How AI features relate to HR Leave

Core app (ch.01–12) — tables, BR, Flow, ACLs
        │
        ▼
Platform AI (this chapter) — Now Assist, Agents, VA, Search
        │
        ▼
Optional: HR Leave skills/topics referencing req_v2 / bal_v2 data

Rule 0 — Before any object (Application + Update Set + Studio)

Repeat these checks before creating or editing any application file. This section is also section 00 Update Set — create the capture set first.

1. Application picker — top-right globe / application menu → select HR Leave (x_786686_hr_leave). Never leave Global selected when creating scoped artifacts.
2. Update Set — All → System Update Sets → Local Update Sets → New:

   Field	Value
   Name	HR Leave – Build
   Application	HR Leave
   State	In Progress
   Description	Captures all HR Leave lab artifacts

   After submit: open the record → click Make Current. Direct list: {BASE}/sys_update_set_list.do
3. Open Studio — {BASE}/$studio.do → confirm header shows HR Leave.

Deep-dive sections below — original AI implementation guide content.

14 — AI in ServiceNow + Agents — Complete Implementation Guide

اقرأ أولاً (ملخص سريع): MASTER_ALL_IN_ONE_SUMMARY.md · HTML: html/MASTER_ALL_IN_ONE_SUMMARY.html
الدليل الشامل الواحد: MASTER_ALL_IN_ONE.md · HTML: html/MASTER_ALL_IN_ONE.html
ممارسة الشاشات (New/Open): MASTER_PRACTICE_GUIDE_ALL_TOPICS.md · html/MASTER_PRACTICE_GUIDE_ALL_TOPICS.html
Platform: ServiceNow Australia · Scope x_786686_hr_leave · sys_id e5f5bc9493a303507520703efaba104d

دليل الذكاء الاصطناعي والوكلاء — تنفيذ كامل من الصفر إلى الإنتاج

Platform: ServiceNow Australia (September 2026) · Next EA: Brazil
App: Custom · Scope: x_786686_hr_leave · sys_id: e5f5bc9493a303507520703efaba104d
Tables: x_786686_hr_leave_employee_v2, x_786686_hr_leave_req_v2, x_786686_hr_leave_bal_v2, x_786686_hr_leave_audit_v2
URL pattern: {BASE} = https://YOUR_INSTANCE.service-now.com

Table of Contents

1. What You Will Build
2. Direct Navigation URLs
3. Part 0 — Quick Map & Prerequisites
4. Part 1 — Plugin Activation & Roles
5. Part 2 — Now Assist Studio (Skills)
6. Part 3 — AI Agent Studio
7. Part 4 — BYO LLM (Claude / GPT)
8. Part 5 — AI Search + GenAI Search
9. Part 6 — Predictive Intelligence
10. Part 7 — Virtual Agent + NLU (Full Topics & Flows)
11. Part 8 — Document Intelligence (IDP)
12. Part 9 — Workflow Studio AI Actions
13. Part 10 — Now Assist for Developers
14. Part 11 — End-to-End HR Leave AI Example
15. Part 12 — AI Governance & Safety
16. Part 13 — Performance & Cost Optimization
17. Part 14 — Hands-On Labs (5 Sessions)
18. Part 15 — Best Practices
19. Part 16 — Production Deployment Checklist
20. Appendices

🎯 What You Will Build — ماذا ستبني

By the end of this guide you will have deployed a complete AI layer on the HR Leave scoped app:

• ✅ 6 Now Assist Skills (summarize, classify, suggest dates, history summary, sentiment, email draft)
• ✅ 1 AI Agent — HR Leave Assistant with 6 tools wired to HRLXAIAgentTools
• ✅ 1 AI Search profile — hrlx_kb_profile with GenAI answers + masked sensitive fields
• ✅ 2 Predictive Intelligence solutions — approval prediction + duration regression
• ✅ 1 NLU model — hrlx_leave_nlu with 8 intents and 200+ utterances
• ✅ 8 Virtual Agent topics — balance, request, cancel, policy, status, handoff, greeting, fallback
• ✅ 1 IDP template — medical certificate extraction wired to sick-leave flow
• ✅ 3 Flow Designer AI Actions — summarize history, classify type, draft manager email
• ✅ Governance — AI policies, PII redaction, audit to x_786686_hr_leave_audit_v2
• ✅ Monitoring dashboard — skill latency, agent containment, PI drift, search zero-results

Style in this doc: English instruction line → Arabic explanation on its own line where helpful. All code in fenced blocks.

📍 Direct Navigation URLs — روابط التنقل المباشرة

Replace {BASE} with your instance URL (e.g. https://dev12345.service-now.com).

Now Assist & AI Agents

AI Search

Predictive Intelligence

Virtual Agent & NLU

Document Intelligence

HR Leave App Tables (Grounding & Tools)

Development & Deployment

Part 0 — Quick Map & Prerequisites

0.1 Decision matrix — which AI tool when?

ابدأ بـ Skill للمهام أحادية الخطوة. انتقل إلى Agent عندما يحتاج المستخدم محادثة متعددة الخطوات مع أدوات.

0.2 Prerequisites checklist

Before starting any AI work:

[ ] HR Leave v2 tables exist and have ≥ 50 test leave requests
[ ] Roles assigned: x_786686_hr_leave.employee, .manager, .hr_admin
[ ] Update Set created: "HR Leave AI v1.0"
[ ] Application picker set to HR Leave (x_786686_hr_leave) in Studio
[ ] At least 5 KB articles tagged topic=HR Leave (for AI Search)
[ ] Integration user NOT used for AI testing — use real employee personas
[ ] Instance has outbound HTTPS to LLM APIs (if BYO) — check ACL network rules
[ ] Admin access to activate plugins (requires admin role)

0.3 Verify app scope (Background Script)

URL: {BASE}/sys.scripts.do

var app = new GlideRecord('sys_app');
app.addQuery('scope', 'x_786686_hr_leave');
app.query();
if (app.next()) {
    gs.info('App: ' + app.name + ' | sys_id: ' + app.sys_id);
} else {
    gs.error('HR Leave app not found — create via 01_DEVELOPMENT.md first');
}

var tables = ['x_786686_hr_leave_req_v2', 'x_786686_hr_leave_bal_v2',
              'x_786686_hr_leave_employee_v2', 'x_786686_hr_leave_audit_v2'];
tables.forEach(function (t) {
    var gr = new GlideRecord(t);
    gs.info(t + ' exists: ' + gr.isValid());
});

Part 1 — Plugin Activation & Roles

Step 1.1: Activate required plugins

Navigation: All > System Definition > Plugins — Open direct
Direct URL: {BASE}/v_plugin_list.do

Detailed activation steps:

1. Open {BASE}/v_plugin_list.do
2. Filter: Name contains "Now Assist" OR "AI Search" OR "Virtual Agent"
3. For each plugin in the table above:
   a. Click the plugin name
   b. Click "Activate" (or "Install" on some instances)
   c. UNCHECK "Load demo data" in production-like sub-prod
   d. Wait for activation (may take 2–10 minutes)
   e. Note any dependent plugins auto-activated
4. After all activations: logout and login (refresh session roles)
5. Open Now Assist Setup wizard: {BASE}/now/assist/admin
6. Accept terms → select data region → Enable HRSD skill pack
7. Verify: All > Now Assist > Now Assist Admin shows "Configured"

Step 1.2: Assign AI admin roles

Navigation: All > User Administration > Users — Open direct
Direct URL: {BASE}/sys_user_list.do

Step 1.3: Create system properties (AI config)

Navigation: All > System Properties — Open direct
Direct URL: {BASE}/sys_properties_list.do

Create these properties in scope x_786686_hr_leave:

Verification script:

var props = ['ai.enabled', 'llm.daily_budget_usd', 'ai_search.profile'];
props.forEach(function (p) {
    var v = gs.getProperty('x_786686_hr_leave.' + p, 'NOT SET');
    gs.info('x_786686_hr_leave.' + p + ' = ' + v);
});

Part 2 — Now Assist Studio (Skills)

2.1 Architecture overview

Step 2.1: Open Now Assist Studio

URL: {BASE}/now/assist/studio

1. App picker (top-left) → select "HR Leave" scope
2. Click "Skills" in left nav
3. Click "New" → "Create Skill"
4. Confirm Application = x_786686_hr_leave on every record

Step 2.2: Skill 1 — hrlx_summarize_reason

Prompt template (SYSTEM + USER):

SYSTEM:
You are an HR assistant for a multinational company. Summarize the employee's
leave reason in ONE sentence, maximum 25 words, neutral professional tone.
Do NOT include medical diagnoses, medication names, or sensitive health details.
If reason is empty, return exactly: "No reason provided."
Respond in the same language as the reason text (English or Arabic).

USER:
Employee: {{employee_name}}
Leave type: {{u_leave_type}}
Start: {{u_start_date}} | End: {{u_end_date}} | Days: {{u_days_count}}
Reason:
"""
{{u_reason}}
"""

Variable mapping:

Grounding config:

1. Skill > Grounding tab > Enable
2. Table: x_786686_hr_leave_req_v2
3. Key: sys_id
4. Map each {{variable}} to column above
5. Save

Test in Playground (5 samples):

Playground steps:
1. Skills > hrlx_summarize_reason > Test
2. Paste sys_id of test request OR fill variables manually
3. Run → check latency (target ≤ 2s)
4. If bad output → edit prompt → Save as version 1.1
5. Repeat until 90%+ acceptable on 20 samples
6. Status → Active

Step 2.3: Skill 2 — hrlx_classify_leave_type

SYSTEM:
Classify the leave request text into EXACTLY ONE label:
annual, sick, parental, unpaid
Return ONLY the label, lowercase, no punctuation or explanation.

USER:
Leave type already selected (may be wrong): {{u_leave_type}}
Employee reason: {{u_reason}}
Additional notes: {{u_notes}}

Wire to Business Rule (before insert/update when u_leave_type empty):

(function executeRule(current, previous) {
    if (!gs.getProperty('x_786686_hr_leave.ai.enabled', 'true') === 'true') return;
    if (current.u_leave_type) return;
    if (!current.u_reason) return;

    try {
        var ext = new sn_one_extend.OneExtendUtil();
        var res = ext.execute({
            executionRequests: [{
                capabilityId: 'hrlx_classify_leave_type',
                payload: {
                    u_leave_type: String(current.u_leave_type),
                    u_reason: String(current.u_reason),
                    u_notes: String(current.u_notes || '')
                }
            }]
        });
        var label = (res.capabilities[0] || {}).response;
        if (/^(annual|sick|parental|unpaid)$/.test(label))
            current.u_leave_type = label;
    } catch (e) {
        gs.warn('hrlx_classify_leave_type failed: ' + e);
    }
})(current, previous);

Step 2.4: Skill 3 — hrlx_suggest_dates

SYSTEM:
Suggest adjusted leave dates avoiding public holidays.
Return ONLY valid JSON:
{"start":"YYYY-MM-DD","end":"YYYY-MM-DD","notes":"..."}
No markdown fences.

USER:
Country: {{u_country}}
Requested start: {{u_start_date}}
Requested end: {{u_end_date}}
Leave type: {{u_leave_type}}
Public holidays next 90 days:
{{holidays_json}}

Pre-step Script Include to build holidays_json:

var HRLXHolidayUtil = Class.create();
HRLXHolidayUtil.prototype = {
    initialize: function() {},
    getHolidaysJson: function(country, days) {
        // Replace with your holiday table or IntegrationHub spoke
        var out = [];
        var gr = new GlideRecord('sn_hr_core_federal_holiday'); // if HRSD installed
        if (gr.isValid()) {
            gr.addQuery('country', country);
            gr.addQuery('date', '>=', new GlideDate());
            gr.setLimit(days || 90);
            gr.query();
            while (gr.next()) {
                out.push({ date: String(gr.date), name: String(gr.name) });
            }
        }
        return JSON.stringify(out);
    },
    type: 'HRLXHolidayUtil'
};

Step 2.5: Skill 4 — hrlx_summarize_history

Summarizes last N leave requests for a manager email.

SYSTEM:
Summarize this employee's leave history for a manager briefing.
Max 150 words. Bullet format. Include totals by type. Neutral tone.

USER:
Employee: {{employee_name}}
Lookback days: {{lookback_days}}
History records:
{{history_text}}

Step 2.6: Skill 5 — hrlx_detect_sentiment

SYSTEM:
Classify employee leave reason sentiment: positive, neutral, frustrated, urgent.
Return ONLY the label.

USER:
Reason: {{u_reason}}

Use output to route frustrated → human HR in Flow Designer.

Step 2.7: Skill 6 — hrlx_draft_manager_email

SYSTEM:
Draft a short email from HR system to employee about their leave request.
Professional, empathetic. Max 120 words. Include request number and dates.
Do not invent approval status — use {{u_state}} exactly.

USER:
Employee name: {{employee_name}}
Request: {{u_number}}
Type: {{u_leave_type}}
Dates: {{u_start_date}} to {{u_end_date}}
State: {{u_state}}
Manager note: {{manager_comment}}

Step 2.8: Activate skills per channel

URL: {BASE}/sys_one_extend_activation_list.do

For each skill create activation:

Step 2.9: UI Action — Summarize on form

URL: {BASE}/sys_ui_action_list.do

(function executeAction(current) {
    var ext = new sn_one_extend.OneExtendUtil();
    var res = ext.execute({
        executionRequests: [{
            capabilityId: 'hrlx_summarize_reason',
            payload: { record: current.getUniqueValue() }
        }]
    });
    var summary = (res.capabilities[0] || {}).response || 'No summary';
    gs.addInfoMessage('AI Summary: ' + summary);
})(current);

Part 3 — AI Agent Studio

3.1 What AI Agent Studio provides

AI Agent Studio builds agentic assistants: an LLM that plans steps, selects tools, executes them, observes results, and responds — across multiple conversation turns.

Direct URL: {BASE}/now/aia/studio

Step 3.1: Create Script Include HRLXAIAgentTools

URL: {BASE}/sys_script_include_list.do
Application: HR Leave

var HRLXAIAgentTools = Class.create();
HRLXAIAgentTools.prototype = {
    initialize: function() {
        this.SCOPE = 'x_786686_hr_leave';
        this.T_EMP = 'x_786686_hr_leave_employee_v2';
        this.T_REQ = 'x_786686_hr_leave_req_v2';
        this.T_BAL = 'x_786686_hr_leave_bal_v2';
        this.T_AUD = 'x_786686_hr_leave_audit_v2';
    },

    _getEmployeeByUser: function(userId) {
        var gr = new GlideRecord(this.T_EMP);
        gr.addQuery('user', userId);
        gr.setLimit(1);
        gr.query();
        return gr.next() ? gr : null;
    },

    getBalance: function(userId) {
        var emp = this._getEmployeeByUser(userId);
        if (!emp) return { error: 'employee_not_found' };
        var out = { annual: 0, sick: 0, parental: 0, unpaid: 0 };
        var bal = new GlideRecord(this.T_BAL);
        bal.addQuery('u_employee_ref', emp.getUniqueValue());
        bal.query();
        while (bal.next()) {
            var t = String(bal.u_leave_type);
            if (out.hasOwnProperty(t)) out[t] = Number(bal.u_remaining_days) || 0;
        }
        return out;
    },

    requestLeave: function(payload) {
        try {
            if (gs.getProperty(this.SCOPE + '.ai.enabled', 'true') !== 'true')
                return { ok: false, error: 'ai_disabled' };

            var emp = this._getEmployeeByUser(payload.user_id);
            if (!emp) return { ok: false, error: 'employee_not_found' };

            var start = new GlideDateTime(payload.start_date);
            var end = new GlideDateTime(payload.end_date);
            if (start.compareTo(end) > 0)
                return { ok: false, error: 'start_after_end' };

            var bal = this.getBalance(payload.user_id);
            var days = Math.ceil((end.getNumericValue() - start.getNumericValue()) / 86400000) + 1;
            if ((bal[payload.type] || 0) < days)
                return { ok: false, error: 'insufficient_balance', needed: days, available: bal[payload.type] || 0 };

            var req = new GlideRecord(this.T_REQ);
            req.initialize();
            req.u_requested_for_user = payload.user_id;
            req.u_employee_ref = emp.getUniqueValue();
            req.u_leave_type = payload.type;
            req.u_start_date = payload.start_date;
            req.u_end_date = payload.end_date;
            req.u_reason = payload.reason || '';
            req.u_state = 'submitted';
            req.u_source = 'ai_agent';
            var sysId = req.insert();
            this._audit('ai_request_leave', sysId, payload);
            return { ok: true, request_id: sysId, number: req.getValue('u_number') };
        } catch (e) {
            gs.error('HRLXAIAgentTools.requestLeave: ' + e);
            return { ok: false, error: 'exception' };
        }
    },

    checkPolicy: function(question) {
        var profile = gs.getProperty(this.SCOPE + '.ai_search.profile', 'hrlx_kb_profile');
        try {
            var sn = new sn_ais.AISearch();
            var results = sn.search(profile, question, { limit: 5, genai: true });
            return {
                answer: results.genai_answer || '',
                citations: results.citations || []
            };
        } catch (e) {
            return { answer: '', citations: [], error: String(e) };
        }
    },

    escalate: function(userId, summary) {
        var emp = this._getEmployeeByUser(userId);
        if (!emp) return { ok: false, error: 'employee_not_found' };
        var hr = new GlideRecord('sn_hr_core_case');
        if (!hr.isValid()) {
            // Fallback: create audit + notify
            this._audit('ai_escalate_no_hrsd', userId, { summary: summary });
            return { ok: true, case_id: null, note: 'HRSD not installed — audit only' };
        }
        hr.initialize();
        hr.opened_for = userId;
        hr.short_description = 'AI escalation: ' + (summary || '').substring(0, 80);
        hr.description = summary;
        hr.subject_person = userId;
        var sysId = hr.insert();
        this._audit('ai_escalate', sysId, { summary: summary });
        return { ok: true, case_id: sysId };
    },

    getMyRequests: function(userId, days) {
        var emp = this._getEmployeeByUser(userId);
        if (!emp) return [];
        var since = new GlideDateTime();
        since.addDaysLocalTime(-1 * (days || 90));
        var gr = new GlideRecord(this.T_REQ);
        gr.addQuery('u_employee_ref', emp.getUniqueValue());
        gr.addQuery('sys_created_on', '>=', since);
        gr.orderByDesc('sys_created_on');
        gr.query();
        var out = [];
        while (gr.next()) {
            out.push({
                id: gr.getUniqueValue(),
                number: String(gr.u_number),
                type: String(gr.u_leave_type),
                start: String(gr.u_start_date),
                end: String(gr.u_end_date),
                state: String(gr.u_state)
            });
        }
        return out;
    },

    cancelRequest: function(userId, requestId) {
        var gr = new GlideRecord(this.T_REQ);
        if (!gr.get(requestId)) return { ok: false, error: 'not_found' };
        if (gr.u_requested_for_user != userId) return { ok: false, error: 'not_owner' };
        if (gr.u_state != 'submitted') return { ok: false, error: 'not_cancellable' };
        gr.u_state = 'cancelled';
        gr.update();
        this._audit('ai_cancel_request', requestId, { user: userId });
        return { ok: true };
    },

    _audit: function(action, recordId, payload) {
        var a = new GlideRecord(this.T_AUD);
        a.initialize();
        a.u_action = action;
        a.u_record = recordId;
        a.u_payload = JSON.stringify(payload || {});
        a.u_actor = gs.getUserID();
        a.u_source = 'ai_agent';
        a.insert();
    },

    type: 'HRLXAIAgentTools'
};

Step 3.2: Define tools in AI Agent Studio

URL: {BASE}/sn_aia_tool_list.do

Tool 1 — GetMyBalance

Tool 2 — RequestLeave

Confirmation gate: Enable “Require user confirmation” for this tool in Agent Studio.

Tool 3 — CheckPolicy

Tool 4 — GetMyRequests

| Method | getMyRequests |
| Input | user_id, days (default 90) |

Tool 5 — CancelRequest

| Method | cancelRequest |
| Input | user_id, request_id |

Tool 6 — EscalateToHR

| Method | escalate |
| Input | user_id, summary |

Step 3.3: Create agent “HR Leave Assistant”

URL: {BASE}/sn_aia_agent_list.do

System prompt:

You are "HR Leave Assistant" for employees of this company.

CAPABILITIES (use tools):
- GetMyBalance: show remaining leave days
- RequestLeave: submit a new request (ALWAYS confirm dates/type first)
- CheckPolicy: answer policy questions from official KB (cite sources)
- GetMyRequests: list recent requests
- CancelRequest: cancel a submitted request owned by the user
- EscalateToHR: create HR case when user is frustrated or issue is complex

RULES:
1. Identity comes from session — never ask for employee ID number.
2. NEVER reveal another employee's data.
3. Out of scope → "I can only help with leave. Contact HR at [email protected]."
4. Match user language (English or Arabic).
5. Before RequestLeave: summarize dates, type, days count, ask "Shall I submit?"
6. If tool returns error insufficient_balance, explain clearly with numbers.
7. For medical/sick policy, remind user a certificate may be required.

STYLE: concise, friendly, professional. Use bullet lists for balances.

Agent JSON export reference:

{
  "name": "HR Leave Assistant",
  "scope": "x_786686_hr_leave",
  "model": "now-llm-pro-v1",
  "tools": ["GetMyBalance", "RequestLeave", "CheckPolicy", "GetMyRequests", "CancelRequest", "EscalateToHR"],
  "knowledge_sources": [{ "type": "ai_search_profile", "name": "hrlx_kb_profile" }],
  "triggers": [
    { "type": "chat", "channel": "employee_center" },
    { "type": "chat", "channel": "service_portal" }
  ],
  "memory": { "session": true, "persistent": { "enabled": true, "max_turns": 8 } },
  "guardrails": {
    "pii_redaction": true,
    "escalation_keywords": ["urgent", "complaint", "discrimination", "harassment", "lawyer"]
  },
  "audit": { "store_inputs": true, "store_outputs": true, "retention_days": 365 }
}

Step 3.4: Enable agent on Employee Center

1. Open {BASE}/now/assist/admin
2. Channels > Employee Center
3. Enable "AI Agents"
4. Add agent: HR Leave Assistant
5. Position: HR Leave category in chat launcher
6. Save
7. Test: {BASE}/esc → open chat → "How many annual days do I have?"

Step 3.5: Test agent (10 test utterances)

Part 4 — BYO LLM (Claude / GPT)

Step 4.1: When to use BYO LLM

Step 4.2: Configure Anthropic provider

URL: {BASE}/sys_one_extend_provider_config_list.do

1. New > Provider type: Anthropic (or Generic REST)
2. Name: hrlx_anthropic
3. API endpoint: https://api.anthropic.com/v1/messages
4. Credential: link to Generic API Key credential (x-api-key)
5. Default model: claude-sonnet-4-5
6. Test connection
7. Per skill: LLM Provider = hrlx_anthropic

Step 4.3: REST Message — Claude_API

URL: {BASE}/sys_rest_message_list.do

Body template:

{
  "model": "${model}",
  "max_tokens": ${max_tokens},
  "system": "${system}",
  "messages": [{ "role": "user", "content": "${user_message}" }]
}

Step 4.4: Script Include — HRLXAIProvider (fallback chain)

var HRLXAIProvider = Class.create();
HRLXAIProvider.prototype = {
    initialize: function() {
        this.SCOPE = 'x_786686_hr_leave';
        this.TIMEOUT_MS = parseInt(gs.getProperty(this.SCOPE + '.llm.timeout_ms', '15000'), 10);
    },

    _prop: function(name, def) { return gs.getProperty(this.SCOPE + '.' + name, def); },

    _budgetCheck: function(provider) {
        var budget = parseFloat(this._prop('llm.daily_budget_usd', '10'));
        var key = 'llm.spent.' + provider + '.' + new GlideDateTime().getDate().getDisplayValue();
        return parseFloat(gs.getProperty(this.SCOPE + '.' + key, '0')) < budget;
    },

    askClaude: function(prompt, model, maxTokens, system) {
        if (!this._budgetCheck('anthropic')) return { error: 'budget_exceeded' };
        try {
            var rm = new sn_ws.RESTMessageV2('Claude_API', 'post');
            rm.setHttpTimeout(this.TIMEOUT_MS);
            rm.setStringParameterNoEscape('model', model || this._prop('llm.anthropic.model', 'claude-sonnet-4-5'));
            rm.setStringParameterNoEscape('max_tokens', String(maxTokens || 400));
            rm.setStringParameterNoEscape('system', system || 'You are a helpful HR assistant.');
            rm.setStringParameterNoEscape('user_message', prompt);
            var resp = rm.execute();
            if (resp.getStatusCode() >= 400) return { error: 'http_' + resp.getStatusCode() };
            var json = JSON.parse(resp.getBody());
            var text = '';
            (json.content || []).forEach(function(c) { if (c.type === 'text') text += c.text; });
            return { text: text, usage: json.usage || {} };
        } catch (e) { return { error: 'exception', message: String(e) }; }
    },

    askWithFallback: function(prompt, options) {
        options = options || {};
        var chain = options.chain || ['anthropic', 'openai', 'now_llm'];
        for (var i = 0; i < chain.length; i++) {
            var res = (chain[i] === 'anthropic') ? this.askClaude(prompt, options.model, options.max_tokens, options.system) : null;
            if (res && !res.error && res.text) return res;
        }
        return { error: 'all_providers_failed' };
    },

    type: 'HRLXAIProvider'
};

Step 4.5: HRLXRagService (custom RAG)

var HRLXRagService = Class.create();
HRLXRagService.prototype = {
    initialize: function() {
        this.PROVIDER = new HRLXAIProvider();
        this.PROFILE = gs.getProperty('x_786686_hr_leave.ai_search.profile', 'hrlx_kb_profile');
    },

    answer: function(question) {
        var sn = new sn_ais.AISearch();
        var res = sn.search(this.PROFILE, question, { limit: 5, genai: true, min_similarity: 0.65 });
        if (res.genai_answer) return { text: res.genai_answer, citations: res.citations || [] };
        return { text: 'I could not find an answer in HR Leave policies.', citations: [] };
    },

    type: 'HRLXRagService'
};

Part 5 — AI Search + GenAI Search

5.1 AI Search vs Zing

Step 5.1: Create search source — hrlx_kb

URL: {BASE}/ais_search_source_list.do

1. New Search Source
2. Name: hrlx_kb
3. Table: kb_knowledge
4. Condition: topic=HR Leave^workflow_state=published^active=true
5. Indexed fields:
   - title (boost 2.0)
   - short_description (boost 1.5)
   - text (boost 1.0)
   - meta (boost 0.5)
6. Index transform script (PII mask):

(function transform(record, indexed) {
    indexed.title = record.short_description + '';
    var body = (record.text || '') + '';
    indexed.body = body.replace(/\b(SSN|PASSPORT|MRN|diagnosis)\b[: ]?\S+/gi, '[REDACTED]');
    indexed.u_kb_number = record.number + '';
    return indexed;
})(record, indexed);

7. Save → Full Index (wait for completion)
8. Verify: AI Search Admin > Index Status = Green

Step 5.2: Create source — hrlx_policy_attachments

Step 5.3: Create profile — hrlx_kb_profile

URL: {BASE}/ais_search_profile_list.do

Step 5.4: Create search application (Portal)

URL: {BASE}/ais_search_application_list.do

1. New Application: hrlx_portal_search
2. Profile: hrlx_kb_profile
3. UI: Service Portal
4. Placeholder: "Ask about leave policy, sick days, carry-over..."
5. Enable GenAI answer panel + citations
6. Link to page: {BASE}/sp?id=hrlx_policy_search

Widget HTML (Service Portal):

<div class="hrlx-aisearch">
  <sp-search
    profile="hrlx_kb_profile"
    placeholder="Ask about leave policy"
    show-genai-answer="true"
    show-citations="true">
  </sp-search>
</div>

Step 5.5: Boost rules & synonyms

URL for synonyms: {BASE}/ais_dictionary_term_list.do

Step 5.6: Search analytics dashboard

Build PA indicators from:

Target: zero-result rate < 5% after 30 days.

Part 6 — Predictive Intelligence

6.1 Solution types for HR Leave

Step 6.1: Add ML output fields to req_v2

URL: {BASE}/sys_dictionary_list.do?sysparm_query=name=x_786686_hr_leave_req_v2

Step 6.2: Backfill training label (one-time script)

Run after you have historical approvals:

var gr = new GlideRecord('x_786686_hr_leave_req_v2');
gr.addQuery('u_state', 'IN', 'approved,rejected');
gr.query();
while (gr.next()) {
    var created = new GlideDateTime(gr.sys_created_on);
    var updated = new GlideDateTime(gr.sys_updated_on);
    var hours = (updated.getNumericValue() - created.getNumericValue()) / 3600000;
    var approved = gr.u_state == 'approved' && hours <= 24;
    gr.u_approved_within_24h = approved ? 'yes' : 'no';
    gr.setWorkflow(false);
    gr.update();
}
gs.info('Backfill complete');

Minimum training rows: ≥ 1,000 historical requests (≥ 100 per class for classification).

Step 6.3: Create classification solution — hrlx_predict_approval_24h

URL: {BASE}/ml_solution_definition_list.do

1. New Solution Definition
2. Name: hrlx_predict_approval_24h
3. Table: x_786686_hr_leave_req_v2
4. Solution type: Classification
5. Input fields:
   - u_leave_type
   - u_days_count
   - u_department (from employee)
   - u_employee_tenure_days (calculated field or BR)
   - day_of_week (derived from u_start_date)
   - u_has_reason (boolean: u_reason length > 5)
6. Output field: u_approved_within_24h
7. Prediction fields: u_predicted_approval_24h, u_predicted_confidence_24h
8. Training window: last 18 months
9. Exclusion: u_state=draft
10. Click Train → wait (15–60 min)
11. Review accuracy report: target ≥ 0.80 accuracy, F1 ≥ 0.65 minority class
12. Activate solution

Step 6.4: Business Rule — predict on insert

(function executeRule(current, previous) {
    if (current.u_predicted_approval_24h) return;
    try {
        var ml = new sn_ml.SolutionStore().get('hrlx_predict_approval_24h');
        if (!ml || !ml.isActive()) return;

        var startDt = new GlideDateTime(current.u_start_date);
        var input = {
            u_leave_type: String(current.u_leave_type),
            u_days_count: parseInt(current.u_days_count || '0', 10),
            u_department: String(current.u_department || ''),
            u_employee_tenure_days: parseInt(current.u_employee_tenure_days || '0', 10),
            day_of_week: startDt.getDayOfWeekUTC(),
            u_has_reason: (current.u_reason && current.u_reason.length > 5) ? 1 : 0
        };
        var pred = ml.predict([input]);
        if (pred && pred[0]) {
            current.u_predicted_approval_24h = pred[0].predicted_value;
            current.u_predicted_confidence_24h = pred[0].confidence;
        }
    } catch (e) {
        gs.warn('PI predict failed: ' + e);
    }
})(current, previous);

Step 6.5: Regression solution — hrlx_predict_duration

Same wizard, type = Regression, output = u_days_count (actual), prediction = u_predicted_duration.

Step 6.6: Solution Monitor

URL: {BASE}/ml_solution_monitor_list.do

Schedule monthly re-train via Scheduled Script Execution.

Part 7 — Virtual Agent + NLU (Full Topics & Flows)

7.1 Virtual Agent vs AI Agent — coexistence

Recommended pattern: VA handles structured flows; fallback / policy Q&A routes to AI Agent.

Step 7.1: Create NLU model — hrlx_leave_nlu

URL: {BASE}/now/nlu/workbench

Step 7.2: Intents (8 total)

Step 7.3: Entities

Step 7.4: Utterances — check_balance (30 examples)

how many leave days do I have
what is my annual leave balance
show my remaining vacation days
how much PTO do I have left
check my sick leave balance
remaining parental leave
how many days off can I take
what's my leave balance
do I have any unpaid leave left
show balance for annual leave
how many annual days remaining
tell me my vacation balance
what leave do I have available
check my time off balance
how much annual leave is left
show me sick days remaining
balance inquiry for leave
how many days of leave remain
what is my current leave entitlement
remaining days for sick leave
how much leave have I got left
annual balance please
sick leave days left
parental leave balance check
unpaid leave remaining
how many holidays can I still take
leave days available
my remaining leave
check PTO balance
how many days can I still book off

Step 7.5: Utterances — request_leave (35 examples)

I want to take leave next Monday
request 3 days annual leave
book sick leave for tomorrow
I need time off from June 5 to June 10
apply for parental leave starting next month
request vacation days
take 2 days off next week
I want to book annual leave
schedule sick leave for Friday
need unpaid leave for 5 days
book me 3 days annual starting Monday
I would like to request leave
can I take leave next Tuesday and Wednesday
request time off for a wedding
apply for 10 days annual leave in July
I need sick leave for 2 days
book parental leave from September 1
request leave for family emergency
take annual leave Dec 20 to Dec 31
I want 1 week off
schedule 4 days sick leave
request unpaid leave next month
book leave type annual for 3 days
I need to take Friday off
request half day sick leave
apply for leave starting {date_range}
book {days_count} days of {leave_type} leave
I want {leave_type} leave from {date_range}
need {days_count} days off
request leave please
take time off
book my vacation
I want to go on leave
schedule my days off
can I have next week off

Step 7.6: Utterances — check_policy (25 examples)

what is the maternity leave policy
can I carry over unused annual leave
how many sick days am I allowed per year
what is the parental leave entitlement
do I need a medical certificate for sick leave
what is the notice period for annual leave
can managers reject my leave request
what happens to unused leave at year end
is unpaid leave available
what is the maximum consecutive leave days
how does sick leave accrue
what is the leave policy for new joiners
can I take leave during probation
what documents are required for sick leave
explain the annual leave policy
how far in advance must I request leave
what is the carry forward limit
are public holidays deducted from leave
what is bereavement leave policy
can I split parental leave
what is the leave approval process
how is leave calculated for part time staff
what is the sick leave waiting period
tell me about leave encashment
what are the rules for emergency leave

Step 7.7: Utterances — cancel_leave (20 examples)

cancel my leave request
withdraw my leave application
I want to cancel HRLX0001234
cancel my upcoming vacation
withdraw the leave I submitted
delete my leave request
cancel submitted leave
I no longer need those days off
please cancel my time off request
revoke my leave request
cancel annual leave for next week
withdraw sick leave request
cancel request number {request_number}
I want to undo my leave request
cancel my pending leave
stop my leave application
cancel the leave I booked
withdraw my days off
cancel leave starting Monday
I need to cancel my vacation request

Step 7.8: Utterances — check_status (20 examples)

what is the status of my leave request
has my leave been approved
track my leave application
where is my request HRLX0001234
is my leave approved yet
check approval status
status of my vacation request
has manager approved my leave
what happened to my leave request
show status of pending leave
is my sick leave approved
check my submitted leave
approval status please
where is my annual leave request
has HR approved my leave
track request {request_number}
what state is my leave in
is my leave still pending
did my leave get rejected
update on my leave request

Step 7.9: Train and publish NLU

1. NLU Workbench > hrlx_leave_nlu
2. Add all utterances above (minimum 15 per intent; we have 20–35)
3. Annotate entities in utterances (highlight dates, types)
4. Click Train → wait for model metrics
5. Target: Intent accuracy ≥ 0.92 on test set
6. Publish model version 1.0
7. Set property x_786686_hr_leave.nlu.model = hrlx_leave_nlu

Step 7.10: Topic catalog (8 topics)

URL: {BASE}/now/virtual-agent/designer

TOPIC FLOW 1 — HRLX Check Balance (block-by-block)

Topic: HRLX Check Balance
Channel: Employee Center, Service Portal
NLU: check_balance (confidence ≥ 0.7)

BLOCK 1 — Script Action "Fetch Balance"
  Script:

(function execute() {
    var userId = vaSystem.getUserId();
    var tools = new x_786686_hr_leave.HRLXAIAgentTools();
    var bal = tools.getBalance(userId);
    if (bal.error) {
        vaInputs.has_error = true;
        vaInputs.error_msg = 'Employee profile not found. Contact HR.';
        return;
    }
    vaInputs.has_error = false;
    vaInputs.balance_text =
        'Annual: ' + bal.annual + ' days\n' +
        'Sick: ' + bal.sick + ' days\n' +
        'Parental: ' + bal.parental + ' days\n' +
        'Unpaid: ' + bal.unpaid + ' days';
})();

BLOCK 2 — Decision
  Condition: has_error == true
  True → Bot Message: "${error_msg}" → End
  False → continue

BLOCK 3 — Bot Message
  Text: "Here is your current leave balance:\n${balance_text}"

BLOCK 4 — Quick Reply
  Question: "Would you like to request leave?"
  Options: Yes | No

BLOCK 5 — Branch
  Yes → Link to Topic "HRLX Request Leave"
  No → Bot Message "Have a great day!" → End

TOPIC FLOW 2 — HRLX Request Leave (slot filling)

Topic: HRLX Request Leave
NLU: request_leave

BLOCK 1 — Collect leave_type (if not in NLU entities)
  Prompt: "Which type of leave? (annual / sick / parental / unpaid)"
  Slot: leave_type
  Validation: must be in [annual, sick, parental, unpaid]
  Max retries: 2 → on fail go to Fallback

BLOCK 2 — Collect start_date
  Prompt: "What is the start date? (YYYY-MM-DD or 'next Monday')"
  Slot: start_date
  Validation: date >= today

BLOCK 3 — Collect end_date
  Prompt: "What is the end date?"
  Slot: end_date
  Validation: end_date >= start_date

BLOCK 4 — Collect reason (optional)
  Prompt: "Reason for leave? (optional — press Skip)"
  Slot: reason
  Required: false

BLOCK 5 — Script Action "Calculate Days Preview"

(function execute() {
    var start = new GlideDateTime(vaVars.start_date);
    var end = new GlideDateTime(vaVars.end_date);
    var days = Math.ceil((end.getNumericValue() - start.getNumericValue()) / 86400000) + 1;
    vaInputs.preview_days = days;
    var tools = new x_786686_hr_leave.HRLXAIAgentTools();
    var bal = tools.getBalance(vaSystem.getUserId());
    vaInputs.enough = (bal[vaVars.leave_type] || 0) >= days;
    vaInputs.available = bal[vaVars.leave_type] || 0;
})();

BLOCK 6 — Decision insufficient balance
  Condition: enough == false
  True → Message: "You need ${preview_days} days but only have ${available}. Choose shorter dates."
         → Go to BLOCK 2 (re-collect dates)
  False → continue

BLOCK 7 — Confirmation
  Message: "Please confirm:\nType: ${leave_type}\nFrom: ${start_date} To: ${end_date}\nDays: ${preview_days}\nReason: ${reason}"
  Quick Reply: Confirm | Cancel

BLOCK 8 — Branch Cancel → End polite

BLOCK 9 — Script Action "Submit Request"

(function execute() {
    var tools = new x_786686_hr_leave.HRLXAIAgentTools();
    var res = tools.requestLeave({
        user_id: vaSystem.getUserId(),
        type: vaVars.leave_type,
        start_date: vaVars.start_date,
        end_date: vaVars.end_date,
        reason: vaVars.reason || ''
    });
    vaInputs.ok = !!res.ok;
    vaInputs.req_number = res.number || '';
    vaInputs.err = res.error || '';
})();

BLOCK 10 — Decision ok
  True → "Your leave request ${req_number} has been submitted for approval."
  False → "Sorry, submission failed: ${err}. Try again or contact HR."
  → End

TOPIC FLOW 3 — HRLX Check Policy

BLOCK 1 — Decision: use AI Agent for open questions
  If utterance length > 50 chars OR contains "?"
    → Transfer to AI Agent "HR Leave Assistant" with context
  Else
    → Script Action CheckPolicy with vaInputs.last_utterance

BLOCK 2 — Display answer + citations
BLOCK 3 — "Was this helpful?" Yes/No → log to analytics

TOPIC FLOW 4 — HRLX Cancel Leave

BLOCK 1 — Collect request_number (or pick from list)
  Script Action: getMyRequests → build choice list

BLOCK 2 — Confirm cancellation

BLOCK 3 — Script Action cancelRequest

BLOCK 4 — Success/fail message

TOPIC FLOW 5 — HRLX Live Agent Handoff

BLOCK 1 — Collect issue summary (optional)

BLOCK 2 — Live Agent Queue
  Queue: HR Leave Support
  Pre-chat fields: user_id, last_intent, last_5_messages

BLOCK 3 — Fallback if no agent available
  Script Action escalate()
  Message: "An HR partner will contact you. Case reference logged."

Step 7.11: Enable VA on channels

1. Virtual Agent > Administration > Channels
2. Employee Center: ON, Bot name from x_786686_hr_leave.va.bot_name
3. Service Portal: add Virtual Agent widget to sp_page
4. Topic recommendations: enable HR Leave topics
5. Test: {BASE}/esc

Step 7.12: VA analytics KPIs

Part 8 — Document Intelligence (IDP)

8.1 Use case — medical certificate for sick leave

When u_leave_type = sick and u_days_count > 2, require medical certificate attachment. IDP extracts fields automatically.

Step 8.1: Create IDP template — hrlx_medical_cert

URL: {BASE}/sn_idp_template_list.do

1. New Template
2. Name: hrlx_medical_cert
3. Document type: Medical Certificate
4. Upload 8–12 sample PDFs (anonymized):
   - 3 hospital letterheads
   - 3 clinic formats
   - 2 Arabic certificates
   - 2 poor-scan quality (for robustness)
5. For each sample, annotate bounding boxes:
   - patient_name
   - certificate_date
   - doctor_name
   - days_recommended
   - clinic_name
   - condition_summary
6. Save annotations

Step 8.2: Train extraction model

URL: {BASE}/sn_idp_model_list.do

1. New Model → Template: hrlx_medical_cert
2. Train/Test split: 80/20
3. Click Train (30–90 minutes)
4. Review per-field F1 scores (target ≥ 0.90 each)
5. If condition_summary F1 low: add more samples or mark optional
6. Activate model v1.0
7. Set property x_786686_hr_leave.idp.template = hrlx_medical_cert

Step 8.3: Flow — run IDP on attachment

URL: {BASE}/now/flow-designer

Flow: HRLX IDP Medical Certificate
Trigger: Attachment added to x_786686_hr_leave_req_v2
Condition: u_leave_type=sick^u_days_count>2

Step 1 — Get attachment sys_id
Step 2 — Action: Document Intelligence Extract
  Template: hrlx_medical_cert
  Input: attachment_sys_id
Step 3 — Script (update request):

(function(inputs, outputs) {
    var fields = inputs.extraction_result.fields || {};
    var req = new GlideRecord('x_786686_hr_leave_req_v2');
    if (!req.get(inputs.request_sys_id)) return;

    req.u_cert_date = fields.certificate_date || '';
    req.u_cert_days = fields.days_recommended || '';
    req.u_cert_doctor = fields.doctor_name || '';
    req.u_cert_clinic = fields.clinic_name || '';
    // Mask sensitive summary for display
    req.u_cert_summary = (fields.condition_summary || '').substring(0, 200);
    req.update();

    var aud = new GlideRecord('x_786686_hr_leave_audit_v2');
    aud.initialize();
    aud.u_action = 'idp_extract_cert';
    aud.u_record = req.getUniqueValue();
    aud.u_payload = JSON.stringify(fields);
    aud.u_actor = gs.getUserID();
    aud.u_source = 'idp';
    aud.insert();
})(inputs, outputs);

Step 4 — Decision: days_recommended < u_days_count → notify HR for review
Step 5 — Email employee if extraction failed (manual upload retry)

Step 8.4: Standalone IDP test

1. IDP Admin > Test Extraction
2. Upload new PDF not in training set
3. Verify all fields extracted
4. Latency target: ≤ 8 seconds per document

Step 8.5: IDP metrics dashboard

Part 9 — Workflow Studio AI Actions

Step 9.1: Built-in actions to reuse

Step 9.2: Custom action — HRLX Summarize Leave History

URL: {BASE}/now/flow-designer → Custom Actions

1. New Custom Action: HRLX Summarize Leave History
2. Application: HR Leave
3. Inputs:
   - employee_ref (Reference: x_786686_hr_leave_employee_v2)
   - lookback_days (Integer, default 365)
4. Outputs:
   - summary_text (String)
5. Steps:
   a. Look Up Records: x_786686_hr_leave_req_v2
      Query: u_employee_ref = employee_ref AND sys_created_on >= lookback
   b. Transform Data: build history_text lines
   c. Now Assist Skill: hrlx_summarize_history
   d. Return summary_text
6. Guardrails: PII redaction ON, max_tokens 250

Step 9.3: Flow — weekly manager digest

Trigger: Scheduled weekly Monday 08:00
For Each: managers with pending approvals
  Action: HRLX Summarize Leave History (direct reports, 90 days)
  Action: Send Email with summary + link to workspace

Step 9.4: Flow — auto-classify on submit

Trigger: Record created x_786686_hr_leave_req_v2, u_state=submitted
Action: Now Assist Skill hrlx_classify_leave_type
Decision: if output != u_leave_type → set u_ai_suggested_type + notify HR
Action: PI predict approval
Decision: if confidence < 0.6 → route to senior manager

Part 10 — Now Assist for Developers

Step 10.1: Generate Script Include from prompt

In Studio ({BASE}/$studio.do):

Prompt:
"Create Script Include HRLXLeaveValidation in scope x_786686_hr_leave.
Methods: validateDates(start, end), validateBalance(userId, type, days).
Use GlideRecord on x_786686_hr_leave_bal_v2. Return {ok, error} objects.
No client callable. Include null checks."

Review → Tab accept → add audit comment → ATF test.

Step 10.2: Inline shortcuts

Step 10.3: Generate ATF from Script Include

1. Open HRLXAIAgentTools in Studio
2. Now Assist > Generate Tests
3. Select: getBalance, requestLeave, cancelRequest
4. Save tests under sys_atf_test scoped to HR Leave
5. Run Test Runner: {BASE}/sys_atf_test_runner.do

Step 10.4: Code review checklist for AI-generated code

[ ] No hard-coded sys_ids
[ ] Scoped table names only (x_786686_hr_leave_*)
[ ] ACL-safe (respects current user)
[ ] No gs.print in production code
[ ] Error handling with gs.error/gs.warn
[ ] Human reviewer name in header comment

Part 11 — End-to-End HR Leave AI Example

Scenario: Employee submits sick leave via chat with certificate

Actors: Employee (ESS), HR Leave Assistant (AIA), VA topics, IDP, PI, Flow

Step 1 — Employee opens {BASE}/esc
Step 2 — "I need sick leave Thu-Fri" → VA Topic Request Leave OR AIA
Step 3 — Slot fill: type=sick, dates confirmed
Step 4 — AIA calls RequestLeave → req HRLX001234 created
Step 5 — Employee uploads medical PDF on form
Step 6 — Flow HRLX IDP Medical Certificate runs → fields extracted
Step 7 — BR runs PI → u_predicted_approval_24h = yes (0.87)
Step 8 — Manager opens record → UI Action Summarize reason (AI)
Step 9 — Manager approves → Flow sends hrlx_draft_manager_email (optional)
Step 10 — Audit rows in x_786686_hr_leave_audit_v2 for ai_request_leave, idp_extract_cert

Verification queries

// All AI-touched requests last 7 days
var gr = new GlideRecord('x_786686_hr_leave_req_v2');
gr.addQuery('u_source', 'ai_agent');
gr.addQuery('sys_created_on', '>=', gs.daysAgoStart(7));
gr.query();
gs.info('AI agent requests: ' + gr.getRowCount());

var aud = new GlideRecord('x_786686_hr_leave_audit_v2');
aud.addQuery('u_source', 'IN', 'ai_agent,idp');
aud.addQuery('sys_created_on', '>=', gs.daysAgoStart(7));
aud.query();
gs.info('AI audit rows: ' + aud.getRowCount());

Part 12 — AI Governance & Safety

Step 12.1: Create AI policy — no_raw_medical_text

URL: {BASE}/now/ai/policy

{
  "policy": "hrlx_no_raw_medical_text",
  "scope": "x_786686_hr_leave",
  "rules": [
    {
      "id": "redact_diagnosis",
      "applies_to": ["input", "output"],
      "regex": "(?i)(diagnosis|condition|illness|medication):\\s*[^\\n]+",
      "action": "redact",
      "replacement": "[REDACTED-MEDICAL]"
    },
    {
      "id": "block_external_emails",
      "applies_to": ["output"],
      "regex": "\\b[A-Z0-9._%+-]+@(?!company\\.com)[A-Z0-9.-]+\\.[A-Z]{2,}\\b",
      "action": "block"
    }
  ]
}

Step 12.2: Human-in-the-loop matrix

Step 12.3: Compliance notes

Step 12.4: Audit retention

Configure 365-day retention on sn_aia_audit and sys_one_extend_audit. Domain audit x_786686_hr_leave_audit_v2 retained per HR policy (typically 7 years).

Part 13 — Performance & Cost Optimization

Step 13.1: Response cache table

Create table x_786686_hr_leave_ai_cache (hash, skill, response, expires_at).

var HRLXAICache = Class.create();
HRLXAICache.prototype = {
    initialize: function() { this.TABLE = 'x_786686_hr_leave_ai_cache'; },
    _hash: function(s) { return new GlideDigest().getSHA256Base64(s); },
    get: function(skill, input) {
        var gr = new GlideRecord(this.TABLE);
        gr.addQuery('u_hash', this._hash(skill + '|' + input));
        gr.addQuery('u_expires_at', '>', new GlideDateTime());
        gr.setLimit(1);
        gr.query();
        return gr.next() ? String(gr.u_response) : null;
    },
    put: function(skill, input, response, ttlMin) {
        var gr = new GlideRecord(this.TABLE);
        gr.initialize();
        gr.u_hash = this._hash(skill + '|' + input);
        gr.u_skill = skill;
        gr.u_response = response;
        var exp = new GlideDateTime();
        exp.addSeconds(60 * (ttlMin || 60));
        gr.u_expires_at = exp;
        gr.insert();
    },
    type: 'HRLXAICache'
};

Use for hrlx_classify_leave_type (deterministic input).

Step 13.2: Model right-sizing

Step 13.3: Monitoring dashboard widgets

Part 14 — Hands-On Labs (5 Sessions)

Lab 1 — Now Assist Skill (45 min)

Goal: Deploy hrlx_summarize_reason with UI Action.

| Time | Step |
|------|------|
| 0–5 | Verify prerequisites (§0.2) |
| 5–15 | Create skill in Now Assist Studio (§2.2) |
| 15–25 | Test 5 samples in Playground |
| 25–35 | Create UI Action on req_v2 form |
| 35–40 | Impersonate manager → click Summarize |
| 40–45 | Check sys_one_extend_audit for log row |

Pass: Summary appears ≤ 2s, no medical detail leaked.
URL: {BASE}/now/assist/studio

Lab 2 — AI Search + GenAI (45 min)

Goal: Ask “Can I carry over annual leave?” and get cited answer.

| Time | Step |
|------|------|
| 0–10 | Create 3 KB articles (topic=HR Leave) if missing |
| 10–25 | Create source hrlx_kb + profile hrlx_kb_profile (§5) |
| 25–35 | Full index → wait green |
| 35–40 | Test in AI Search Admin preview |
| 40–45 | Add sp-search widget to portal page |

Pass: GenAI answer cites KB number.
URL: {BASE}/now/ais/admin

Lab 3 — Virtual Agent topic (60 min)

Goal: Complete balance → request leave conversation.

| Time | Step |
|------|------|
| 0–15 | Import NLU utterances for check_balance + request_leave |
| 15–25 | Train + publish hrlx_leave_nlu |
| 25–45 | Build topics per §7 block diagrams |
| 45–55 | Enable on {BASE}/esc |
| 55–60 | Run conversation test script (10 utterances) |

Pass: Request created with u_source=virtual_agent (add field if needed).
URL: {BASE}/now/virtual-agent/designer

Lab 4 — AI Agent + tools (60 min)

Goal: HR Leave Assistant submits leave after confirmation.

| Time | Step |
|------|------|
| 0–15 | Deploy HRLXAIAgentTools Script Include |
| 15–30 | Create 6 tools in Agent Studio |
| 30–45 | Create agent + system prompt |
| 45–55 | Enable on Employee Center |
| 55–60 | Test: "book 2 days annual next week" → confirm → verify req |

Pass: Record in req_v2, audit row written.
URL: {BASE}/now/aia/studio

Lab 5 — IDP + PI (90 min)

Goal: Sick leave + certificate extraction + approval prediction.

| Time | Step |
|------|------|
| 0–20 | Upload samples + annotate hrlx_medical_cert |
| 20–50 | Train IDP model (parallel: backfill PI labels) |
| 50–65 | Create Flow for attachment trigger |
| 65–80 | Train hrlx_predict_approval_24h |
| 80–90 | Submit sick leave with PDF → verify cert fields + prediction |

Pass: u_cert_date populated, u_predicted_approval_24h set.
URL: {BASE}/now/idp/admin , {BASE}/ml_solution_definition_list.do

Part 15 — Best Practices

15.1 Design

✅ One skill = one job (summarize OR classify, not both)
✅ Ground skills to specific tables with field-level mapping
✅ VA for happy-path slots; AIA for long-tail questions
✅ Always show citations for policy answers
✅ Confirm destructive tool calls (RequestLeave, CancelRequest)
✅ Log every AI action to x_786686_hr_leave_audit_v2
❌ Do not send cross-employee data to LLM
❌ Do not auto-approve leave based on PI alone
❌ Do not index u_reason without PII masking
❌ Do not skip human review on AI-generated code

15.2 Prompt engineering

✅ Separate SYSTEM (rules) from USER (data)
✅ Specify output format explicitly (JSON, single label, max words)
✅ Include refusal behavior ("If unsure, say...")
✅ Test in Arabic and English if bilingual workforce
✅ Version prompts; never edit production without version bump

15.3 Operations

✅ Weekly: review zero-result AI Search queries → add KB
✅ Monthly: review skill latency + cost dashboards
✅ Quarterly: re-train PI + NLU with new utterances
✅ Per release: run 5 labs as regression suite
✅ Maintain eval set (50 questions + expected answers) per skill

15.4 Security

✅ Mask medical text in prompts sent to external LLMs
✅ Use instance ACLs — tools run as authenticated user
✅ Budget caps on BYO LLM properties
✅ Disable x_786686_hr_leave.ai.enabled kill switch tested quarterly

Part 16 — Production Deployment Checklist

Phase A — Pre-deployment (sub-prod complete)

[ ] All 5 labs passed on sub-prod
[ ] Update Set "HR Leave AI v1.0" complete + peer reviewed
[ ] ATF tests green for HRLXAIAgentTools
[ ] Eval set ≥ 90% pass rate per skill
[ ] NLU intent accuracy ≥ 0.92
[ ] IDP field F1 ≥ 0.90
[ ] PI accuracy ≥ 0.80
[ ] AI Search zero-result rate < 10%
[ ] Agent containment ≥ 70% in UAT
[ ] Security review: PII redaction verified
[ ] Cost model approved (projected RU/month)
[ ] Rollback plan documented (disable ai.enabled property)

Phase B — Deployment day

[ ] Clone Update Set to production via DevOps/CD pipeline
[ ] Set production LLM API keys (password2 properties)
[ ] Activate plugins (if not already) — maintenance window
[ ] Now Assist Admin → enable skills for production scope
[ ] AI Search → full re-index on production
[ ] NLU model published to production instance
[ ] IDP model activated
[ ] PI solutions activated (monitor only first 24h)
[ ] Enable VA on Employee Center (feature flag gradual)
[ ] Enable AI Agent on Employee Center (pilot group first)
[ ] Verify kill switch: x_786686_hr_leave.ai.enabled = true
[ ] Smoke test: 10 utterances from §3.5 + §7
[ ] Monitor logs 2 hours: syslog, sn_aia_execution, sys_one_extend_audit

Phase C — Post-deployment (week 1)

[ ] Day 1: review all sn_aia_audit for errors
[ ] Day 2: review AI Search zero-result queries
[ ] Day 3: review VA containment + handoff rate
[ ] Day 5: review LLM spend vs budget
[ ] Day 7: stakeholder sign-off meeting
[ ] Document known issues + FAQ for HR helpdesk
[ ] Schedule monthly PI re-train
[ ] Schedule quarterly prompt review

Rollback procedure

1. Set x_786686_hr_leave.ai.enabled = false (immediate stop)
2. Disable AI Agent channel in Now Assist Admin
3. Disable VA topics (set inactive)
4. Deactivate PI solutions (optional — predictions stop)
5. Revert Update Set if Script Include defects found
6. Communicate to HR: "AI assistant temporarily offline"

Appendices

Appendix A — Key tables reference

Appendix B — Script Includes

Appendix C — REST Messages

Appendix D — Roles

Appendix E — Plugin quick search terms

Now Assist, AI Agents, AI Search, Predictive Intelligence,
Document Intelligence, Virtual Agent, NLU Workbench, Now Assist for Creator

Appendix F — Troubleshooting

Appendix G — Related documentation

End of guide. Start at Part 0 → pick your track (Skill, Agent, VA, Search, PI, IDP) → run matching Lab → complete Deployment Checklist.

للأسئلة: ابدأ من Quick Map → نفّذ Lab المناسب → راجع Troubleshooting (Appendix F).