Generative AI (Now Assist), AI Agents (autonomous), and Virtual Agent (chatbot) — designed, integrated, and governed for the enterprise. We deliver the full AI stack, from prompt engineering to production rollout.
Now Assist is ServiceNow's Generative AI platform — a family of LLM-powered features embedded across every product (ITSM, HR, CSM, SecOps, Creator). It runs on ServiceNow's own foundation models, with fallbacks to Azure OpenAI. As of the Xanadu / Yokohama release wave, Now Assist AI Agents also enable autonomous, multi-step workflow execution.
Virtual Agent is the chatbot front-end that leverages Now Assist under the hood. It handles both structured topic flows (NLU-based) and open-ended generative responses (LLM-based) — deployable across Teams, Slack, Employee Center, the customer portal, and mobile.
The hard part isn't turning Now Assist on — it's designing it well. Prompt engineering, guardrails, data protection, cost control, adoption strategy, and human-in-the-loop gates are what separate a real AI deployment from a demo that fails in production.
VanPaulTek has delivered Now Assist across ITSM, HRSD, and CSM, and built Virtual Agent chatbots across Teams and Employee Center. We know where the platform shines and where you need to slow down.
Every AI capability ServiceNow ships — plus custom skills via Skill Kit — designed, deployed, and governed by one team.
Generative AI features embedded in the incident, problem, and change workflows.
AI features that reduce HR agent load and improve employee experience in Employee Center.
Customer-service AI — case + agent + article generation for B2B and B2C support.
AI for security incident triage, response, and playbook execution.
Developer-facing AI — generates Business Rules, Script Includes, UI Actions, Flow Designer steps.
Autonomous, multi-step AI agents that plan and execute across ServiceNow workflows — the newest wave (2024–2025).
The customization layer — build your own Now Assist skills with prompt engineering + retrieval + guardrails.
Multi-channel chatbot combining NLU topics with Now Assist LLM answers — the primary self-service surface.
Design, architect, develop, implement, and support — five phases, one accountable team.
AI + chatbot strategy — the biggest failure point if skipped.
The technical + governance architecture that makes AI trustworthy.
Building AI skills + chatbot topics with real engineering discipline.
Rollout must be phased — AI in production has unique risks.
AI + chatbot are living systems — support is ongoing tuning.
Sample roadmap based on real implementations — adjustable to your scope, but grounded in what actually works. Not vendor marketing timelines.
Practical fixes that don't need a project charter. Ordered by timeframe and impact — the stuff experienced practitioners just do.
One config toggle. Fulfillers get a paragraph summary of long incidents in seconds. Adoption is nearly instant; MTTR drops 5-10%.
Employee Center gets AI-powered semantic search over knowledge. Users find answers 2-3x faster. Deflection lifts immediately.
Ship VA to EC with top-10 topics (PW reset, PTO balance, benefits FAQ, IT ticket). Deflects 20-30% of common inquiries.
Regex + LLM guardrail catches SSN, credit cards, names in outputs. Non-negotiable for compliance. Do this before scale.
Now Assist drafts resolution note based on case activity. Agent reviews + edits. AHT drops 20%+ within 30 days.
Deploy VA to MS Teams for IT ticket creation + HR FAQ. Where users already work = adoption. Immediate deflection.
Now Assist scores every change's risk based on history + CMDB context. CAB spends time on high-risk, not standard.
After N similar resolved cases, Now Assist drafts a KB article. Human review, publish. Knowledge grows organically.
RAG-based policy lookup with citations. Reduces 'what's our policy on X?' questions to HR/legal by 40%+.
When VA hands off to live agent, LLM summarizes the conversation as the case description. Agent starts informed.
Real KPIs and targets from mature implementations. Track these; if they trend the wrong way, something is off.
% of AI suggestions accepted by users (as-is or with minor edits). Below 40% = tuning needed.
% of VA conversations resolved without live-agent handoff. Requires topic + LLM tuning.
% of AI outputs judged correct by SMEs. Below 85% = prompt or RAG tuning needed.
PII leaked in any AI output. Zero-tolerance — investigate any incident. Requires guardrails.
Average Now Assist cost per interaction. Above $0.10 = review model routing + caching.
% of chatbot inputs correctly matched to intent. Below 75% = NLU tuning or topic scope issue.
% of VA sessions escalating to human. Below 20% = VA over-confident; above 50% = topics inadequate.
% of AI-generated content (articles, responses) used as-is or with minor edits. Adoption signal.
Honest warnings from many deliveries — the mistakes that cost time, money, and adoption. These aren't in vendor guides.
Why it fails: AI cost + governance + accuracy all break under simultaneous rollout. First skill fails, tars everything.
Do this instead: One skill, one user group, review gates on. Prove value + tune, then expand. Sequential, not parallel.
Why it fails: LLMs hallucinate. A confidently-wrong response to a customer = trust destroyed + PR risk.
Do this instead: Human review gate on all customer-facing text. Selectively auto-post once accuracy is proven (95%+ on category).
Why it fails: Sending customer PII into LLM prompts = compliance violation (GDPR, CCPA, HIPAA). Real regulatory exposure.
Do this instead: PII redaction guardrail on prompt inputs + outputs. Auditable. Non-negotiable — do this before any launch.
Why it fails: Per-request licensing scales with usage. First month bill = 3x forecast is common if unmonitored.
Do this instead: Cost dashboards + usage caps + model routing to cheaper models where accuracy allows. Weekly review.
Why it fails: VA hits its limit + user is stuck. Frustration = abandonment + brand damage.
Do this instead: Handoff to live agent as first-class capability. Full context transfer. User rarely stuck > 3 VA turns.
Why it fails: 50+ topics = NLU accuracy drops + user confusion. Chatbot becomes worse than search.
Do this instead: Start with 10-15 focused topics. Add only after usage data justifies. Retire poor-performing topics quarterly.
Why it fails: RAG index over 30 days old on high-change data = confidently wrong answers.
Do this instead: RAG refresh cadence matched to content velocity. Knowledge = daily; policy = weekly; product info = per release.
Why it fails: Autonomous agents executing production actions without gates = ONE bad decision = production outage.
Do this instead: Approval gates on destructive/costly actions. Loosen as trust proven per action type. Never fully unattended.
Why it fails: Feature enabled + no comms = users don't discover + AI value never realized. Bill still due.
Do this instead: Adoption plan: comms, training, showcase wins, quarterly usage reports. Change management is 50% of AI ROI.
Why it fails: Changing prompts silently = A/B accuracy changes = no way to attribute regressions.
Do this instead: Prompt template versioning + change log + rollback. Prompt = code, treat it that way.
Enable Incident Summarization, Resolution Notes, Change Risk across the ITSM practice.
Deploy VA to Teams for IT + HR common requests. Meet users where they work.
Modern portal with Now Assist AI Search — deflect 40%+ of internal inquiries.
Now Assist for customer service — response drafts, article gen, sentiment triggers.
Now Assist for SecOps — alert triage, IOC context, executive briefs.
Purpose-built AI skill using your data + prompts + retrieval + guardrails.
Multi-step AI agents that plan + execute — with human gates on high-impact actions.
Build the review, monitoring, and quality-control framework that keeps AI trustworthy.
ServiceNow's own foundation model (developed with NVIDIA + Anthropic partnerships) is the default. Azure OpenAI (GPT-4/4o) is available as a fallback or alternative. Custom models can be routed via Skill Kit for specialized domains.
PII redaction guardrails are configurable (regex + LLM-based). Prompts + outputs can be logged or not per policy. Data residency respects your ServiceNow instance region. We design the data-protection posture as part of every deployment.
Per-interaction licensing (Now Assist Requests). Different skills consume different amounts. We design a cost model + usage caps as part of implementation to prevent surprise bills.
Hybrid. Structured topics use NLU (deterministic, fast, reliable). Open questions fall back to Now Assist LLM. The mix is configurable per topic; we tune this per use case.
Yes — both are first-class channels. Also Web, Employee Center, mobile app, and customer portals. Same underlying topics + skills; channel-specific tuning available.
Virtual Agent = conversational chatbot (user-initiated). AI Agents (2024+) = autonomous workflow executors that can plan multi-step tasks + make decisions across workflows (with human-in-loop gates). Different capability, complementary.
Yes — via Skill Kit + model routing. You can call custom LLM endpoints from a Now Assist skill. Governance and cost stay under your control.
Typical: 15-30% agent productivity gain in ITSM/CSM, 20-40% deflection lift on Virtual Agent, 20-40% knowledge article authoring time saved. Actual varies by adoption discipline + prompt tuning quality.
Layered approach: RAG grounding, human-review gates for customer-facing text, factuality guardrails, SME feedback loops, prompt tuning. Never fully unattended for high-impact outputs.
Separate licensing on top of core ServiceNow. Sold in request bundles + product-specific packs (Now Assist for ITSM, HR, CSM). We help you size the right entitlement based on projected use.
Reach out — we get back within 1 business day.
