Three Types of AI — What's the Difference?
ServiceNow since 2017
Artificial Intelligence (AI)
Broad umbrella covering all machine intelligence. Enables intelligent experiences, automation, and optimisation. Analyses existing data to inform decisions — does not create new content.
Now Assist · Current releases
Generative AI (GenAI)
Specialised branch that creates original, realistic content based on training patterns. Uses neural networks to generate text, code, and summaries that didn't exist in training data.
Agentic AI · Latest releases
Agentic AI
Autonomously sets goals, analyses data, and takes initiative to execute tasks end-to-end without waiting for user input. Plans, decides, and acts — not just responds.
ServiceNow AI Value Pillars
Intelligent Experiences
- Virtual Agent chatbot for self-service
- AI Search for fast knowledge retrieval
- NLU-powered intent recognition
- Personalised, contextual responses
Intelligent Automation
- Routine task automation to cut errors
- Predictive routing and categorisation
- Agentic end-to-end workflow execution
- Frees agents for higher-value work
Intelligent Optimisation
- Real-time visibility and trend insights
- Anticipate issues before they escalate
- Rapid, data-backed decision making
- Continuous service improvement signals
💡 Key framing for clients: AI, GenAI, and Agentic AI are not competing choices — they are complementary layers. AI underpins the platform since 2017. GenAI generates content. Agentic AI orchestrates execution. Position the full stack as "return on intelligence" — Bill McDermott's framing.
Language Intelligence
NLU — Natural Language Understanding
NLU Inference & Workbench
Understands word meanings and context to infer intent and entities from user utterances. Enables machines to respond to natural language. Supports 17 languages as of Zurich release.
Virtual Agent
Issue Auto Resolution
AI Search
NLQ — Natural Language Query
NLQ Engine & Genius Results
Translates plain-language requests into structured Glide record queries (JSON or visual). Active by default. Surfaces records across tables without requiring knowledge of table structure.
Analytics Q&A
CMDB Query Builder
Global Search
List V2 Filter
Predictive Intelligence (PI)
Machine Learning Frameworks
Three PI Frameworks
Uses ML algorithms to learn from contextual data, understand patterns, and predict outcomes without manual intervention. Drives automation and continual service improvement.
ITSM
CSM
HRSD
Classification
Sets field values during record creation. Auto-categorise incidents, route cases to correct assignment groups based on short description.
Similarity
Surfaces similar tasks and content. Expose Similar Resolved Incidents to agents helping resolve current tickets more efficiently.
Clustering
Reveals patterns not obvious via traditional reporting. Group records by attributes — e.g. highlight commonalities in failed change requests.
Document & Task Intelligence
Document Intelligence — DocIntel
Automated Document Extraction
Goes beyond OCR to extract, validate, and integrate structured data from JPEG, PNG, and PDF documents into automation workflows. Processes single and multi-page typed documents — forms, invoices, ID documents.
com.snc.docintel_admin
Invoice Processing
Forms
Task Intelligence — TI
ML-Driven Task Workflows
Uses ML and platform data to automate creation, triage, and investigation of tasks. Tracks impact across creation, deflection, triage, remediation, and optimisation moments. Reduces MTTR.
sn_ti_admin
ITSM
CSM
🏗️ Architect tip — NLU vs NLQ vs Now Assist: NLU handles intent discovery (what does the user want to do). NLQ handles data retrieval (query tables in plain language). Now Assist LLM topics replace NLU for VA conversations, offering better accuracy with less configuration. Position each based on use case complexity and available configuration resource.
Now Assist Skills by Audience
👩💻 Developers
Intelligent code generation, workflow suggestions, and agentic multi-step development task orchestration to accelerate app delivery.
⚙️ Admins
Improve app delivery, reduce maintenance overhead, and accelerate time to value with AI-assisted configuration and monitoring.
🎧 Customer Support
Case summarisation, recommended actions, and contextual knowledge surfacing — enabling agents to deliver faster, more personalised service.
🔧 Agents
AI-driven overviews, insights, and work notes that reduce manual effort and allow focus on high-complexity resolutions.
👤 Employees
Natural language self-service, intelligent search, and multi-turn conversational catalogue ordering — get what you need, when you need it.
📈 ITSM / CSM / SPM / ITOM
Domain-specific skills growing with every release. Each skill activatable and configurable through the Now Assist Admin Console.
Large Language Model (LLM) Options
Proprietary
Now LLM
ServiceNow's own LLM based on StarCoder2, co-developed with Hugging Face and Nvidia. Purpose-built for the Now Platform via the Now LLM Service.
Third-Party
External LLMs
Bring your own GenAI model — GPT-4, Google Gemini Pro, and other supported providers. Configured via the Generative AI Controller.
Custom
Customer LLMs
Organisations can develop and host their own fine-tuned LLMs and integrate them into the Now Assist framework to tailor AI behaviour.
Generative AI Controller
→
ServiceNow's intelligence connection layer between the Now Platform and all LLMs (proprietary or external)→
Foundation for all Generative AI products, and available for custom Gen AI applications and workflows→
Accessible directly in Flow Designer, Virtual Agent Designer, background scripts, and business rules→
Configured and monitored through the Now Assist Admin Console — single pane of glass for activation, config, and monitoringAvailability restriction: Some Now Assist products/features are currently unavailable for FedRAMP, NSC DOD IL5, Australia IRAP-Protected environments, and self-hosted customers. Always verify availability via KB0743854 on Now Support before recommending.
Agentic AI Architecture Components
AI Agent Orchestrator
Central coordinator that ensures multiple AI agents work together efficiently. Routes requests, manages dependencies, facilitates collaboration, and resolves multi-step processes requiring input from several agents or systems.
AI Agent Fabric & Integration
Connects ServiceNow's native and third-party AI agents, tools, and data sources. Enables seamless integration across platforms. Supports interoperability to extend automation beyond ServiceNow via APIs and connectors.
Channels & Data Flow
Environments where AI agents interact with users and systems (Now Assist panel, employee apps). Data flow manages real-time information movement between agents, workflows, and external systems for timely, accurate, secure outcomes.
AI Agent Studio
Central platform for creating, managing, and optimising agentic workflows and AI agents. Provides templates, agent configuration, testing environment, Now Assist Guardian settings, and analytics dashboard in a unified interface.
Operational Flow — Four Stages
1
Trigger: Specific events initiate workflows — user actions, system events (incident creation, data updates), or scheduled processes. Configurable to respond to any business scenario.2
Execution: The agentic workflow coordinates one or more AI agents through defined action sequences, decision points, and interactions managed by the AI Agent Orchestrator.3
Orchestration: Multiple agents collaborate, share information, and delegate subtasks. The orchestration layer handles exceptions, escalation, and requests for additional data as needed.4
Integration: Workflows connect with external enterprise systems via APIs, connectors, and the Workflow Data Fabric for end-to-end automation across the broader IT ecosystem.Out-of-Box Agentic Workflow Use Cases
ITSM
Generate Resolution Plans
Fetches record details, generates resolution steps, and updates work notes automatically for incidents, cases, and tasks.
CSM
Analyse & Improve Services
Analyses customer feedback and derives service metrics to identify improvement opportunities with actionable recommendations.
CMDB
CMDB Governance
Validates lifecycle policy management, enforces data certification standards, and monitors health metrics across CI classes.
Asset
Hardware Asset Requests
Analyses and fulfils hardware requests — consuming local stock, transferring from stockrooms, or procuring from vendor automatically.
HRSD
HR Case Response
Auto-answers non-critical HR cases using existing knowledge articles and catalogue items, minimising agent intervention.
🤖 Agentic AI vs. Traditional AI — key client message: Traditional AI agents follow predefined rules and require external commands. Agentic AI sets its own objectives, breaks them into tasks, and adapts strategies as new data emerges. This is the difference between automation that waits and automation that acts.
Topic Discovery — Three Approaches
K
Keywords: Simple trigger-word matching. Fast to configure, low accuracy. Use only for highly predictable, low-variation inputs.N
NLU: Intent and entity recognition from natural language utterances. Precise and controllable. Requires training data, ongoing tuning, and an NLU admin role. Supports 17 languages.L
LLM (Now Assist in VA): Large Language Model-based intent discovery. Low-code setup, high flexibility. Configured in minutes from the Admin Console. Cannot run on the same portal as NLU.
💡 Architect decision: NLU offers control and precision. LLM offers flexibility and scalability. Many organisations benefit from a hybrid approach — NLU for complex structured flows, LLM for knowledge search and catalogue. Recommend Now Assist in VA for greenfield deployments with a strong knowledge base.
Key Roles Required
Business Analyst
- Maps business processes to VA intents
- Collects and curates utterance datasets
- Native speaker awareness & dialect sensitivity
NLU Administrator
- Builds, trains, tunes NLU models
- Monitors prediction performance over time
- Controls model release lifecycle
VA Administrator
- Develops, tests, and maintains conversations
- JavaScript skills for moderate-complex topics
- ServiceNow VA Fundamentals + Implementer certified
Implementation Lifecycle
Initiate
Readiness checklist · Project kickoff
Plan
Training · OOB topic review · Top use case identification · Channel & branding decisions
Execute
Activate plugins · Configure branding · Build & test VA flows · Train NLU model iteratively
Deliver
UAT · Comms plan · Re-train NLU · Publish conversations · Go-live
Close
Support transition · Project close-down · Hypercare period
Monitor
Review deflection & adoption · NLU performance · Identify next automation opportunities
Success Metrics That Matter
Cost Savings
- Cost per contact by channel
- Deflection rate — issues resolved by VA
- Self-service rate increase
- $ savings from deflected T2 incidents
Customer Satisfaction
- CSAT — Customer Satisfaction Score
- CES — Customer Effort Score
- NPS — Net Promoter Score
- Transfer-to-live-agent rate (trending down)
NLU Performance
- Utterance miss rate
- Intent prediction accuracy
- Conversation completion vs. abandonment
- Repeat user and active user growth
Release Strategy: Use a dedicated scoped application for VA topics. Never build in production. Use the ServiceNow Application Repository or Git over Update Sets for VA deployment. Separate scoped apps per business domain (ITSM, HR, CSM) if teams have independent release cycles.
Four Responsible AI Principles
Human-Centered
Clear guidance on AI usage. Customer choice empowered — identifies when AI is used and provides responsible deployment guidelines across all products.
Transparency
Open communication through product documentation and model cards. Clarity about how AI models work and an honest acknowledgement of their limitations.
Diversity & Bias
AI teams and training datasets reflect diversity. Continuous testing to understand demographic impacts and control for undesired bias in predictions and outputs.
Accountability
Trust is central. External and internal oversight structures are in place for governance, ensuring customers can rely on AI capabilities with confidence.
Now Assist Guardian
🛡️
Configurable via AI Agent Studio → Settings. Applies safety guardrails to all AI agents and agentic workflows in your instance.🚫
Offensiveness detection — flags and blocks harmful or inappropriate AI-generated content before it surfaces to users or agents.💉
Prompt injection attempt detection — identifies and rejects adversarial inputs designed to manipulate AI agent behaviour.🧠
Long-term memory controls — governs what context AI agents retain across sessions, balancing personalisation with privacy requirements.🔧
LLM selection governance — manage which LLMs are accessible to agents, ensuring organisational AI policy is enforced at the platform layer.
🌐 CTA positioning: Responsible AI is not a feature checkbox — it is a governance conversation you must have with every customer before deploying Now Assist or Agentic AI. Address model transparency, data residency, bias testing, and oversight structures as part of the architecture design, not as an afterthought.
Takeaway 01
AI Layers Are Complementary, Not Competing
AI, GenAI, and Agentic AI each play distinct roles — AI powers intelligent automation, GenAI creates new content, and Agentic AI autonomously plans and executes tasks. Advise customers on the right layer for each use case rather than positioning them as alternatives.
Takeaway 02
Match Capability to Business Outcome
NLU, NLQ, Predictive Intelligence, Document Intelligence, and Task Intelligence each target specific problem types. Identifying the right capability requires understanding the process bottleneck first — then selecting the AI tool that removes it.
Takeaway 03
Responsible AI Governance Underpins All of It
Human-centred, transparent, diversity-aware, and accountable AI principles must be embedded from the start — not added later. Use Now Assist Guardian and the governance conversation as part of every AI implementation design session.