diff --git a/run_scheduler.py b/run_scheduler.py
index 4b4b273..178c61c 100644
--- a/run_scheduler.py
+++ b/run_scheduler.py
@@ -127,6 +127,10 @@ def _register_schedules():
     schedule.every().day.at("00:05").do(run_cost_throttle_reset_if_new_month)
     logger.info("📅 每日 00:05：cost_throttle_reset_if_new_month")
 
+    # Phase 24: ROI 月報（每日 09:00 跑，內部判斷是否月初第 1 日）
+    schedule.every().day.at("09:00").do(run_roi_monthly_report_if_new_month)
+    logger.info("📅 每日 09:00：roi_monthly_report（月初第 1 日才送）")
+
     schedule.every().day.at("03:00").do(run_db_backup_task)
     logger.info("📅 每日 03:00：db_backup")
 
@@ -270,6 +274,24 @@ def run_cost_throttle_reset_if_new_month():
         logger.error(f"[CostThrottle] reset failed: {e}", exc_info=True)
 
 
+def run_roi_monthly_report_if_new_month():
+    """每日 09:00 — Phase 24 ROI 月報（內部判斷月初第 1 日才送）
+
+    對比上月 ai_calls 統計 vs 戰前 baseline，推 Telegram「節省 X tokens / $Y」。
+    寫入 ai_insights (type='roi_monthly_report') 作長期記錄。
+    """
+    try:
+        from datetime import datetime
+        if datetime.now().day != 1:
+            return  # 非月初第 1 日 skip
+        from services.roi_report_service import generate_and_send_roi_report
+        result = generate_and_send_roi_report()
+        logger.info("[ROIReport] sent=%s period=%s",
+                    result.get('sent'), result.get('period', '?'))
+    except Exception as e:
+        logger.error(f"[ROIReport] task failed: {e}", exc_info=True)
+
+
 def run_embed_consistency_check():
     """每週日 04:30 — BGE-M3 跨主機一致性驗證（ADR-033 護欄 #3）。
 
diff --git a/services/roi_report_service.py b/services/roi_report_service.py
new file mode 100644
index 0000000..63be39a
--- /dev/null
+++ b/services/roi_report_service.py
@@ -0,0 +1,231 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+services/roi_report_service.py
+Operation Ollama-First v5.0 / Phase 24 — ROI 月報生成器
+
+設計原則：
+- 每月 1 日 09:00 跑（與 daily_report 同時段）
+- SQL 統計上月 ai_calls × mcp_calls × rag_query_log
+- 對比戰前 baseline（hardcode 戰前估算數字）
+- 推 Telegram「上月節省 X tokens / $Y / 攔截 Z 次 LLM 呼叫」
+- 寫進 ai_insights（type='roi_monthly_report'）作長期記錄
+
+戰前 baseline（戰役 v5.0 啟動前估算）:
+- Gemini 月 ~50M tokens / $20
+- NIM 月 ~5.8M tokens
+- Ollama 月 ~30M tokens（自架免費）
+- Total ~$25/月（含 OpenRouter）
+"""
+
+from __future__ import annotations
+import os
+import logging
+from datetime import datetime, timedelta
+from calendar import monthrange
+from typing import Dict, Any, Optional
+
+logger = logging.getLogger(__name__)
+
+# 戰前 baseline（v5.0 啟動前估算，後續可從 ai_call_budgets 表動態算）
+BASELINE = {
+    'gemini_monthly_tokens': 50_000_000,
+    'gemini_monthly_cost_usd': 20.0,
+    'nim_monthly_tokens': 5_800_000,
+    'ollama_monthly_tokens': 30_000_000,
+    'total_monthly_cost_usd': 25.0,
+}
+
+
+def _last_month_range(today: Optional[datetime] = None) -> tuple:
+    """取上月 1 號 00:00 到 本月 1 號 00:00"""
+    today = today or datetime.now()
+    this_month_start = datetime(today.year, today.month, 1)
+    if today.month == 1:
+        last_month_start = datetime(today.year - 1, 12, 1)
+    else:
+        last_month_start = datetime(today.year, today.month - 1, 1)
+    return last_month_start, this_month_start
+
+
+def query_last_month_stats() -> Dict[str, Any]:
+    """SQL 查上月 ai_calls / mcp_calls / rag_query_log 統計"""
+    try:
+        from sqlalchemy import text as sa_text
+        from database.manager import get_session
+    except Exception as exc:
+        logger.warning('[ROI] DB import failed: %s', exc)
+        return {}
+
+    last_start, this_start = _last_month_range()
+    period_label = last_start.strftime('%Y年%m月')
+
+    session = get_session()
+    try:
+        # 1. ai_calls 統計
+        ai = session.execute(
+            sa_text("""
+                SELECT
+                  COALESCE(SUM(input_tokens + output_tokens), 0) AS total_tokens,
+                  COALESCE(SUM(cost_usd), 0) AS total_cost,
+                  COUNT(*) AS total_calls,
+                  COUNT(*) FILTER (WHERE provider IN ('gcp_ollama','ollama_secondary','ollama_111')) AS ollama_calls,
+                  COUNT(*) FILTER (WHERE provider = 'gemini') AS gemini_calls,
+                  COUNT(*) FILTER (WHERE provider = 'claude') AS claude_calls,
+                  COUNT(*) FILTER (WHERE provider IN ('nim','nim_via_elephant')) AS nim_calls,
+                  COUNT(*) FILTER (WHERE rag_hit) AS rag_hit_calls,
+                  COUNT(*) FILTER (WHERE cache_hit) AS cache_hit_calls,
+                  COALESCE(SUM(input_tokens + output_tokens) FILTER (WHERE provider = 'gemini'), 0) AS gemini_tokens,
+                  COALESCE(SUM(cost_usd) FILTER (WHERE provider = 'gemini'), 0) AS gemini_cost,
+                  COALESCE(SUM(cost_usd) FILTER (WHERE provider = 'claude'), 0) AS claude_cost
+                FROM ai_calls
+                WHERE called_at >= :start AND called_at < :end
+            """),
+            {'start': last_start, 'end': this_start},
+        ).fetchone()
+
+        # 2. mcp_calls 統計
+        try:
+            mcp = session.execute(
+                sa_text("""
+                    SELECT COUNT(*) AS total, COUNT(*) FILTER (WHERE cache_hit) AS cache_hits
+                    FROM mcp_calls
+                    WHERE called_at >= :start AND called_at < :end
+                """),
+                {'start': last_start, 'end': this_start},
+            ).fetchone()
+        except Exception:
+            mcp = None
+
+        # 3. rag_query_log 統計
+        try:
+            rag = session.execute(
+                sa_text("""
+                    SELECT
+                      COUNT(*) AS total_queries,
+                      COUNT(*) FILTER (WHERE saved_call) AS saved_calls,
+                      AVG(feedback_score) FILTER (WHERE feedback_score IS NOT NULL) AS avg_feedback
+                    FROM rag_query_log
+                    WHERE queried_at >= :start AND queried_at < :end
+                """),
+                {'start': last_start, 'end': this_start},
+            ).fetchone()
+        except Exception:
+            rag = None
+
+        return {
+            'period_label': period_label,
+            'period_start': last_start,
+            'period_end': this_start,
+            'ai_total_tokens': int(ai[0] or 0) if ai else 0,
+            'ai_total_cost': float(ai[1] or 0) if ai else 0.0,
+            'ai_total_calls': int(ai[2] or 0) if ai else 0,
+            'ollama_calls': int(ai[3] or 0) if ai else 0,
+            'gemini_calls': int(ai[4] or 0) if ai else 0,
+            'claude_calls': int(ai[5] or 0) if ai else 0,
+            'nim_calls': int(ai[6] or 0) if ai else 0,
+            'rag_hit_calls': int(ai[7] or 0) if ai else 0,
+            'cache_hit_calls': int(ai[8] or 0) if ai else 0,
+            'gemini_tokens': int(ai[9] or 0) if ai else 0,
+            'gemini_cost': float(ai[10] or 0) if ai else 0.0,
+            'claude_cost': float(ai[11] or 0) if ai else 0.0,
+            'mcp_total': int(mcp[0] or 0) if mcp else 0,
+            'mcp_cache_hits': int(mcp[1] or 0) if mcp else 0,
+            'rag_total': int(rag[0] or 0) if rag else 0,
+            'rag_saved': int(rag[1] or 0) if rag else 0,
+            'rag_avg_feedback': float(rag[2] or 0) if rag and rag[2] else 0.0,
+        }
+    except Exception as exc:
+        logger.error('[ROI] query failed: %s', exc)
+        return {}
+    finally:
+        session.close()
+
+
+def render_roi_report(stats: Dict[str, Any]) -> str:
+    """組 Telegram 訊息（HTML format）"""
+    if not stats:
+        return "⚠️ ROI 月報資料查詢失敗，請查 logs"
+
+    period = stats['period_label']
+    gemini_saved_tokens = max(0, BASELINE['gemini_monthly_tokens'] - stats['gemini_tokens'])
+    gemini_saved_cost = max(0.0, BASELINE['gemini_monthly_cost_usd'] - stats['gemini_cost'])
+    saved_pct = (
+        gemini_saved_tokens / BASELINE['gemini_monthly_tokens'] * 100
+        if BASELINE['gemini_monthly_tokens'] else 0
+    )
+
+    return (
+        f"📊 <b>ROI 月報 {period}</b>\n"
+        f"━━━━━━━━━━━━━━━━━━━━\n"
+        f"💰 <b>成本攔截</b>\n"
+        f"  Gemini: {stats['gemini_tokens']:,} tokens / ${stats['gemini_cost']:.2f}\n"
+        f"  vs 戰前: {BASELINE['gemini_monthly_tokens']:,} / ${BASELINE['gemini_monthly_cost_usd']:.2f}\n"
+        f"  ✅ 攔截: <b>{gemini_saved_tokens:,} tokens / ${gemini_saved_cost:.2f} ({saved_pct:.1f}%)</b>\n"
+        f"\n"
+        f"🤖 <b>Provider 分布</b>\n"
+        f"  Ollama (自架免費): {stats['ollama_calls']:,} calls\n"
+        f"  Gemini: {stats['gemini_calls']:,} calls\n"
+        f"  Claude: {stats['claude_calls']:,} calls / ${stats['claude_cost']:.2f}\n"
+        f"  NIM: {stats['nim_calls']:,} calls\n"
+        f"\n"
+        f"🧠 <b>RAG 自主學習</b>\n"
+        f"  查詢: {stats['rag_total']:,} 次\n"
+        f"  攔截 LLM: {stats['rag_saved']:,} 次（saved_call=true）\n"
+        f"  反饋分: {stats['rag_avg_feedback']:.2f} / 5\n"
+        f"\n"
+        f"🔧 <b>MCP + Cache</b>\n"
+        f"  MCP 呼叫: {stats['mcp_total']:,}\n"
+        f"  Cache 命中: {stats['cache_hit_calls']:,} ai_calls + {stats['mcp_cache_hits']:,} mcp_calls\n"
+        f"\n"
+        f"📈 <b>戰役 v5.0 KPI</b>\n"
+        f"  Gemini -23.5% 目標：{'✅ 達標' if saved_pct >= 23 else f'⚠️ {saved_pct:.1f}%'}\n"
+        f"  RAG 命中 ≥25% 目標：{'✅' if stats['rag_total'] > 0 and stats['rag_saved']/max(stats['rag_total'],1) >= 0.25 else '⏳'}"
+    )
+
+
+def generate_and_send_roi_report() -> Dict[str, Any]:
+    """每月 1 日 09:00 cron 呼叫主入口"""
+    stats = query_last_month_stats()
+    if not stats:
+        logger.warning('[ROI] no stats; skip')
+        return {'sent': False, 'reason': 'no_stats'}
+
+    msg = render_roi_report(stats)
+
+    # 推 Telegram
+    try:
+        from services.telegram_templates import _send_telegram_raw
+        _send_telegram_raw(msg)
+        logger.info('[ROI] %s 月報已推 Telegram', stats['period_label'])
+    except Exception as exc:
+        logger.warning('[ROI] telegram push failed: %s', exc)
+
+    # 寫 ai_insights 長期記錄
+    try:
+        from sqlalchemy import text as sa_text
+        from database.manager import get_session
+        session = get_session()
+        try:
+            session.execute(
+                sa_text("""
+                    INSERT INTO ai_insights (insight_type, content, status, created_by, confidence)
+                    VALUES ('roi_monthly_report', :content, 'approved', 'roi_report_service', 0.95)
+                """),
+                {'content': msg},
+            )
+            session.commit()
+        finally:
+            session.close()
+    except Exception as exc:
+        logger.warning('[ROI] persist failed: %s', exc)
+
+    return {'sent': True, 'period': stats['period_label'], 'msg_chars': len(msg)}
+
+
+__all__ = [
+    'query_last_month_stats',
+    'render_roi_report',
+    'generate_and_send_roi_report',
+    'BASELINE',
+]
diff --git a/tests/test_roi_report_service.py b/tests/test_roi_report_service.py
new file mode 100644
index 0000000..99a100f
--- /dev/null
+++ b/tests/test_roi_report_service.py
@@ -0,0 +1,120 @@
+"""
+tests/test_roi_report_service.py
+─────────────────────────────────────────────────────────────────
+Operation Ollama-First v5.0 / Phase 24 — ROI 月報生成器驗證
+"""
+
+from datetime import datetime
+from unittest.mock import patch
+
+import pytest
+
+
+def test_baseline_constants():
+    """戰前 baseline 應有 5 個必要欄位"""
+    from services.roi_report_service import BASELINE
+
+    required = {'gemini_monthly_tokens', 'gemini_monthly_cost_usd',
+                'nim_monthly_tokens', 'ollama_monthly_tokens',
+                'total_monthly_cost_usd'}
+    assert required.issubset(BASELINE.keys())
+    assert BASELINE['gemini_monthly_tokens'] == 50_000_000
+    assert BASELINE['gemini_monthly_cost_usd'] == 20.0
+
+
+def test_last_month_range_january():
+    """1 月跑時上月應為去年 12 月"""
+    from services.roi_report_service import _last_month_range
+
+    today = datetime(2026, 1, 5)
+    last_start, this_start = _last_month_range(today)
+    assert last_start == datetime(2025, 12, 1)
+    assert this_start == datetime(2026, 1, 1)
+
+
+def test_last_month_range_normal_month():
+    from services.roi_report_service import _last_month_range
+
+    today = datetime(2026, 5, 15)
+    last_start, this_start = _last_month_range(today)
+    assert last_start == datetime(2026, 4, 1)
+    assert this_start == datetime(2026, 5, 1)
+
+
+def test_render_roi_report_with_savings():
+    """戰役後成功攔截 Gemini → 訊息含「攔截」+ 達標 ✅"""
+    from services.roi_report_service import render_roi_report
+
+    stats = {
+        'period_label': '2026年04月',
+        'period_start': datetime(2026, 4, 1),
+        'period_end': datetime(2026, 5, 1),
+        'ai_total_tokens': 80_000_000,
+        'ai_total_cost': 12.0,
+        'ai_total_calls': 5000,
+        'ollama_calls': 4000,
+        'gemini_calls': 800,
+        'claude_calls': 100,
+        'nim_calls': 100,
+        'rag_hit_calls': 200,
+        'cache_hit_calls': 150,
+        'gemini_tokens': 35_000_000,  # 戰前 50M → 省 15M (30%)
+        'gemini_cost': 14.0,           # 戰前 $20 → 省 $6
+        'claude_cost': 5.0,
+        'mcp_total': 180,
+        'mcp_cache_hits': 50,
+        'rag_total': 800,
+        'rag_saved': 250,              # 31% RAG 攔截率
+        'rag_avg_feedback': 4.2,
+    }
+    msg = render_roi_report(stats)
+
+    assert '2026年04月' in msg
+    assert '攔截' in msg
+    assert '15,000,000' in msg or '15M' in msg or '15,000' in msg
+    assert '$6' in msg or '6.00' in msg
+    assert '✅' in msg  # Gemini -23.5% 目標達標
+
+
+def test_render_roi_report_below_target():
+    """未達 -23.5% 目標 → 訊息含 ⚠️"""
+    from services.roi_report_service import render_roi_report
+
+    stats = {
+        'period_label': '2026年04月',
+        'period_start': datetime(2026, 4, 1),
+        'period_end': datetime(2026, 5, 1),
+        'ai_total_tokens': 80_000_000, 'ai_total_cost': 18.0, 'ai_total_calls': 5000,
+        'ollama_calls': 1000, 'gemini_calls': 3000, 'claude_calls': 500, 'nim_calls': 500,
+        'rag_hit_calls': 50, 'cache_hit_calls': 30,
+        'gemini_tokens': 45_000_000,  # 戰前 50M → 只省 10% < 23%
+        'gemini_cost': 18.0, 'claude_cost': 5.0,
+        'mcp_total': 100, 'mcp_cache_hits': 10,
+        'rag_total': 200, 'rag_saved': 30, 'rag_avg_feedback': 3.5,
+    }
+    msg = render_roi_report(stats)
+
+    assert '⚠️' in msg  # 未達標標記
+
+
+def test_render_empty_stats():
+    """空 stats → 預設失敗訊息"""
+    from services.roi_report_service import render_roi_report
+
+    assert '⚠️' in render_roi_report({})
+
+
+def test_query_last_month_stats_db_fail_returns_empty(monkeypatch):
+    """DB 失敗應回 {} 不 raise"""
+    from services.roi_report_service import query_last_month_stats
+
+    class _BrokenSession:
+        def execute(self, *a, **kw):
+            raise RuntimeError('DB connection lost')
+        def close(self):
+            pass
+
+    import services.roi_report_service as ros
+    monkeypatch.setattr('database.manager.get_session', lambda: _BrokenSession())
+    result = ros.query_last_month_stats()
+    assert result == {}