Standalone article · part of a sequenced guide
What you'll unlock: Now Assist is not one feature — it is a skill framework on a managed LLM layer with ACL-grounded context. Production value comes from skill selection, retrieval quality, human approval on output, and admin discipline — not from enabling every toggle.
Now Assist (GenAI Core)
The deepest dive into ServiceNow's flagship AI capability — from architecture to production configuration
Chapter context
Most enterprises bought Now Assist before they understood skills, grounding, or governance. The result: low adoption, security stalls, and finance asking why GenAI spend rose without MTTR moving. This chapter is the flagship deep dive. Pair it with Chapter 1's ecosystem map and future AI Search chapter when retrieval underperforms.
Is this chapter for you?
Are you the Now Assist admin or platform owner?
Read Concepts 5–6 fully — skill tuning and governance are your job. Concepts 1–2 for architecture credibility.
Are you rolling out ITSM agent assist?
Concept 2 + section 2.7 PDI walkthrough — start with incident summary only; expand after knowledge hygiene.
Is your scope CSM, HR, or employee portal deflection?
Concept 3 — persona, field exclusion, and legal review before any employee-facing send.
Are you a ServiceNow developer or platform builder?
Concept 4 — code/Flow/ATF assist with mandatory peer review; never skip ATF on accepted code.
Chapter 1 gave you the ecosystem map. This chapter goes inside Now Assist — the capability most organisations license first and configure last. You will learn how skills are built and routed, how ITSM/CSM/HR skills differ, how developers use assist without creating security debt, and how admins run a governed program. By the end you can activate skills on PDI, explain the prompt pipeline in an interview, and publish a governance policy your CISO will sign.
Chapter insight
Now Assist is not one feature — it is a skill framework on a managed LLM layer with ACL-grounded context. Production value comes from skill selection, retrieval quality, human approval on output, and admin discipline — not from enabling every toggle.
Reference diagrams
Now Assist request pipeline
Every assist invocation walks this path. Debugging bad output means tracing which stage failed — not re-prompting blindly.
Now Assist by product domain
Skills are product-bound. Architecture maps entitlements to workflows — not one global chat.
Implementation paths
Sequence: architecture literacy → one skill POC → tune → govern → scale by product.
Concept 1
Now Assist Architecture
Skills, LLM routing, prompt assembly, grounding, and the full stack from request to rendered output
1.1
The Now Assist skill framework
How skills are defined, packaged, and activated per product domain
Key takeaway
A Now Assist skill is a governed unit of GenAI capability — prompt template, context sources, output schema, and product binding — activated per domain (ITSM, CSM, HR, Platform).
Why this matters
Interviewers ask 'what is a skill vs a plugin?' Architects who answer with the framework design implementations; admins who answer 'it's a toggle' create fragile rollouts.
Skills are product-scoped packages. ServiceNow ships platform skills; enterprises extend with custom skills where the framework allows. Activation is entitlement + admin configuration + role visibility.
Packaging layers: definition → context binding → surface binding
POC discipline: enable one skill per workflow, measure adoption, then expand — not all ITSM skills on day one.
Workflow — do this next
- 01On PDI: Now Assist Admin → list available skills by product.
- 02Map each skill to a workflow step and owning team.
- 03Document activation status in a SKILL_REGISTRY spreadsheet.
Real example
Skill registry prevents duplicate pilots
A global IT org had three regions independently enabling incident summary with different KB scopes. Central architecture published SKILL_REGISTRY.md — one owner per skill, standard context sources, shared guardrails. Reduced config drift and made security review repeatable.
1.2
The LLM connection layer
How ServiceNow routes to model providers and what the fallback chain looks like
Key takeaway
ServiceNow operates managed LLM connectivity — you configure skills and policies; the platform handles provider routing, failover, and operational SLAs rather than you wiring raw API keys in production.
Why this matters
Security and procurement ask 'which model?' and 'what if it's down?' This layer is your answer.
Enterprises typically do not point Now Assist at arbitrary public endpoints. ServiceNow's managed connection layer abstracts provider details while exposing admin controls for skill behaviour.
Fallback thinking: if primary model latency spikes or a region is unavailable, the platform may route to alternate capacity — exact behaviour is release-dependent. Architects document degraded mode in runbooks.
Contrast with IntegrationHub calling external LLM APIs: possible for custom apps, but Now Assist skills use the native layer — different governance path.
Workflow — do this next
- 01Confirm with ServiceNow account team which AI SKUs and regions apply.
- 02Define degraded-mode SOP: what agents do when assist panel errors.
- 03Never build production dependency on unaudited external LLM keys for core ITSM.
Real example
Regional outage — degraded mode saved the shift
EU instance experienced temporary GenAI latency. Runbook switched agents to manual templates for external comms; incident summaries deferred. No policy violations from rushed paste jobs. Recovery in 90 minutes — architecture had planned for absence, not just presence.
1.3
The prompt construction pipeline
How context, schema, record data, and user input assemble into a prompt
Key takeaway
Now Assist builds prompts dynamically: system instructions + skill template + ACL-filtered record context + optional retrieval snippets + user action — not a single static string.
Why this matters
Debugging bad output starts here. 'Fix the prompt' without knowing assembly order wastes weeks.
Pipeline stages: system layer → skill template → record context → retrieval → user input
Schema awareness: skills often receive structured field metadata so the model knows field labels vs values — reducing confusion on custom tables.
Interview tip: draw this pipeline on a whiteboard — instant senior signal.
Workflow — do this next
- 01Pick one skill and list every input source (fields, KB, user text).
- 02On PDI, trigger skill on a test record — note which fields change output.
- 03Remove a context source — observe hallucination increase (grounding lesson).
Real example
Custom field not in skill context
Now Assist summaries ignored u_business_impact on incidents. Root cause: field not in skill context contract. Added to context mapping — summaries suddenly aligned with exec reporting. Pipeline literacy turned a 'bad AI' ticket into a config fix.
1.4
The grounding strategy
How Now Assist anchors responses in instance data rather than model hallucination
Key takeaway
Grounding = retrieval + record context + output constraints. Ungrounded skills answer from parametric model knowledge — dangerous for policy, pricing, and technical procedures.
Why this matters
Hallucination complaints are usually grounding complaints. This section is how you fix them.
Grounding sources: record fieldsknowledge articlessimilar records. Skills should instruct the model to prefer sources over invention.
Techniques: require citations, 'answer only from context' instructions, refusal when context insufficient, and human review on external-facing text.
Pair grounding with AI Search index quality — Ch 1 warned; Ch 2 implements.
Workflow — do this next
- 01Audit top 10 KB articles used by assist — retire outdated.
- 02Enable citation-style output where product supports it.
- 03Red-team: ask for policy not in KB — verify refusal or uncertainty.
Real example
VPN policy hallucination — grounding fix
Assist cited obsolete split-tunnel policy. Retrieval indexed 2019 article. Fix: retire article, boost 2025 policy, add skill guardrail 'if no matching KB, say unknown and link HR portal.' Hallucinations on policy dropped to near zero in UAT.
1.5
Token management in Now Assist
How ServiceNow controls context window consumption and cost at scale
Key takeaway
Long work notes, huge related lists, and bloated retrieval eat context budget — causing truncation, higher cost, and quality drop. Architects design what enters the window.
Why this matters
Finance asks why AI spend scaled linearly with ticket volume. Token discipline is the answer.
Context window is finite. ServiceNow applies summarisation and truncation strategies to fit limits. Custom skills must explicitly limit inputs.
Cost drivers: number of assist invocations × context size × skill complexity. High-volume L1 summary on every refresh burns budget — trigger on open or on demand, not every keystroke.
Design pattern: summarise history once, store in a field, feed summary to subsequent assists — amortise tokens.
Workflow — do this next
- 01Measure assist invocations per ticket type in POC.
- 02Limit retrieval to top-k chunks with relevance threshold.
- 03Trigger assist on agent action, not auto on every form load — unless ROI proves out.
Real example
Auto-summary on every save — cost surprise
Team enabled incident summary on every work-note save. 400% increase in GenAI consumption in one month. Changed to 'on open' + manual refresh. Quality unchanged; cost normalised. Token management is product design.
1.6
The output rendering layer
How GenAI responses surface in forms, portals, and workspaces
Key takeaway
Now Assist outputs render as inline suggestions, editable drafts, side panels, or inserted field values — UX shape affects adoption more than model quality.
Why this matters
Agents ignore assist that fights their workflow. Rendering design is change management.
Agent workspace: typically draft in panel — never silent auto-post to customer without policy.
Portals and employee experience: shorter responses, links to KB, escalation CTAs. Developer surfaces: code blocks in Script Editor with accept/reject.
Accessibility and mobile: rendering must work on small screens — field teams adopt mobile-first designs.
Workflow — do this next
- 01Shadow agents for one shift — watch where assist appears and whether they use it.
- 02Adjust UX: move panel, reduce clicks to accept, show citations inline.
- 03Track accept rate vs ignore rate per skill.
Real example
Accept rate 12% → 68% after UX change
Resolution note draft required 4 clicks to insert. Moved to one-click 'insert and edit' with citation footnotes. Adoption followed UX, not model upgrade.
1.7
The feedback loop
How thumbs-up/down data feeds into skill quality improvement
Key takeaway
Explicit feedback signals (helpful / not helpful) plus implicit signals (edit distance, accept without edit) guide skill tuning — not automatic retraining of foundation models.
Why this matters
Continuous improvement programs need a feedback story. This is how ServiceNow skills get better operationally.
Thumbs feedback aggregates to skill owners — patterns of dislike on certain categories indicate grounding or tone issues.
Implicit signals often matter more: edit distance, time-to-accept, abandon rate. High edit distance → fix template or context, not 'retrain AI.'
Close loop: weekly review with top agents; feed quotes into prompt/guardrail updates; version skills in dev → test → prod.
Workflow — do this next
- 01Export feedback metrics weekly in POC.
- 02Tag negative feedback by category and missing KB.
- 03Assign owner to iterate skill monthly.
Real example
Feedback drove guardrail on legal language
CSM agents thumbs-downed 34% of drafts for overly casual tone. Added system tone parameter and prohibited phrases guardrail. Thumbs-down fell to 8% in three weeks.
1.8
Now Assist in the API
How developers call GenAI capabilities programmatically from scripts and flows
Key takeaway
Platform APIs and Flow actions expose GenAI for custom apps — same governance, different surface. Use for bespoke UX, batch jobs, and integration patterns skills don't cover.
Why this matters
Builders interview on 'can I call Now Assist from Flow?' Yes — with architecture discipline.
Patterns: Flow AI actions, scripted calls, custom UI
Same rules apply: ACL context, logging, rate limits, human approval on external output. API is not a bypass for governance.
POC: wrap GenAI call in subflow with error handling and audit field updates.
Workflow — do this next
- 01Identify use case not covered by native skill — justify custom vs configure.
- 02Build subflow: gather context → GenAI action → parse output → conditional update.
- 03Log request id and output hash on record for audit.
Ready-to-use artifacts
Complete templates — paste directly into your AI tool or automation workflow.
GenAI subflow pattern (conceptual)
Architecture checklist for Flow + Now Assist integration.
## Subflow: GenAI_Assist_IncidentSummary 1. Input: incident sys_id 2. Validate ACL + entitlement 3. Assemble context (fields, notes cap, KB retrieval) 4. Invoke GenAI action / API 5. Parse structured output 6. Write to u_ai_summary + u_ai_summary_at 7. On error → log + notify + skip write Never auto-close incident from GenAI output in v1.
Concept 2
Now Assist for ITSM
Incident summaries, resolution notes, change risk, and the PDI walkthrough for production ITSM skills
2.1
Incident summarisation
How Now Assist reads incident history and produces a structured summary for incoming agents
Key takeaway
Incident summary skills ingest description, work notes, related CIs, and recent updates — outputting a scannable brief so agents skip 10 minutes of scroll-back.
Why this matters
Highest-ROI ITSM skill for mature ticket volume. Interviewers ask for your summary architecture.
Inputs typically include short description, priority, assignment, chronology of work notes, and linked CMDB items. Output should be structured — not a prose wall.
Trigger design: on assignment change or on open — avoids regenerating on every minor edit unless agent requests refresh.
Quality gate: if incident has <3 notes, summary may add little value — consider minimum context threshold.
Workflow — do this next
- 01Baseline: time for new assignee to understand ticket (sample 20 incidents).
- 02Enable summary skill on PDI with test incidents containing long threads.
- 03Compare summary accuracy to supervisor gold standard.
Real example
Major incident handoff — 14 min to 3 min
P1 reassignment at 2am. Incoming manager read AI summary with timeline, current hypothesis, and pending actions — validated against last two notes. Faster situational awareness without reading 60 entries.
2.2
Resolution note generation
The AI that drafts resolution notes from work notes and closure codes
Key takeaway
Resolution note skills synthesise closure narrative from work history and resolution codes — agents edit and approve before resolve.
Why this matters
Reopen rates often trace to vague resolution notes. This skill standardises closure quality.
Skill maps resolution code and work notes into customer-appropriate and audit-ready closure text.
Guardrail: never include internal-only jargon in customer-visible fields — separate internal vs external note skills where product allows.
Metric: edit distance on resolution notes before/after assist.
Workflow — do this next
- 01Collect 10 exemplar resolution notes from senior agents — tone reference.
- 02Configure skill to match structure: cause, fix, verification.
- 03Require agent accept before state moves to Resolved.
Real example
Reopen rate drop after resolution standardisation
Reopen rate 11% on password incidents. Resolution notes were one-liners. Assist drafted verification steps from work notes; agents added one line. Reopen rate fell to 6% in 60 days.
2.3
Similar incident surfacing
How GenAI surfaces relevant historical incidents with natural language context
Key takeaway
Combines semantic similarity (AI Search / PI) with GenAI explanation of why incidents relate — not just a list of numbers.
Why this matters
Differentiates platform approach from 'search sys_id.' Agents get narrative context for faster pattern match.
Retrieval finds candidate incidents; GenAI explains shared symptoms, resolution pattern, and caveats — grounded in record fields.
Works best when historical data is labelled with good resolution codes and CI linkage.
Interview: distinguish PI similarity scores vs GenAI narrative layer.
Workflow — do this next
- 01Test on recurring incident types (VPN, Outlook, access requests).
- 02Verify cited incidents exist and resolutions were successful.
- 03Add dismiss feedback when similarity is wrong — tune retrieval.
Real example
Outlook sync — pattern in 90 seconds
Agent opened new Outlook sync ticket. Assist surfaced three resolved incidents from same patch cycle with shared fix — certificate update. Agent applied known fix in 12 minutes vs 45 average for that category.
2.4
Change risk assessment
How Now Assist evaluates a change request and assigns a risk narrative
Key takeaway
Change assist reads description, affected CIs, schedule, backout plan, and history — producing risk narrative for CAB context, not replacing CAB judgment.
Why this matters
Change managers face document overload. Assist accelerates review; humans retain approval authority.
Output: risk factors, missing information checklist, comparison to similar past changes — advisory only
Integrate with change policy: flag missing backout, blackout conflicts, high-impact CI classes.
Do not auto-approve changes from GenAI risk score — compliance nightmare.
Workflow — do this next
- 01Pilot on standard changes first, not emergency.
- 02CAB reads assist narrative as prep — not sole vote input.
- 03Log when CAB overrides assist recommendation — improve prompts.
Real example
CAB prep time halved
Change manager used assist summary before weekly CAB. Missing backout plan flagged on 3 of 12 changes pre-meeting — fixed before vote. Meeting shortened 40 minutes.
2.5
Problem management assist
The AI-driven root cause analysis capability and how to configure it
Key takeaway
Problem assist correlates incident clusters, change history, and known errors — drafting RCA hypotheses for problem managers to validate.
Why this matters
Problem management is understaffed everywhere. Assist makes problem records viable again.
Inputs: linked incidents, work notes, recent changes on affected CIs, known error database. Output: hypothesis tree, suggested problem statement, recommended investigations.
Human validates causation — correlation from AI is not root cause until proven.
Configure linkage discipline: incidents must relate to problems for assist to work.
Workflow — do this next
- 01Select incident cluster with suspected common cause.
- 02Create problem record; link incidents; run assist.
- 03Problem manager validates hypothesis with evidence before publishing RCA.
Real example
Recurring Wi-Fi drops — problem record in one day
47 incidents across two sites. Assist highlighted shared controller firmware and change window overlap. Problem manager confirmed in 4 hours; permanent fix in 2 weeks. Without assist, cluster might have stayed as isolated incidents.
2.6
Interaction summary
How Now Assist summarises full conversation history before live agent handoff
Key takeaway
When Virtual Agent or chat escalates, interaction summary gives the live agent transcript digest — intent, tried steps, customer sentiment — reducing repeat questions.
Why this matters
Handoff friction destroys CSAT. This skill is the bridge between deflection and human support.
Summary includes: customer goal, bot actions taken, data collected, escalation reason, open slots — attached to incident or interaction record.
Agent should see summary before first customer message — not buried in related lists.
Pair with VA containment metrics — measure handoff quality, not just handoff rate.
Workflow — do this next
- 01Configure VA escalation to create incident with transcript attachment.
- 02Enable interaction summary on handoff event.
- 03Agent survey: 'Did summary reduce repeat questions?'
Real example
Chat handoff CSAT +12 points
Customers hated repeating VPN error codes after bot escalation. Interaction summary populated custom field on incident. Agents acknowledged issue in first reply. Repeat-question complaints dropped sharply.
2.7
Configuration walkthrough
Activating, configuring, and testing each ITSM skill on PDI
Key takeaway
PDI path: verify entitlements → enable ITSM skills → configure knowledge scope → assign roles → test with synthetic incidents → measure before promoting.
Why this matters
Hands-on PDI steps are what separate playbook readers from cert collectors.
Step 1: Confirm Now Assist plugins/entitlements on PDI (may require developer instance AI features).
Step 2: Now Assist Admin — enable incident summary, resolution note, similar incidents per release catalog.
Step 3: Scope AI Search / knowledge sources to ITSM KB spaces.
Step 4: itil / itil_admin roles — verify skill visibility in Agent Workspace.
Step 5: Create test incidents with varied note length; document expected vs actual; file defects vs config gaps.
Workflow — do this next
- 01Open PDI → Now Assist configuration module (per your release).
- 02Enable one skill — incident summary only for week 1.
- 03Run 10 test scenarios; capture screenshots for playbook wiki.
- 04Expand to resolution notes week 2 after feedback review.
Ready-to-use artifacts
Complete templates — paste directly into your AI tool or automation workflow.
PDI ITSM Now Assist test cases
Minimum test pack before UAT sign-off.
| # | Scenario | Pass criteria | |---|----------|---------------| | 1 | Long thread (20+ notes) | Summary captures timeline | | 2 | New incident (1 note) | Graceful short summary or prompt to refresh later | | 3 | Wrong category | Summary still factually accurate | | 4 | Restricted field in notes | Field not leaked to summary | | 5 | Resolve with code | Resolution draft matches code tone | | 6 | Similar incidents exist | Cites real records with valid resolutions | | 7 | VA handoff | Interaction summary on incident | | 8 | Non-English KB (if applicable) | Correct language policy |
2.8
Real use case: ITSM transformation at scale
The architecture, the metrics, and the change management
Key takeaway
Enterprise ITSM Now Assist rollout: phased skills, champion network, knowledge prerequisite, metrics on handle time and quality — not big-bang enablement.
Why this matters
Case study format for interviews and business cases.
Architecture: AI Search cleanup → incident summary + similar incidents → resolution notes → change assist → VA handoff summaries. PI routing ran parallel — not replaced.
Metrics: handle time −17%, resolution note quality score +23%, agent NPS +11, GenAI cost per ticket tracked monthly.
Change: 6 champion agents per region, weekly office hours, feedback loop to skill owners, executive dashboard on adoption not just licenses.
Workflow — do this next
- 01Phase 0: knowledge hygiene (4 weeks).
- 02Phase 1: summary + similar (4 weeks).
- 03Phase 2: resolution + handoff (4 weeks).
- 04Phase 3: change assist for standard changes.
Real example
8,000-agent IT organisation — 12-month rollout
Global bank phased Now Assist across three regions. Phase 0 retired 2,400 stale KB articles. Phase 1 champions in Mumbai, Dublin, Dallas. Year-one result: $4.2M annualised productivity estimate, 0 critical audit findings on AI outputs due to human-approve policy on external text.
Concept 3
Now Assist for CSM and HR
Case summaries, response drafting, HR data handling, deflection, coaching, and persona setup
3.1
Case summarisation in CSM
How agents receive context-rich summaries at case open
Key takeaway
CSM summaries emphasise customer identity, product context, SLA, sentiment history, and open commitments — different shape than ITSM incident briefs.
Why this matters
CSM agents juggle relationship context, not just technical state. Skills must reflect that.
Context includes account tier, product entitlements, prior cases, escalations, and comms channel. Output tone is customer-facing aware.
Integrate with Customer Service Management workspace — summary panel at case open.
ACL: agents see only accounts in their domain — summary inherits restrictions.
Workflow — do this next
- 01Map CSM case fields required in summary template.
- 02Test on high-touch account with long case history.
- 03Validate no cross-account data leakage in multi-tenant scenarios.
Real example
Enterprise SaaS — case open in 90 seconds
Tier-1 customer with 6 prior cases. Summary highlighted open billing dispute and last exec escalation — new agent avoided tone-deaf response. CSAT saved on reactive call.
3.2
Response drafting
Now Assist composing outbound communication from case context and tone guidelines
Key takeaway
Draft skills produce email, portal reply, or chat response from case facts and brand tone parameters — human sends after review.
Why this matters
CSM throughput is communication-bound. Drafting is the killer app if tone is controlled.
Configure tone guidelines and prohibited phrases (legal liability, competitor mentions).
Include case resolution steps and KB links — grounded citations reduce escalations.
Track edit distance by agent tier — juniors benefit most; tune training accordingly.
Workflow — do this next
- 01Collect brand voice examples — feed into skill tuning.
- 02Enable draft for email channel first; chat second.
- 03Supervisor spot-check 20 outbound emails weekly in month one.
Real example
B2B support — 22% more cases per agent
Draft assist on email channel cut composition time. Seniors edited 10%; juniors 35%. Training focused on edit patterns from seniors. Throughput rose without CSAT drop.
3.3
HR case assist
Employee experience use cases and special data handling considerations
Key takeaway
HR skills handle sensitive life events — benefits, payroll, conduct — with stricter field exclusion, legal review, and often no external LLM experimentation outside platform controls.
Why this matters
HR is highest-risk domain for GenAI. Architects win trust by designing restrictions first.
Use cases: case summarisation for HR agents, policy-aligned drafts, onboarding checklists — not performance judgment or medical advice.
Exclude: disciplinary details, medical notes, compensation fields from GenAI context unless legal approves.
Coordinate with works council / union on augmentation messaging — Ch 1 pattern applies stronger here.
Workflow — do this next
- 01Legal review list of prohibited topics for HR assist.
- 02Field-level exclusion on HR case tables.
- 03Human approval mandatory on all employee-facing HR text.
Real example
HRSD — parental leave cases
Assist summarised case and drafted policy-cited responses from approved KB only. Medical details in secure attachment excluded from context. Zero PII leakage in 6-month audit sample.
3.4
Self-service deflection
How Now Assist powers the portal experience to contain cases before creation
Key takeaway
Employee and customer portals use Now Assist + AI Search to answer in natural language — creating cases only when intent unresolved or policy requires human.
Why this matters
Deflection is a portal design problem dressed as AI. Now Assist is the language layer on search.
Flow: employee asks question → retrieval → grounded answer → 'did this help?' → case creation on no.
Measure deflection with record truth — not chat session end.
Stale HR/policy KB kills deflection — content ops before model ops.
Workflow — do this next
- 01Identify top 20 HR/CSM portal queries.
- 02Ensure AI Search returns correct articles.
- 03Enable conversational entry with Now Assist answers.
- 04Track case creation rate before/after.
Real example
Employee portal — benefits questions
41% of benefits cases were password and FAQ variants. Portal assist + KB cleanup deflected 33% before case create. Remaining cases had higher complexity — appropriate escalation.
3.5
Agent coaching
How GenAI surfaces recommended next steps and knowledge articles in real time
Key takeaway
Coaching skills suggest next best action — check entitlement, offer refund policy, escalate to tier 2 — based on case state and playbooks, not generic advice.
Why this matters
Reduces variance between junior and senior agents. Popular in BPO and high-volume CSM.
Ground coaching in playbooks — avoid free-form 'you should' without policy link.
Show article links alongside suggestion — agent learns while working.
Disable coaching on legal/compliance cases if policy requires specialist-only handling.
Workflow — do this next
- 01Encode top 10 case types as playbook articles.
- 02Enable coaching skill with playbook retrieval.
- 03Measure suggestion acceptance rate.
Real example
Telco BPO — coaching on billing disputes
Coaching suggested credit policy check before escalation. Junior agents followed playbook 2x more often. Escalation rate down 14% on billing category.
3.6
Sentiment analysis integration
Using GenAI to flag emotional signals in case communication
Key takeaway
Sentiment signals — frustration, urgency, churn risk — flag cases for priority handling or supervisor review; not for automated punitive actions.
Why this matters
Revenue-at-risk accounts need human judgment; sentiment is a triage input.
GenAI can infer sentiment from customer messages and case history — surface as field or banner: 'high frustration detected.'
Ethical guardrail: do not use sentiment to deny service or auto-close.
Combine with SLA and account tier for routing — sentiment alone is noisy.
Workflow — do this next
- 01Define thresholds for supervisor alert.
- 02Test on historical escalated cases — recall vs false positive.
- 03Train agents on how to respond to sentiment flags.
Real example
Churn risk — exec outreach trigger
Sentiment + enterprise tier triggered supervisor playbook on renewal case. Proactive exec call retained $2.1M account. Flag was input, not auto-decision.
3.7
Configuration and persona setup
Adapting Now Assist tone and scope for HR vs customer service contexts
Key takeaway
Personas bundle tone, vocabulary, allowed topics, and channel defaults — HR empathetic and policy-cite; CSM brand-aligned and commercial-aware.
Why this matters
One global persona fails regulated HR and flashy consumer brand alike.
HR persona: formal, inclusive language, mandatory KB citation, no speculation on individual compensation.
CSM persona: brand voice guide, empathy without over-apology, product terminology glossary.
Separate personas per region/language where local policy differs.
Workflow — do this next
- 01Workshop with HR comms and CSM marketing on tone docs.
- 02Configure persona parameters per product.
- 03UAT with native speakers for non-English locales.
Real example
Dual persona — same platform, different guardrails
Multinational configured HR persona with strict citation mode and CSM persona with promotional tone limits. Shared platform team, separate skill owners. Audit passed for HR; brand team approved CSM drafts.
3.8
Real use case: HR service delivery transformation
What changes, what stays, and what the numbers show
Key takeaway
HRSD transformation: assist augments case handling and portal deflection; humans retain decisions on conduct, medical accommodation, and compensation. Metrics on case duration and employee satisfaction.
Why this matters
HR leaders need realistic narrative — not 'AI replaces HR business partners.'
Changed: faster case summaries, policy-aligned drafts, portal deflection on FAQs, coaching on routine cases.
Unchanged: human decision on accommodations, investigations, complex grievances.
Numbers: case handle time −15%, employee CSAT +8 points on routine inquiries, zero unsupervised external sends.
Workflow — do this next
- 01Start with top 5 routine case types only.
- 02Legal sign-off on context exclusions.
- 0312-week pilot on one country before global.
Real example
15,000-employee manufacturer — HRSD pilot
Germany-only pilot: benefits and leave cases. Assist + portal deflection. Works council approved augmentation framing. Expanded to UK after DPIA. HR business partners spent saved time on complex employee relations — not headcount reduction.
Concept 4
Now Assist for Developers
Code generation, Flow generation, ATF tests, explanation, docs, refactoring, and PDI enablement
4.1
Code generation in Script Editor
How Now Assist writes GlideRecord queries, Business Rules, and client scripts
Key takeaway
Developer Now Assist generates ServiceNow-native code patterns — GlideRecord, gs, g_form — from natural language, with accept/reject in Script Editor.
Why this matters
Admins and developers interview on 'does it write good GlideRecord?' Know limits and review discipline.
Strengths: boilerplate queries, field validation stubs, simple business rules. Weaknesses: complex transactions, scope chains, performance-critical bulk operations — always review.
Always validate: addQuery ACL impact, null checks, async patterns, and update() in before rules.
Team standard: generated code goes through same peer review as human code.
Workflow — do this next
- 01On PDI: open Script Editor → invoke Now Assist for Development.
- 02Prompt: 'GlideRecord query open incidents assigned to Network group.'
- 03Compare output to team coding standard — note gaps.
Real example
Junior admin — first business rule in 20 minutes
New admin used assist for date validation on change request. Senior corrected getValue() pattern. Still faster than blank page — learning accelerated with review, not blind trust.
4.2
Flow generation
How natural language descriptions produce Flow Designer flows
Key takeaway
Flow generation translates intent into draft flows — triggers, conditions, actions — that builders refine in Flow Designer; not production-ready without human validation.
Why this matters
Flow is where citizen developers collide with AI — governance matters.
Describe: 'When incident priority becomes 1, notify major incident manager and create problem task.' Assist proposes flow skeleton.
Review: error handling, subflow reuse, ACL on actions, idempotency on re-triggers.
Citizen developer policy: generated flows require admin approval before promote.
Workflow — do this next
- 01Generate draft flow from plain English on PDI.
- 02Walk through test plan — happy path and edge cases.
- 03Add logging and failure notifications before activate.
Real example
Catalog request flow — 2 hours vs 1 day
Builder generated 80% of hardware request flow from description. Manual work: integration action tuning and approval branch. Net savings one day; quality gate prevented missing rollback email.
4.3
Test generation
How Now Assist creates ATF test cases from flow and script descriptions
Key takeaway
Assist drafts Automated Test Framework steps — open form, set field, submit, assert — accelerating regression coverage for flows and critical scripts.
Why this matters
Test debt blocks upgrades. ATF generation is an underused Now Assist developer win.
Input: flow name or script purpose. Output: ATF step sequence with suggested assertions.
Human adds: test data setup, teardown, environment constraints, flaky UI waits.
Integrate with CI pipeline — tests generated in sprint, not hoarded in draft.
Workflow — do this next
- 01Pick one production-critical flow.
- 02Generate ATF draft; run on PDI.
- 03Fix selectors and timing; commit to test suite.
Real example
Upgrade regression — 12 new ATF tests in a sprint
Team generated ATF coverage for 12 high-risk flows before Vancouver upgrade. Caught 3 breakages in test env. Assist did not replace test strategy — accelerated authoring.
4.4
Code explanation
How developers use Now Assist to understand inherited scripts and legacy code
Key takeaway
Explain mode walks through legacy Business Rules, Script Includes, and client scripts — invaluable for acquisitions and consultant handoffs.
Why this matters
Every ServiceNow shop has tribal knowledge in opaque scripts. Explanation reduces bus factor.
Paste or point assist at script — ask for line-by-line explanation, side effects, and tables touched.
Verify explanation against execution — assist may miss dynamic table names.
Document output into technical debt register — explanation becomes spec for refactor.
Workflow — do this next
- 01Select top 5 highest-incident legacy scripts.
- 02Generate explanations; senior dev validates.
- 03Add comments or wiki pages from validated explanations.
Real example
Acquisition integration — 400 Script Includes
Acquired company's instance had minimal documentation. Assist explanation accelerated mapping of integration Script Includes. Humans flagged 12 dangerous patterns assist missed — still 3x faster than manual-only read.
4.5
Documentation generation
Producing technical documentation from code and flow definitions
Key takeaway
Assist generates runbooks, field dictionaries, and flow narratives from artifacts — keeping docs closer to truth if regenerated on change.
Why this matters
Auditors and handoffs need docs. AI lowers the cost of staying current.
Generate: flow diagram narrative, input/output contract, error paths, related tables.
Store in knowledge or Git-linked docs — version with release.
Mark auto-generated sections — human owner signs off quarterly.
Workflow — do this next
- 01Pick flow without documentation.
- 02Generate doc; owner edits for operational truth.
- 03Link doc from flow properties or team wiki.
Real example
SOX audit — flow documentation in one week
Audit required documentation for 40 financial flows. Assist generated first draft; owners corrected approval paths. Audit passed with minor edits — vs months of manual doc debt.
4.6
Refactoring assist
How Now Assist suggests improvements to existing scripts
Key takeaway
Refactor suggestions target readability, GlideRecord best practices, and duplication reduction — not automatic apply in production.
Why this matters
Technical debt sprints benefit from AI pair programming — with senior review.
Ask: 'refactor for performance' or 'replace nested GlideRecord with single query' — compare suggestions to platform best practices docs.
Watch for: over-abstraction, breaking subtle business logic, deprecated APIs.
Pair with ATF — refactor without tests is gambling.
Workflow — do this next
- 01Run ATF baseline on target script's flows.
- 02Apply refactor suggestion in dev.
- 03Re-run ATF; performance test on large dataset.
Real example
GlideRecord get() in loop — assist caught it
Legacy incident update rule used get() per row in loop. Assist suggested bulk query pattern. 40s → 3s on test batch. Senior approved after ATF green.
4.7
The developer productivity case
Time saved, error rates, and the realistic developer uplift
Key takeaway
Realistic uplift: 15–30% on authoring tasks (scripts, flows, tests, docs); near-zero on judgment-heavy architecture. Error rate rises if review skipped.
Why this matters
Platform leaders need honest numbers — not '10x developer' keynote math.
Measure: story points, cycle time, defect escape rate, ATF coverage — before/after assist adoption.
Biggest gains: onboarding new developers, documentation debt, repetitive ATF.
Smallest gains: net-new architecture, complex integrations — still human-led.
Workflow — do this next
- 01Track assist usage in dev team for 4 sprints.
- 02Survey: where did it help vs waste time?
- 03Adjust guidelines — when assist is required vs optional.
Real example
Platform team — 6-month developer metrics
8 developers, assist enabled. Flow authoring time −22%, script defects in UAT −11% (more review), doc coverage 45% → 78%. Architecture spikes unchanged — assist did not replace design meetings.
4.8
Configuration walkthrough
Enabling and using Now Assist for Development on PDI
Key takeaway
PDI: enable developer assist SKU/plugin → Script Editor and Flow Designer assist icons → configure org coding standards in prompt hints where supported → pilot on non-prod apps only.
Why this matters
Hands-on enablement closes the chapter for builders.
Enable Now Assist for Development in instance settings (per release documentation).
Verify roles: admin, developer, or custom role with assist entitlement.
Exercise: generate GlideRecord, explain script, draft ATF, generate flow — four labs in one afternoon.
Workflow — do this next
- 01PDI → enable developer assist.
- 02Lab 1: GlideRecord from NL prompt.
- 03Lab 2: explain unknown Script Include.
- 04Lab 3: ATF for simple catalog flow.
- 05Lab 4: document that flow.
Ready-to-use artifacts
Complete templates — paste directly into your AI tool or automation workflow.
PDI Developer Now Assist — half-day lab
Self-paced lab checklist.
## Morning labs 1. GlideRecord: query + update with guard 2. Business Rule: before insert validation 3. Explain inherited Script Include (paste) ## Afternoon labs 4. Flow from NL description + manual fix 5. ATF generation + run 6. Doc export to team wiki ## Exit criteria - [ ] Peer review checklist used on all accepts - [ ] No assist-accepted code in prod without ATF - [ ] Team coding standard doc linked in wiki
Concept 5
Skill Configuration and Tuning
Prompt templates, guardrails, domain tuning, multilingual setup, A/B tests, and versioning
5.1
Skill anatomy
Prompt templates, context sources, output formats, and tone parameters
Key takeaway
Every skill decomposes into: input contract, prompt template, context sources, output schema, tone, and guardrails — tune each independently.
Why this matters
Custom skill work starts here. Interview question: 'how would you build a custom summary skill?'
Input contract: which fields, max note count, related lists. Prompt template: task instruction and format. Output schema: bullets, JSON, max length.
Tone parameters: formal, concise, empathetic — map to brand or HR policy.
Version all components in dev — diff before promote.
Workflow — do this next
- 01Decompose one native skill into anatomy diagram.
- 02Identify which parts are admin-tunable vs platform-fixed.
- 03Document your org's standard output format per skill type.
Real example
Custom outage comms skill
Built custom skill: input = incident + affected CI list; output = three-paragraph exec template; tone = calm/factual; guardrail = no root cause claim without verified field. Anatomy doc became template for 4 more skills.
5.2
System prompt customisation
How to tailor Now Assist responses to your organisation's language and standards
Key takeaway
Organisation-level instructions inject vocabulary, acronym glossaries, mandatory disclaimers, and formatting rules into skills — consistency without per-agent prompting.
Why this matters
Global enterprises need 'how we write' embedded in AI — not reinvented per user.
Examples: always use 'colleague' not 'user' in HR; include ticket number in summaries; spell product names correctly from glossary.
Avoid conflict: org instructions should not fight skill-specific tasks — layer general → specific.
Review quarterly with comms and legal — stale org prompts cause drift.
Workflow — do this next
- 01Draft ORG_AI_STYLE_GUIDE.md with legal/comms.
- 02Apply via admin system prompt settings (per release).
- 03Sample 20 outputs monthly for compliance.
Real example
Banking — mandatory disclaimer injection
Org prompt required regulatory disclaimer on customer-facing drafts. 100% compliance in spot check vs 60% when agents typed manually under pressure.
5.3
Guardrail configuration
What topics, formats, and outputs to prohibit and how to enforce them
Key takeaway
Guardrails block: competitor mentions, medical/legal advice, credential disclosure, speculative blame, off-topic generation — enforced at skill and org level.
Why this matters
One viral bad output kills a program. Guardrails are insurance.
Techniques: prohibited topic lists, required citations, max length, refusal instructions when context empty, block on restricted field patterns.
Test guardrails with red-team prompts monthly.
Log guardrail triggers — high rate means retrieval or policy gap, not just attacks.
Workflow — do this next
- 01List 20 prohibited outputs for your industry.
- 02Configure guardrails; run red-team script.
- 03Fix retrieval before weakening guardrails.
Real example
Healthcare — no diagnosis guardrail
HR health benefits skill refused diagnosis prompts and redirected to nurse line. Red-team found 0 successful diagnosis generations in 500 attempts. Legal signed off rollout.
5.4
Domain-specific tuning
Adapting skills for specialised industries (legal, healthcare, finance)
Key takeaway
Regulated domains need narrower retrieval, specialist glossaries, stricter human approval, and often geographic persona splits.
Why this matters
Generic ITSM tuning fails in legal and finance — architects show domain fluency.
Legal: no contract interpretation; cite only approved playbooks; privilege-aware field exclusion.
Finance: numbers must match record fields — instruct model never to calculate independently.
Healthcare: HIPAA-aligned context minimisation.
Workflow — do this next
- 01Domain SME workshop — allowed vs forbidden assist tasks.
- 02Tune retrieval to approved libraries only.
- 03Dual approval on external comms in regulated domains.
Real example
Insurance claims — field-anchored numbers
Skill instructed: all amounts must copy claim field values verbatim. Hallucinated payout amounts dropped to zero in UAT. Adjusters still approve every letter.
5.5
Multi-language configuration
Enabling Now Assist in non-English environments
Key takeaway
Multilingual assist requires: language detection, KB in target languages, persona per locale, and native-speaker UAT — not just translate-English-output.
Why this matters
Global rollouts fail when HQ English skills ship to LATAM unchanged.
Configure language preference on user or portal profile where supported.
Maintain parallel KB — translated articles, not machine-translated stale content.
UAT with native speakers for tone and policy accuracy.
Workflow — do this next
- 01Pick pilot language with complete KB subset.
- 02Enable locale persona.
- 03Compare CSAT vs English baseline.
Real example
LATAM HR portal — Spanish assist
Spanish KB subset 120 articles. Assist configured for es-MX tone. Deflection 28% vs 31% English — parity achieved after content investment, not model change.
5.6
A/B testing skills
How to run controlled experiments on skill variations before full deployment
Key takeaway
Test prompt variants on cohorts — team A vs team B, or percentage rollout — measuring accept rate, edit distance, and task time before global default.
Why this matters
Data-driven tuning separates architects from guessers.
Design: hypothesis, metric, duration, sample size, ethical approval for employee-facing tests.
Compare variant B only if statistically better on primary metric without harming quality guardrails.
Document winner in skill version notes.
Workflow — do this next
- 01Define metric: edit distance primary.
- 02Split 50 agents — control vs variant prompt.
- 03Run 2 weeks; analyse; promote winner.
Real example
Resolution note — bullet vs paragraph format
Variant A bullets, B short paragraphs. Bullets won: 12% lower edit distance, faster scan by agents. Promoted globally in skill v2.1.
5.7
Feedback analysis
Using feedback data to identify skill underperformance and iterate
Key takeaway
Aggregate thumbs-down by category, skill, region, and topic — pattern analysis drives prompt and retrieval fixes, not random tweaks.
Why this matters
Continuous improvement needs an analytics ritual.
Weekly dashboard: feedback rate, top negative themes, correlation with missing KB, agent tenure.
Qualitative: sample 10 negative cases — root cause tag: grounding, tone, length, wrong task.
Close loop with release notes to agents — visibility builds trust.
Workflow — do this next
- 01Build feedback export (report or admin view per release).
- 02Tag root causes in spreadsheet.
- 03Monthly skill owner review meeting.
Real example
Negative spike after upgrade
Thumbs-down doubled post upgrade. Analysis: retrieval index lag — not model change. Reindex fixed 80% of negatives in 48 hours. Feedback analysis avoided wrongful rollback.
5.8
Skill versioning
Managing skill changes across development, test, and production instances
Key takeaway
Skills move dev → test → prod with version tags, rollback plan, and change window — same discipline as code promotion.
Why this matters
Untracked skill edits in prod are how governance breaks.
Store: version id, change summary, approver, rollback steps, ATF/UAT evidence.
Use update sets or platform deployment tools per your release practice.
Never hot-edit prod skill during business hours without rollback ready.
Workflow — do this next
- 01Define skill promotion pipeline in CAB.
- 02Test in subprod with production-like data mask.
- 03Tag prod skill config with version metadata field.
Ready-to-use artifacts
Complete templates — paste directly into your AI tool or automation workflow.
Skill version log template
Track skill promotions across instances.
| Version | Skill | Instance | Change | Approver | UAT ref | Rollback | |---------|-------|----------|--------|----------|---------|----------| | 1.0.0 | incident_summary | prod | initial | J.Smith | UAT-441 | disable skill | | 1.1.0 | incident_summary | prod | added CI context | A.Lee | UAT-502 | revert to 1.0.0 |
Concept 6
Now Assist Administration and Governance
Licensing, RBAC, monitoring, data handling, audit, incident response, and enterprise policy templates
6.1
License and entitlement
Understanding the Now Assist licensing model and what each SKU unlocks
Key takeaway
Now Assist capabilities map to product SKUs — ITSM, CSM, HR, Platform/Developer packs differ. Entitlement verification on each instance prevents 'enable but not licensed' gaps.
Why this matters
Procurement and architects must speak SKU — interviews and renewals depend on it.
Work with account executive on Now Assist entitlements — map skills to purchased products. PDI may differ from prod entitlements.
Document entitlement matrix in architecture wiki — which region/instance has which pack.
Renewal: align skill roadmap to licensed products — don't design HR assist without HR SKU.
Workflow — do this next
- 01Export entitlement report from ServiceNow or partner portal.
- 02Map each planned skill to SKU.
- 03Flag gaps before POC charter sign-off.
Real example
POC blocked — missing CSM pack
CSM team built Now Assist case draft POC on prod-like subprod. Discovery: CSM Now Assist not licensed — only ITSM. Architecture paused; procurement added SKU. Saved 6 weeks of rework on wrong instance.
6.2
Role-based access to Now Assist
Who sees which skills and how access is controlled
Key takeaway
Roles and ACLs gate skill visibility — agents see ITSM skills, HR agents see HR skills, developers see dev assist. Principle of least privilege applies to AI surfaces.
Why this matters
Over-broad access causes cross-domain data exposure in assist context.
Map skills to roles: itil, sn_customerservice_agent, sn_hr_core.admin, admin/developer.
Custom roles for pilots — narrow cohort before enterprise rollout.
Review quarterly as roles proliferate in large orgs.
Workflow — do this next
- 01RBAC matrix: role × skill × instance.
- 02Test with low-privilege user — verify denied skills hidden.
- 03Align with identity governance joiner/mover/leaver.
Real example
Contractor role — dev assist without prod HR skills
Contractors received developer assist on dev instance only; prod roles excluded HR case skills. Prevented contractor seeing employee case context in error.
6.3
Usage monitoring
The dashboards and reports that show consumption, cost, and adoption
Key takeaway
Track: invocations per skill, unique users, accept rate, feedback ratio, cost allocation — executive dashboard on adoption, finops view on spend.
Why this matters
You cannot defend renewal without usage proof.
Adoption metrics: % agents using assist weekly, invocations per ticket, skill ranking.
Cost metrics: consumption trends vs ticket volume — spot runaway auto-triggers.
Value metrics: handle time, edit distance, CSAT — tie to business case.
Workflow — do this next
- 01Define monthly Now Assist ops review agenda.
- 02Build dashboard (PA or admin reports per release).
- 03Share with product owner and finance.
Real example
Finops caught runaway skill
Usage report showed 3x spike on change summary skill — misconfigured auto-run on every save. Fixed trigger; saved $40k annualised estimate.
6.4
Data handling and retention
How Now Assist processes data, what it stores, and deletion controls
Key takeaway
Understand: what leaves instance boundary (managed LLM layer), what logs persist, retention periods, and deletion on request — document in DPIA.
Why this matters
Legal asks before InfoSec. Architects bring ServiceNow docs + org policy.
Clarify with ServiceNow trust docs: customer data isolation, subprocessors, region residency.
Minimise sensitive fields in context — design choice, not only product default.
Retention: align AI logs with corporate records policy — may differ from incident retention.
Workflow — do this next
- 01Complete DPIA section on Now Assist data flow.
- 02List fields never sent to GenAI context.
- 03Define deletion process for AI interaction logs if applicable.
Real example
GDPR erasure request — scope defined
Employee erasure request triggered playbook: incident PII redacted per policy; AI interaction logs checked against retention schedule. Legal satisfied because architects pre-documented flows.
6.5
Audit and compliance
The logs that satisfy governance and the export format for auditors
Key takeaway
Audit trail: who triggered assist, on which record, what output was accepted, who edited before send — exportable for SOX, ISO, and internal audit.
Why this matters
Regulated industries require demonstrable control — not 'trust the AI.'
Enable platform logging features per release; integrate with SIEM if required.
Quarterly audit sample: random 50 accepted outputs — verify policy compliance.
Map controls to framework: SOC2 CC, ISO 27001, internal AI policy.
Workflow — do this next
- 01Define audit evidence package for Now Assist.
- 02Run mock audit with internal compliance.
- 03Remediate gaps before external audit season.
Real example
SOX ITGC — assist logs in control matrix
Now Assist accept events logged with user id and timestamp. Control tested: no customer email sent without human accept event. Auditor passed control on first sample.
6.6
The Now Assist admin role
The responsibilities, the tools, and the escalation path
Key takeaway
Now Assist admin owns: entitlements, skill config, RBAC, monitoring, feedback loop, promotion pipeline, and liaison with security/legal — not 'the person who clicked enable.'
Why this matters
Enterprises need a named role — job descriptions are forming around this.
Tools: Now Assist admin module, AI Search admin, update sets, feedback reports, skill version log.
Escalation: platform issues to ServiceNow support; harmful output to security; policy to legal.
Partner with: ITSM/CSM/HR product owners, knowledge manager, finops.
Workflow — do this next
- 01Assign primary and backup Now Assist admin.
- 02Publish RACI for skill changes.
- 03Quarterly skills council meeting.
Real example
Dedicated AI platform admin — Fortune 500
Role split from ServiceNow admin: 0.5 FTE Now Assist admin, 0.5 FTE AI Search/knowledge. Ticket volume to central config team dropped 40%; skill quality improved via ownership.
6.7
Incident response for AI
What to do when Now Assist produces incorrect or harmful output
Key takeaway
AI incident playbook: detect → contain (disable skill or degraded mode) → assess scope → notify stakeholders → root cause (config vs retrieval vs model) → fix → postmortem — same rigor as security incidents.
Why this matters
Programs without AI incident response learn painfully in public.
Severity levels: P1 harmful external comms; P2 widespread wrong policy; P3 isolated bad draft.
Contain: kill switch skill, force manual workflow, comms to agents.
Postmortem: grounding gap, guardrail miss, or user bypass — action items to skill owner.
Workflow — do this next
- 01Draft AI_INCIDENT_RUNBOOK.md before go-live.
- 02Tabletop exercise with security and comms.
- 03Assign severity matrix and on-call rotation.
Real example
Wrong refund policy in draft — 4-hour contain
Agent caught incorrect refund % in draft before send. Reported via feedback. Skill disabled globally in 15 minutes. Root cause: outdated KB article in retrieval. Article retired; skill re-enabled after reindex. No customer impact — process worked.
6.8
Governance framework template
The policy document structure that enterprise deployments need
Key takeaway
Enterprise policy covers: purpose, scope, allowed use cases, prohibited use, data classification, human approval rules, roles, monitoring, incident response, vendor/subprocessor, training, and review cycle.
Why this matters
Ch 2 closes with the artifact enterprises paste into GRC tools.
Align policy with Ch 1 trust model and Ch 2 admin practices — one coherent program.
Board and works council review for employee-facing AI.
Annual policy refresh minimum; quarterly for fast-moving agent features.
Workflow — do this next
- 01Copy template; customise with legal.
- 02Obtain sign-off from CIO, CISO, DPO, HR for HR skills.
- 03Train all assist users on policy acknowledgment.
Ready-to-use artifacts
Complete templates — paste directly into your AI tool or automation workflow.
Now Assist governance policy — outline
Enterprise policy skeleton for GRC adoption.
# Now Assist Governance Policy ## 1. Purpose & scope ## 2. Approved use cases (by product) ## 3. Prohibited uses ## 4. Data classification & field exclusions ## 5. Human approval requirements ## 6. Roles & responsibilities (RACI) ## 7. Skill lifecycle (dev/test/prod) ## 8. Monitoring & metrics ## 9. AI incident response ## 10. Training & acknowledgment ## 11. Vendor & data residency ## 12. Review schedule (annual) Appendix A: Skill registry Appendix B: Entitlement matrix Appendix C: Red-team checklist
Ready-to-use artifacts
Complete templates — paste directly into your AI tool or automation workflow.
Now Assist skill registry (starter)
Central register — one row per skill before production.
| Skill | Product | Owner | Status | Context sources | Approval required | |-------|---------|-------|--------|-----------------|-------------------| | incident_summary | ITSM | @lead | POC | incident fields, KB-IT | internal only | | resolution_draft | ITSM | @lead | planned | work notes, resolution codes | yes external | | case_draft | CSM | @csm | disabled | case, account, KB-CSM | yes always |
AI incident response — quick stub
Paste into ops wiki before go-live.
## Detect Feedback spike, harmful draft reported, policy violation ## Contain (15 min) Disable affected skill globally | Notify agents | Degraded mode SOP ## Assess Records affected, external send? | Retrieval or guardrail root cause? ## Recover Fix KB/config | Reindex | UAT | Re-enable with comms ## Postmortem Owner, timeline, preventive action in skill version log
GenAI subflow pattern (conceptual)
Architecture checklist for Flow + Now Assist integration.
## Subflow: GenAI_Assist_IncidentSummary 1. Input: incident sys_id 2. Validate ACL + entitlement 3. Assemble context (fields, notes cap, KB retrieval) 4. Invoke GenAI action / API 5. Parse structured output 6. Write to u_ai_summary + u_ai_summary_at 7. On error → log + notify + skip write Never auto-close incident from GenAI output in v1.
PDI ITSM Now Assist test cases
Minimum test pack before UAT sign-off.
| # | Scenario | Pass criteria | |---|----------|---------------| | 1 | Long thread (20+ notes) | Summary captures timeline | | 2 | New incident (1 note) | Graceful short summary or prompt to refresh later | | 3 | Wrong category | Summary still factually accurate | | 4 | Restricted field in notes | Field not leaked to summary | | 5 | Resolve with code | Resolution draft matches code tone | | 6 | Similar incidents exist | Cites real records with valid resolutions | | 7 | VA handoff | Interaction summary on incident | | 8 | Non-English KB (if applicable) | Correct language policy |
PDI Developer Now Assist — half-day lab
Self-paced lab checklist.
## Morning labs 1. GlideRecord: query + update with guard 2. Business Rule: before insert validation 3. Explain inherited Script Include (paste) ## Afternoon labs 4. Flow from NL description + manual fix 5. ATF generation + run 6. Doc export to team wiki ## Exit criteria - [ ] Peer review checklist used on all accepts - [ ] No assist-accepted code in prod without ATF - [ ] Team coding standard doc linked in wiki
Skill version log template
Track skill promotions across instances.
| Version | Skill | Instance | Change | Approver | UAT ref | Rollback | |---------|-------|----------|--------|----------|---------|----------| | 1.0.0 | incident_summary | prod | initial | J.Smith | UAT-441 | disable skill | | 1.1.0 | incident_summary | prod | added CI context | A.Lee | UAT-502 | revert to 1.0.0 |
Now Assist governance policy — outline
Enterprise policy skeleton for GRC adoption.
# Now Assist Governance Policy ## 1. Purpose & scope ## 2. Approved use cases (by product) ## 3. Prohibited uses ## 4. Data classification & field exclusions ## 5. Human approval requirements ## 6. Roles & responsibilities (RACI) ## 7. Skill lifecycle (dev/test/prod) ## 8. Monitoring & metrics ## 9. AI incident response ## 10. Training & acknowledgment ## 11. Vendor & data residency ## 12. Review schedule (annual) Appendix A: Skill registry Appendix B: Entitlement matrix Appendix C: Red-team checklist
Regional IT + HR — unified Now Assist program
A multinational enabled Now Assist on ITSM and HRSD without shared admin, skill registry, or retrieval hygiene. IT agents ignored summaries; HR legal blocked rollout after one draft cited wrong leave policy.
Before
Parallel pilots, no grounding standard, no kill switch, developer assist enabled in prod without review policy.
After
Central Now Assist admin, skill registry, Ch 2 governance policy, IT incident summary POC first, HR pilot with field exclusions and legal-approved KB only. Developer assist restricted to subprod.
- IT incident summary accept rate 71% after UX fix
- HR pilot: 0 policy violations in 90-day audit sample
- GenAI cost per ticket tracked — 18% under budget after trigger tuning
- AI incident runbook used once — contained in 4 hours, no customer impact
What goes wrong
Enabling all ITSM skills without knowledge or search prerequisite
Concept 1.4 grounding + Ch 1 AI Search warning — clean KB before scale.
HR or CSM drafts sent without human approval
Concepts 3.3, 6.5 — field exclusion and mandatory accept on external comms.
Developers accept generated code without review or ATF
Concept 4 — peer review and ATF mandatory; assist is pair programmer, not deploy button.
No owner, no monitoring, no AI incident runbook
Concept 6 — dedicated admin role, usage dashboard, governance policy template.
Vetted by Krishna KumarCurator, FactorBeam
Discussion
Discussion coming soon
Shared comments for this playbook are not live yet. When they are, you'll be able to ask questions, share what worked, and see replies from other readers.