AI Sales Recommendation & Follow-Up Assistant for MSME

This research archive documents the journal-informed architecture pivot from an AI Lead Follow-Up Assistant into an AI Sales Recommendation & Follow-Up Assistant for MSME.

The goal is to connect a scientific journal reference about MSME e-commerce and recommendation systems with a practical AI agent architecture for customer inquiry processing, product/service recommendation, sales follow-up, CRM tracking, and human-reviewed decision support.

---

🔬 Key Objectives

• Translate a scientific journal problem into an engineering architecture.
• Map MSME e-commerce and recommendation-system concepts into an AI agent workflow.
• Extend the existing lead follow-up assistant into a sales recommendation assistant.
• Preserve human review and owner control before any customer-facing message is sent.
• Build a stronger portfolio case study based on business problem, system architecture, and journal-informed reasoning.

---

📚 Journal Reference

Paper:

Implementasi E-Commerce dengan Sistem Informasi Rekomendasi menggunakan Metode Collaborative Filtering untuk Pengembangan Penjualan pada UMKM

Journal:

Jurnal Sistem Informasi Bisnis, Vol. 15 No. 1, 2025

Core Topic:

E-commerce, MSME, recommendation information system, collaborative filtering, sales development

Journal Link:

📄 https://ejournal.undip.ac.id/index.php/jsinbis/article/view/68333

---

🧩 Journal-Based Problem

The journal describes several important MSME problems:

1. Many MSME still run traditional business processes.
2. MSME need better digital sales channels through e-commerce.
3. Many MSME lack proper promotional platforms.
4. Customers need help choosing MSME products quickly and accurately.
5. A recommendation system can help customers find relevant products.
6. Collaborative filtering can use MSME data, consumer data, and rating data to generate recommendations.

The journal uses collaborative filtering to recommend products based on user ratings and product similarity.

The recommendation process includes:

rating data
-> average user rating
-> product similarity calculation
-> weighted prediction
-> highest predicted product recommendation

---

🧠 Business Case Interpretation

The business case is not only:

MSME need e-commerce.

The deeper business case is:

MSME need a system that can understand customer interest, recommend the most relevant product or service, and help the owner follow up with the right offer.

This is where the existing AI Lead Follow-Up Assistant becomes valuable.

The old system already solves:

lead inquiry
-> lead classification
-> follow-up draft
-> CRM status
-> human review

The journal adds a new layer:

customer/product interaction
-> recommendation engine
-> next best offer
-> sales follow-up

The combined system becomes:

customer inquiry
-> intent extraction
-> lead qualification
-> product/service recommendation
-> next best offer
-> follow-up draft
-> owner review
-> CRM and feedback tracker

---

🧱 Existing Project Reuse Mapping

The current project is not wasted.

It becomes the foundation layer of the new architecture.

---

🧭 New Project Direction

Previous Direction

AI Lead Follow-Up Assistant

Main purpose:

Help small service businesses classify leads, draft follow-up messages, and track CRM status with human review.

New Direction

AI Sales Recommendation & Follow-Up Assistant for MSME

Main purpose:

Help MSME understand customer interest, recommend relevant products or services, generate next-best-offer follow-up drafts, and track owner-reviewed sales actions.

---

❗ Updated Problem Statement

MSME often struggle not only because they reply slowly, but because customer interest is not converted into a relevant sales recommendation.

Incoming messages, product interest, customer preferences, budget signals, and prior interactions are often scattered.

As a result:

• Owners respond manually.
• Customer intent is not structured.
• Product or service recommendations are not data-informed.
• Follow-up messages are not tied to the best offer.
• Customer response history is not reused.
• Sales opportunities can be lost.

---

✅ Updated Solution Statement

AI Sales Recommendation & Follow-Up Assistant for MSME is a journal-informed AI agent architecture that combines:

• Customer inquiry intake.
• Intent extraction.
• Lead qualification.
• Product/service catalog matching.
• Recommendation engine.
• Next best offer generation.
• Follow-up draft generation.
• QA guard.
• Human review.
• CRM and feedback tracking.

The system is designed to help MSME turn customer inquiries and interaction data into structured, relevant, and owner-reviewed sales actions.

---

🎯 Updated Result Goal

The final engineering snapshot should demonstrate:

customer inquiry
-> customer intent extraction
-> lead qualification
-> product/service recommendation
-> next best offer
-> follow-up draft
-> human review
-> tracker update
-> customer feedback loop

The result is not just faster reply generation.

The result is a more structured sales decision workflow.

---

🧬 Journal-to-System Mapping

---

🏗️ Final Target Architecture

Customer Channels
-> Lead & Interaction Intake
-> Data Normalization
-> Customer Intent Extraction Agent
-> Lead Qualification Agent
-> Product / Service Catalog
-> Recommendation Engine
-> Next Best Offer Agent
-> Follow-Up Drafting Agent
-> QA Guard Agent
-> Human Review
-> CRM / Google Sheet Tracker
-> Customer Response & Rating Feedback
-> Recommendation Data Store

---

🗺️ Architecture Diagram

[Instagram DM / WhatsApp / Form / E-Commerce Interaction]
                    |
                    v
        [Lead & Interaction Intake Layer]
                    |
                    v
          [Normalization & Validation]
                    |
                    v
        [Customer Intent Extraction Agent]
                    |
                    v
        [Lead Qualification Agent]
                    |
                    v
        [Product / Service Catalog]
                    |
                    v
        [Recommendation Engine]
        - rule-based matching
        - collaborative filtering simulation
        - rating / interaction scoring
                    |
                    v
        [Next Best Offer Agent]
                    |
                    v
        [Follow-Up Drafting Agent]
                    |
                    v
        [QA Guard Agent]
                    |
                    v
        [Human Review Layer]
                    |
                    v
        [CRM / Google Sheet Tracker]
                    |
                    v
        [Customer Response / Rating / Feedback]
                    |
                    v
        [Recommendation Learning Data]

---

New Core Modules

---

1. Lead & Interaction Intake Layer

Purpose:

Capture customer inquiry and product/service interaction.

Input examples:

• Instagram DM.
• WhatsApp message.
• Google Form.
• Product inquiry.
• Service inquiry.
• Rating.
• Manual owner entry.

---

2. Customer Intent Extraction Agent

Purpose:

Extract customer intent from raw inquiry text.

Output example:

{
  "customer_intent": "looking_for_service",
  "interest_type": "social_media_management",
  "problem_signal": "needs help managing Instagram",
  "preference_signal": "monthly support",
  "urgency_signal": "this month"
}

---

3. Lead Qualification Agent

Purpose:

Score and classify the lead using the existing deterministic rules.

This reuses:

docs/lead-scoring-rules.md

Existing score criteria:

• Budget clarity.
• Timeline urgency.
• Service clarity.
• Buying intent.

---

4. Product / Service Catalog Layer

Purpose:

Store available MSME products or services that can be recommended.

Initial MVP can use a service catalog instead of a full product catalog.

Future file:

data/service-catalog.json

Example:

[
  {
    "offer_id": "SVC001",
    "offer_name": "Instagram Feed Design Starter",
    "category": "design",
    "best_for": ["coffee shop", "small brand", "low budget"],
    "price_range": "under 1 juta",
    "required_context": ["number of designs", "style reference"]
  },
  {
    "offer_id": "SVC002",
    "offer_name": "Monthly Social Media Management",
    "category": "social_media_management",
    "best_for": ["skincare brand", "restaurant", "fashion brand"],
    "price_range": "2-5 juta/bulan",
    "required_context": ["current account", "monthly content target"]
  },
  {
    "offer_id": "SVC003",
    "offer_name": "Launch Content Package",
    "category": "campaign",
    "best_for": ["fashion brand", "product launch"],
    "price_range": "3-8 juta/project",
    "required_context": ["launch date", "target audience", "asset needs"]
  }
]

---

5. Recommendation Engine

Purpose:

Recommend the most relevant product or service based on customer intent, catalog fit, and interaction data.

The recommendation engine should evolve in three levels.

---

Level 1 — Rule-Based Recommendation MVP

Use:

customer intent
+ service interest
+ business type
+ budget range
+ timeline

To recommend:

best matching offer

Example:

Skincare Brand
+ Social Media Management
+ 2-3 juta/bulan
+ Bulan ini
= Monthly Social Media Management

---

Level 2 — Interaction-Based Recommendation

Add customer behavior signals:

viewed_offer
asked_price
requested_portfolio
accepted_proposal
rejected_offer
sent_rating

---

Level 3 — Collaborative Filtering Simulation

Use rating or interaction data to simulate collaborative filtering.

Input data:

customer_id
offer_id
rating
interaction_type
created_at

Process:

rating data
-> average customer rating
-> offer similarity
-> weighted prediction
-> highest positive recommendation

---

6. Next Best Offer Agent

Purpose:

Convert recommendation engine output into a sales action.

Output example:

{
  "recommended_offer_id": "SVC002",
  "recommended_offer_name": "Monthly Social Media Management",
  "recommendation_reason": "The lead is a skincare brand asking for monthly Instagram management with clear budget and this-month timeline.",
  "recommendation_confidence": "high",
  "offer_action": "ask_discovery_before_package"
}

---

7. Follow-Up Drafting Agent

Purpose:

Draft a follow-up message using customer intent, lead score, recommended offer, missing information, and human review rule.

Example:

Halo Kak Dinda, bisa banget Kak. Dari kebutuhan Kakak untuk kelola IG skincare brand dengan timeline bulan ini, kemungkinan paling cocok arahnya ke monthly social media management. Sebelum aku kasih rekomendasi paket, boleh aku lihat akun Instagram-nya dan target konten bulan ini seperti apa?

---

8. QA Guard Agent

Purpose:

Validate recommendation and follow-up draft before owner review.

Existing QA rules remain active.

Additional recommendation-specific risk flags:

low_recommendation_confidence
insufficient_interaction_data
cold_start_customer
cold_start_offer
offer_budget_mismatch
recommendation_needs_review

---

9. Human Review Layer

Purpose:

Keep the owner in control.

The system must not auto-send messages.

Workflow:

AI recommends
-> AI drafts
-> owner reviews
-> owner approves / edits / rejects
-> owner sends manually

---

10. CRM and Feedback Tracker

Purpose:

Store lead data, AI output, recommendation output, owner review, customer response, and rating feedback.

Future tracker extension:

tracker/recommendation-tracker-columns.csv

Potential new columns:

customer_id
recommended_offer_id
recommended_offer_name
recommendation_score
recommendation_confidence
recommendation_reason
offer_action
customer_response
customer_rating
interaction_type
feedback_recorded_at

---

🧪 Final Engineering Snapshot Target

The final target is:

Journal-informed AI agent architecture for MSME sales recommendation and follow-up.

The final snapshot should prove:

1. Lead inquiry can be captured.
2. Customer intent can be extracted.
3. Lead priority can be scored.
4. Product/service catalog can be matched.
5. A relevant offer can be recommended.
6. A follow-up message can be drafted.
7. QA guard can detect risk.
8. Owner review is preserved.
9. CRM tracker stores the sales action.
10. Customer response or rating can feed future recommendations.

---

📁 Updated Repository Direction

Add these new artifacts without deleting the previous project foundation:

docs/
  journal-business-case-mapping.md
  final-system-architecture.md
  recommendation-engine-design.md
  customer-interaction-model.md
  recommendation-output-schema.md

data/
  service-catalog.json
  customer-interactions.csv
  sample-ratings.csv

tracker/
  recommendation-tracker-columns.csv

n8n/
  recommendation-node-prompt.md

outputs/
  recommendation-test-C001.json

case-study/
  ai-sales-recommendation-followup-case-study-v1.md

---

🛣️ Roadmap After This Pivot

---

Phase 2A — Stabilize Existing Automation

Continue the current semi-automated workflow:

manual trigger
-> AI processing
-> JSON parsing
-> Google Sheet lead tracker

---

Phase 2B — Add Recommendation Layer

Add:

service catalog
customer interaction schema
rule-based recommendation logic
next best offer agent
recommendation output schema

---

Phase 3 — Collaborative Filtering Simulation

Build a simple simulation:

sample-ratings.csv
-> average rating
-> similarity matrix
-> weighted prediction
-> recommended offer

---

Phase 4 — Portfolio Case Study v1

Upgrade case study from:

AI Lead Follow-Up Assistant

Into:

AI Sales Recommendation & Follow-Up Assistant for MSME

---

🧭 Strategic Decision

The existing project should not be discarded.

The project should be extended.

Final strategic framing:

AI Lead Follow-Up Assistant is the CRM and follow-up foundation.

The journal-based recommendation layer turns it into an AI Sales Recommendation & Follow-Up Assistant for MSME.

This creates a stronger portfolio because it combines:

• AI agent workflow.
• MSME business problem.
• Recommendation system concept.
• Lead qualification.
• Sales follow-up.
• CRM tracking.
• Human review safety.
• Scientific-journal-informed architecture.

---

✅ Definition of Done

This chunk is complete when:

1. docs/journal-business-case-mapping.md exists.
2. The journal problem is summarized.
3. The business case is translated into system architecture.
4. Existing project artifacts are mapped to the new architecture.
5. The new project direction is defined.
6. The final architecture direction is documented.
7. Recommendation layer modules are identified.
8. Human review safety remains preserved.
9. The next roadmap is clear.

---

Final Archive Summary

This research archive confirms that the current AI Lead Follow-Up Assistant project remains useful as the foundation layer.

The stronger direction is to extend it into a journal-informed architecture:

AI Sales Recommendation & Follow-Up Assistant for MSME

This direction aligns technical implementation with a real MSME business problem: not only replying faster, but recommending the right product or service, drafting the right follow-up, preserving owner review, and storing feedback for future recommendation improvement.