Agentic AI Customer Support Assistant

Project Overview

I designed and built an Agentic AI Customer Support Assistant for ITIL-aligned Managed Service Provider (MSP) environments, focused on transforming raw support tickets into structured, actionable insights for technicians.

The system acts as an AI-powered triage and decision-support layer on top of traditional Service Desk workflows, helping teams move from vague or incomplete tickets to clear problem understanding, validated priority, and guided resolution steps.

At its core, the platform combines LLM reasoning, workflow orchestration, and structured output enforcement to deliver reliable, auditable, and production-ready AI assistance in real-world IT operations.

The system operates in a controlled, advisory-only mode, ensuring safe integration into enterprise environments without disrupting existing workflows.

My Role

I led the end-to-end system design and implementation, including:

AI triage architecture and workflow orchestration (n8n-based)

Prompt engineering for deterministic, ITIL-aligned outputs

Backend workflow design and ticket processing pipelines

Structured AI output schema design (ai_intervention_v1)

Integration strategy with Service Desk systems

System safety, governance, and auditability design

This was built as a production-grade advisory system, not a prototype.

Architecture & Approach

The system follows a pipeline-driven, backend-first architecture:

Ticket ingestion from Service Desk (API/Webhook)

Data normalization and enrichment

AI-powered ticket analysis and classification

Structured intervention generation (JSON-based)

Technician notification and review loop

Feedback logging for continuous improvement

A key design principle was strict separation between AI reasoning and system control, ensuring the AI never performs actions directly but instead assists human decision-making.

Tech Stack & Frameworks

Orchestration: n8n (self-hosted workflow engine)

AI Models: Google Vertex AI (Gemini)

Backend Services: Cloud Run (for scalable APIs & KB retrieval)

Database: PostgreSQL (with pgvector for semantic search)

Integration: ServiceDesk (ticket source)

Data Handling: Structured JSON schemas with strict validation

Key Engineering Contributions

1. AI-Powered Ticket Understanding & Normalization

Designed a preprocessing layer that transforms messy, real-world tickets into structured, high-quality inputs by:

Cleaning and normalizing ticket data

Standardizing priority formats (P1–P4)

Detecting categories (VPN, network, database, etc.)

Identifying VIP users and contextual signals

This ensures consistent and reliable AI reasoning regardless of input quality.

2. Structured AI Intervention System (ai_intervention_v1)

Developed a strict, production-grade AI output schema including:

Problem summary

Priority validation & correction

Business impact assessment

Missing information detection

Suggested solutions and workflow steps

Clarification questions for customers

Confidence scoring

All outputs are JSON-only, ensuring:

Deterministic behavior

Easy validation and debugging

Seamless system integration

3. ITIL-Aligned Priority Validation Engine

Built a controlled priority assessment system where AI:

Validates (not blindly overrides) user-defined priority

Uses impact + urgency models

Applies strict rules (e.g., P1 only for confirmed outages)

Avoids over-escalation and false urgency

This aligns AI decisions with real-world ITIL practices.

4. Runbook & Decision Support Generation

Designed a structured troubleshooting system that generates:

Quick actions (fast fixes with probability estimates)

Step-by-step runbooks (diagnostic → resolution)

Escalation conditions

Risk-aware actions (non-destructive first)

This turns AI from a “chat assistant” into a practical engineering tool.

5. Knowledge Base–Augmented Triage (Hybrid RAG)

Implemented a scalable retrieval system that enhances AI reasoning by:

Generating semantic search queries from tickets

Performing vector similarity search (pgvector)

Injecting relevant KB context only when confidence is high

Avoiding low-quality or irrelevant context injection

This significantly improves solution accuracy while maintaining reliability.

6. Safety, Governance & Human-in-the-Loop Design

Built strict guardrails to ensure enterprise safety:

AI cannot modify tickets

AI cannot escalate automatically

AI cannot contact customers without approval

All outputs require technician review

Full logging of AI suggestions and human corrections

This ensures trust, auditability, and controlled adoption in MSP environments.

7. Technician Feedback Loop & Learning System

Designed a feedback capture system where technicians:

Validate AI priority and solutions

Provide corrections

Add notes and improvements

These logs create a foundation for:

Continuous system improvement

Future learning pipelines

Advanced decision-support capabilities

Key Strengths of the System

AI integrated into real ITIL workflows (not standalone chatbot)

Strong separation of reasoning, control, and execution

Handles messy, real-world ticket data

Deterministic, structured, and auditable outputs

Human-in-the-loop validation for enterprise reliability

Scalable architecture (ready for autonomous phases)

Outcome

The platform enables support teams to:

Understand tickets faster and more accurately

Reduce back-and-forth with users

Improve priority classification consistency

Accelerate troubleshooting and resolution

Build a foundation for AI-driven support automation

It effectively acts as an AI-powered technician assistant, enhancing operational efficiency while maintaining full human control over decisions.