes. Input and output are transient, not cached on the compute hub, and not commingled with other customer data. Domain-separated data is also not commingled. Nothing to configure. Know it for the DPO conversation.
Configured control Data Privacy for Now Assist masks sensitive data in generative AI use cases using real-time anonymisation, replacing the previous Sensitive Data Handler solution. It is integrated into the Generative AI Controller and works with all LLMs. Masking is two-way: placeholder or anonymised values are sent to the LLM, then replaced with the original values after the response returns. Platform Team configures. Must be configured. It is not on by default.
Commercial decision The data sharing programme for model improvement, which customers can opt out of from the AI Admin Hub or AI Control Tower. Even when opted in, only prompts and responses already masked and anonymised by Data Privacy for Now Assist are used. Privacy team decides. Platform Team executes.

Data Sharing Opt-Out Procedure

Navigate to AI Admin Hub > Settings > Data sharing and processing > Data sharing. This is per instance and requires the sn_generative_ai.data_steward role. ServiceNow documentation states opt-out can take up to five business days. For GCC and self-hosted instances data sharing is not available, so it is not enabled.

Governance Gate: Confirm the data sharing position in writing before the first production AI use case goes live, not after. Record the decision and the date against the AI Use Case record (11.6). This is a question you will be asked in every AI audit and every customer security review.

Known Limitations

  • Masking in Data Privacy for Now Assist is regex and data-pattern based, not contextual AI-based, and it must be configured.

  • Pattern-based detection can only find what you have explicitly defined a pattern for. Names, freehand addresses, and organisation references in unstructured text are missed, and attachments have historically been a blind spot. A Data Privacy or ServiceNow Vault subscription extends beyond this into broader discovery and protection.

Section 12 Key Documentation

  • AI Native Packaging and Tiers Overview — servicenow.com/docs/r/intelligent-experiences/ai-native-sku-overview.html

  • AI Admin Center Release Notes — servicenow.com/docs/r/release-notes/now-assist-center-rn.html

  • AI Control Tower product page — servicenow.com/products/ai-control-tower.html

  • AI Control Tower (ServiceNow Store) — store.servicenow.com/store/app/ce9dc5b01b192e50a85b16db234bcb4b

↑ Back to navigation index

Document Control

Field Value
Document Version (Enriched) 2.0 (Australia release refresh, August 2026)
Prepared By Varun Gupta
Effective Date August 2026
Applicable Releases ServiceNow Zurich · Australia (GA 5 May 2026, current). Australia Patch 4 released 9 July 2026.
Required Entitlement AI Native tiers: Foundation, Advanced, or Prime (Now Assist, Moveworks, Workflow Data Fabric, and AI Control Tower bundled in all three). Legacy Pro Plus and Enterprise Plus reached end of sale on 1 July 2026. Custom AI skill and agent creation requires Prime.
Store App Versions Now Assist AI Agents store app v6.1.11 (latest confirmed, March 2026) │ Loop Prevention: sn_aia v4.0.38+ │ Role Masking: store app v6.0+ │ AI Control Tower bundled in all AI Native tiers; the five-dimension AICT shipped with Australia and further capabilities reached GA in the June 2026 release. Verify store app versions in your instance.
Review Frequency Each major ServiceNow release (bi-annual)

 

Preview file
1261 KB
  • 30 Views
Version history
Last update:
yesterday
Updated by:
ServiceNow Employee
Contributors